Crane line fault diagnosis system and method based on multi-source data fusion

Abstract

The invention relates to the technical field of crane line fault diagnosis, in particular to a crane line fault diagnosis system and method based on multi-source data fusion, which comprises a data acquisition and processing unit, a mechanical and electrical coupling characteristic unit and a characteristic fusion and fault quantification unit, the three-axis vibration acceleration and the three-phase current waveform of the track are obtained through the data collecting and processing unit, the mechanical and electrical coupling characteristic unit conducts three-dimensional vector synthesis and wavelet packet decomposition on vibration data, and a time-space incidence matrix of harmonic distortion and vibration is constructed. And the feature fusion and fault quantification unit outputs coupling factors by using a bidirectional long-short-term memory network and an attention mechanism, and outputs a fault probability value through a dynamic time warping matching algorithm after time-frequency domain analysis and sample entropy judgment, so that time-space correlation modeling and dynamic fault matching of multi-source data are realized. And the fault positioning precision and the diagnosis accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crane line fault diagnosis, in particular to a crane line fault diagnosis system and method based on multi-source data fusion. BACKGROUND

[0002] Crane line fault intelligent diagnosis is an important technology. In modern industrial logistics and heavy machinery operation, crane line fault diagnosis technology is a core means to ensure the safe operation of hoisting equipment and prevent major accidents by monitoring the running state of the equipment in real time and positioning the line abnormalities.

[0003] With the development of industrial intelligence and the increasing complexity of cranes, building a multi-source data diagnosis system that integrates mechanical vibration and electrical parameters is of great significance to improve fault identification accuracy and reduce unplanned downtime. Traditional single-parameter analysis diagnosis mode has been unable to meet the complex diagnosis needs of mechanical-electrical coupling faults. However, the traditional crane line fault diagnosis technology has the core problems of insufficient correlation analysis of mechanical and electrical parameters and low recognition rate of coupling faults. Existing solutions only analyze single parameters such as vibration or current, and do not establish a spatio-temporal correlation model of mechanical vibration and electrical parameters. When track deformation causes poor contact, the correlation between vibration energy mutation and current harmonic distortion is delayed for more than 200ms, resulting in fault positioning error exceeding the preset threshold. Single signal processing method cannot effectively separate the multi-band characteristics of mechanical vibration and the harmonic components of electrical parameters, resulting in low feature extraction accuracy in the 100-200Hz coupling fault frequency band, and increasing the misjudgment rate of insulation aging and poor contact. In addition, the traditional system lacks a dynamic fault pattern matching mechanism, and cannot adaptively update the fault template when facing track deformation, insulation aging, and poor contact fault types, resulting in low recognition success rate of new fault types. This lack of multi-source data correlation and insufficient dynamic matching capability ultimately makes the comprehensive fault diagnosis accuracy of the traditional solution low under complex working conditions, making it difficult to meet the high reliability operation requirements of hoisting equipment. To solve this problem, we provide a crane line fault diagnosis system and method based on multi-source data fusion. SUMMARY

[0004] The present application aims to provide a crane line fault diagnosis system and method based on multi-source data fusion to solve the problems raised in the background.

[0005] 1. Due to the insufficient correlation analysis of mechanical and electrical parameters in traditional technology, the correlation between vibration and current parameters is delayed for a long time, and the positioning error is large. Therefore, this case constructs a spatio-temporal correlation matrix through a mechanical-electrical coupling feature unit, which can shorten the correlation delay and improve the fault positioning accuracy.

[0006] 2. Due to the low coupling fault recognition rate of traditional technology, the low feature extraction accuracy rate and high misjudgment rate in the 100-200Hz frequency band, the case can improve the feature extraction accuracy rate and reduce the misjudgment rate through feature fusion and fault quantization units.

[0007] To achieve the above object, a crane line fault diagnosis system based on multi-source data fusion is provided, which comprises the following units: The data acquisition and processing unit processes the three-axis vibration acceleration data collected by the three-axis acceleration sensor to generate a vibration vector sequence, collects the three-phase current waveform of the power supply line through the current sensor, synchronously extracts the harmonic components, and calculates the total harmonic distortion rate and the odd-even harmonic energy ratio according to the harmonic components; The mechanical and electrical coupling feature unit performs three-dimensional vector synthesis on the vibration vector sequence to generate an orbit deformation energy distribution matrix, maps it to 32 sub-bands through wavelet packet decomposition, calculates the energy proportion of each sub-band, and constructs a harmonic distortion and vibration space-time correlation matrix according to the energy proportion and the total harmonic distortion rate; The feature fusion and fault quantization unit inputs the space-time correlation matrix and the odd-even harmonic energy ratio into a bidirectional long short-term memory network, dynamically allocates feature weights through an attention mechanism, outputs a coupling factor, performs time-frequency domain joint analysis on the coupling factor to extract an intrinsic mode function, calculates the sample entropy of each intrinsic mode function, determines whether there is an electrical parameter anomaly caused by mechanical vibration according to the sample entropy, and outputs a fault probability value according to the determination result using a dynamic time warping matching algorithm.

[0008] As a further improvement of the technical solution, the method for generating an orbit deformation energy distribution matrix through three-dimensional vector synthesis comprises: The X, Y and Z axis vibration acceleration data collected by the three-axis acceleration sensor are processed to calculate the square value of each axis acceleration, and the square values are integrated and accumulated according to a time window length of 100ms and an overlap rate of 50% to obtain three-axis energy components. The three-axis energy components are stacked in time sequence to generate an orbit deformation energy distribution matrix, and the matrix is normalized.

[0009] As a further improvement of the technical solution, the wavelet packet decomposition method comprises: A wavelet basis function is introduced to decompose the orbit deformation energy distribution matrix by 5 layers to generate 32 sub-bands, calculate the energy proportion of each sub-band, and sort the energy values from high to low. When the absolute value of the natural logarithm of the energy ratio of a certain sub-band to the adjacent frequency band is greater than 1.5, it is determined as an outlier band and is removed.

[0010] As a further improvement of the technical solution, the construction method of the space-time correlation matrix comprises: The Pearson correlation coefficient is calculated for the reserved sub-band energy proportion and the total harmonic distortion rate respectively, a 32x1 dimensional correlation coefficient vector is generated, and a weight vector is calculated according to the energy entropy of each sub-band, the correlation coefficient vector and the weight vector are multiplied to generate a space-time correlation matrix.

[0011] As a further improvement of the technical solution, the bi-directional long short-term memory network first expands the 32-dimensional space-time correlation matrix into a continuous vector by row and concatenates it with the odd-even harmonic energy ratio to form a 33-dimensional input vector, and then calculates the forward hidden state step by step in time sequence, and each time step state is determined by the hidden output of the previous time step and the current input vector, while the backward hidden state is calculated step by step in reverse time sequence, and each time step state is determined by the hidden output of the next time step and the current input vector, and finally the forward and backward hidden states at the same time step are concatenated into a 128-dimensional fusion vector.

[0012] As a further improvement of the technical solution, the attention mechanism first extracts the 128-dimensional fusion vectors of all time steps within the historical time window, calculates the correlation strength score of the current fusion vector and each historical vector through a weight matrix and activates it with a hyperbolic tangent function, converts the score into a probability distribution weight through an exponential function and summation operation, and then generates a context vector representing key historical features by weighting and summing the historical vectors according to the weight, and finally concatenates the current fusion vector and the context vector and outputs a coupling factor in the range of 0 to 1 through an S-shaped growth curve function.

[0013] As a further improvement of the technical solution, the time-frequency domain joint analysis first identifies all maximum and minimum points of the time series of the coupling factor, connects the maximum points to form an upper envelope line and the minimum points to form a lower envelope line using cubic spline interpolation, calculates the point-by-point arithmetic mean of the upper and lower envelope lines as the mean envelope, and subtracts the mean envelope from the original coupling factor and repeatedly selects until the difference between the number of extreme points and the number of zero-crossing points is not more than 1 and the envelope mean of any point is zero to extract the intrinsic mode function, and finally performs an orthogonal transformation on each intrinsic mode function and calculates the instantaneous phase angle change rate as the instantaneous frequency, and superimposes the instantaneous frequencies of all mode functions to generate a time-frequency energy distribution map.

[0014] As a further improvement of the technical solution, the sample entropy determination first constructs a vector sequence by intercepting a continuous subsequence of length 3 from the third intrinsic mode function with a time delay of 5 sampling points, counts the number of matches in all vector pairs with a distance less than 0.2 times the standard deviation of the sequence, and outputs the sample entropy value based on the matching ratio through natural logarithm operation; When the sample entropy value is greater than 1.2 for 3 seconds, the time-frequency energy integral value in the 100-200Hz frequency band is greater than 0.35, and the kurtosis statistic of the third intrinsic mode function is greater than 4.5, it is determined that there is mechanical vibration causing electrical abnormality.

[0015] As a further improvement of the technical solution, the fault probability value generation first presets a physical constraint template, specifically as follows: The track deformation template is: the coupling factor monotonically increases and the change rate continuously exceeds 0.3; The insulation aging template is: the sample entropy is less than 0.8 and the total harmonic distortion rate is higher than 18%; The poor contact template is: the frequency energy integral peak value in the 50-150Hz frequency band exceeds 0.4; The dynamic programming algorithm is used to calculate the minimum cumulative alignment distance of the to-be-tested sequence and each template, and the distance is mapped to a fault probability value of 0 to 1 through an exponential decay function, when the highest fault probability exceeds 0.8 and meets the abnormal judgment condition of claim 8, the corresponding fault type and probability are output.

[0016] The second object of the present application is to provide a method for realizing a crane line fault diagnosis system based on multi-source data fusion, comprising the following steps: S1, collect track three-axis vibration acceleration and three-phase current waveform, calculate the square value of axial acceleration of vibration data and generate three-axis energy components by 100ms time window and 50% overlap rate integration, extract harmonic components of current data to calculate total harmonic distortion rate and odd-even harmonic energy ratio; S2, superimpose three-axis energy components to generate track deformation energy distribution matrix, use wavelet base function for 5-layer decomposition to get 32 sub-bands, eliminate outlying frequency bands with natural logarithm absolute value of adjacent frequency band energy ratio greater than 1.5, calculate the Pearson correlation coefficient vector of the energy proportion of the reserved sub-bands and the total harmonic distortion rate, and generate a space-time correlation matrix by frequency band energy entropy weighting; S3, expand the space-time correlation matrix and splice it with the odd-even harmonic energy ratio to form a 33-dimensional input vector, extract forward / backward time sequence features through a bidirectional long short-term memory network, output a 128-dimensional fusion vector, and generate a coupling factor based on an attention mechanism dynamic weighting history vector, extract the first five intrinsic mode functions of the coupling factor empirical mode decomposition, calculate the sample entropy value of the third intrinsic mode function, and determine the abnormality by combining the time-frequency energy integral greater than 0.35 and the kurtosis greater than 4.5 in the 100-200Hz frequency band; S4, define a physical template, and use dynamic time warping to calculate the minimum cumulative alignment distance, which is mapped to a fault probability through an exponential decay function, and when the highest probability is greater than 0.8, the fault type is output and the template library is updated.

[0017] Compared with the prior art, the present application has the following advantages: The application solves the problem of insufficient correlation analysis of mechanical and electrical parameters in the prior art by a multi-source data fusion mechanism, correlation analysis of three-axis vibration acceleration and three-phase current waveform data, construction of a time-space correlation matrix, shortens the correlation delay of vibration energy mutation and current harmonic distortion, improves the fault positioning accuracy, separates the mechanical vibration multi-band characteristics and electrical parameter harmonic components by wavelet packet decomposition, combines the bidirectional long short-term memory network and the attention mechanism for feature fusion, improves the feature extraction accuracy of the 100-200Hz coupling fault frequency band, reduces the misjudgment rate of track deformation, insulation aging and poor contact, and in addition, through a dynamic time warping matching algorithm, the fault template is adaptively updated according to a preset physical constraint template, the recognition success rate of new fault types is improved, and the comprehensive fault diagnosis accuracy under complex working conditions is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is the overall block diagram of the application; Figure 2 It is the overall flowchart of the application.

[0019] The meanings of various labels in the figure are as follows: 1, data acquisition and processing unit; 2, mechanical and electrical coupling feature unit; 3, feature fusion and fault quantization unit. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0021] The application provides a crane line fault diagnosis system based on multi-source data fusion, please refer to Figure 1 As shown in the figure, it includes the following units: The data acquisition and processing unit 1 processes the track three-axis vibration acceleration data collected by the three-axis acceleration sensor to generate a vibration vector sequence, collects the three-phase current waveform of the power supply line through a current sensor, synchronously extracts the harmonic components, and calculates the total harmonic distortion rate and the odd-even harmonic energy ratio according to the harmonic components; The mechanical and electrical coupling feature unit 2 performs three-dimensional vector synthesis on the vibration vector sequence to generate a track deformation energy distribution matrix, maps it to 32 sub-bands through wavelet packet decomposition, calculates the energy proportion of each sub-band, and constructs a time-space correlation matrix of harmonic distortion and vibration according to the energy proportion and the total harmonic distortion rate; In order to convert the triaxial vibration acceleration data into energy features that can reflect track deformation, the mechanical and electrical coupling feature unit 2 needs to perform three-dimensional vector synthesis on the vibration data to generate a track deformation energy distribution matrix, and the specific implementation is as follows: The method for generating a track deformation energy distribution matrix through three-dimensional vector synthesis includes: Since the square value of vibration acceleration is positively correlated with energy, the square values of X, Y, and Z axial vibration acceleration data collected by the triaxial acceleration sensor are calculated, the square values are integrated and accumulated in a sliding window with a time window length of 100 ms and an overlap rate of 50%, each time window contains 100 ms of square value data (assuming the sampling frequency is 1000 Hz, each window contains 100 data points), and the energy component of the window is obtained by accumulating the square values in the window. When the next window slides, the time length of overlap is 50% (i.e. 50 ms), ensuring the continuity of the energy features. Through this operation, the three axial energy components are obtained, the time domain vibration signal is converted into energy features, and the energy mutation caused by track deformation is more easily identified. In order to comprehensively reflect the deformation energy distribution of the track in three-dimensional space, the three axial energy components are superimposed in time sequence. Specifically, a two-dimensional matrix is constructed with time as the row dimension and axis as the column dimension. For example, at a certain time, the X-axis energy component is 0.8g²・ms, the Y-axis is 0.6g²・ms, and the Z-axis is 1.0g²・ms. Then the row vector corresponding to the time point in the matrix is [0.8, 0.6, 1.0]. With the passage of time, the track deformation energy distribution matrix is generated by filling the rows one by one. The number of rows of the matrix corresponds to the number of time windows, and the number of columns is 3, corresponding to the three axes. This realizes the visualization of the time and space features of vibration energy, and the matrix is normalized. Since there may be differences in the acceleration range of different axes, in order to eliminate the dimensional influence and facilitate subsequent analysis, the generated track deformation energy distribution matrix is normalized. The specific method is as follows: calculate the maximum value of all elements in the matrix, divide each element by the maximum value, and map the value range of the matrix elements to the [0, 1] interval. After normalization, the energy features of each axis are comparable, and the accuracy of subsequent wavelet packet decomposition operations is improved.

[0022] In order to separate the key frequency band features representing mechanical faults from the track deformation energy distribution matrix, the mechanical and electrical coupling feature unit 2 needs to perform wavelet packet decomposition on the matrix and eliminate outlier frequency bands, and the specific implementation is as follows: The wavelet packet decomposition method includes: Since wavelet packet decomposition can subdivide the signal in full frequency band, first, select the wavelet base function suitable for mechanical vibration signal analysis, introduce the wavelet base function to the track deformation energy distribution matrix, and generate 32 sub-bands. The first layer decomposition divides each row vector energy signal into low-frequency approximation component and high-frequency detail component, which correspond to the overall trend and local change of the signal respectively. The second to fifth layer decomposition continues to recursively decompose the low-frequency and high-frequency components of each layer until the fifth layer generates 32 sub-bands. For example, the second layer decomposes the two components of the first layer into two components respectively, obtaining four sub-bands, the third layer further decomposes into eight, and so on until the fifth layer forms 32 equal-width frequency bands, each corresponding to a specific frequency range. This decomposition process separates the frequency domain features of the original energy signal layer by layer through multi-resolution analysis. After decomposition, the energy proportion of each sub-band is calculated to measure its importance. The square sum of all energy data points in each sub-band is calculated to obtain the energy value of the frequency band, which is then divided by the total energy of all sub-bands to obtain the energy proportion. Then, the 32 sub-bands are sorted in descending order of energy value, highlighting the main frequency bands with concentrated energy. The sum of the energy proportions of the top 10 high-frequency sub-bands usually exceeds 60%, providing key analysis objects for subsequent fault feature extraction. To eliminate noise or irrelevant frequency band interference, the energy ratio threshold is used to remove outlier frequency bands. For the sorted sub-bands, the energy ratio of each frequency band to the adjacent frequency band is calculated, and the absolute value of the natural logarithm is taken. When the absolute value of the natural logarithm of the energy ratio of a certain sub-band to the adjacent frequency band is greater than 1.5, it is determined as an outlier frequency band and removed. This mechanism can effectively filter out abnormal frequency bands caused by environmental interference or sensor noise, improve the signal-to-noise ratio of subsequent feature extraction, and provide a reliable frequency domain feature basis for mechanical-electrical fault joint diagnosis.

[0023] To construct a feature matrix that can represent the spatiotemporal correlation of mechanical vibration and electrical parameters, the mechanical-electrical coupling feature unit 2 needs to generate a spatiotemporal correlation matrix based on the retained sub-band energy proportion and total harmonic distortion rate through Pearson correlation coefficient analysis and energy entropy weighting. The specific implementation is as follows: The construction method of the spatiotemporal correlation matrix includes: Pearson correlation coefficients are calculated for the reserved sub-band energy proportion and the total harmonic distortion rate respectively, a 32x1 dimensional correlation coefficient vector is generated, the Pearson correlation coefficient is a statistical measure of the degree of linear correlation between two variables, the value range is-1 to 1, the closer to 1, the stronger the positive correlation, the closer to-1, the stronger the negative correlation, and the closer to 0, the weaker the linear correlation, in the present application, the coefficient is used to quantify the degree of linear correlation between the reserved sub-band energy proportion and the total harmonic distortion rate, to reveal the potential relationship between mechanical vibration energy distribution and electrical parameter abnormalities, for each sub-band reserved after wavelet packet decomposition, a total of 32, the Pearson correlation coefficients of the energy proportion sequence and the total harmonic distortion rate sequence are calculated respectively, specifically, taking a sub-band as an example, the energy proportion data and the corresponding total harmonic distortion rate data in multiple time windows are obtained, the correlation coefficient of the sub-band and the total harmonic distortion rate is obtained by calculating the ratio of the covariance and the standard deviation of the two, and finally a 32x1 dimensional correlation coefficient vector is generated, this step uses the prior art to realize the correlation analysis of mechanical-electrical parameters, quantifies the coupling relationship between vibration energy and current harmonics, and provides a correlation feature basis for the subsequent space-time correlation matrix, in order to measure the uncertainty of the energy distribution of each sub-band, the energy entropy of each sub-band is calculated and a weight vector is generated, the calculation method of energy entropy is: for the energy proportion sequence of a sub-band, the probability distribution of each energy proportion value is calculated first, and then the energy entropy of the frequency band is obtained by the weighted sum of the negative logarithm probability, the greater the energy entropy value, the more complex the energy distribution of the frequency band, and the richer the fault information contained, so a higher weight should be given, a 32x1 dimensional weight vector is generated by this method, the weight value range is 0 to 1, so that the sub-band with high energy entropy obtains greater weight, improves the representation ability of fault features, the correlation coefficient vector and the weight vector are point multiplied to generate a space-time correlation matrix, the 32x1 dimensional correlation coefficient vector and the weight vector are point multiplied to obtain the weighted correlation coefficient corresponding to each sub-band, and finally a 32x1 dimensional space-time correlation matrix is generated, each element of the matrix represents the weighted correlation degree of the energy proportion and the total harmonic distortion rate of the corresponding sub-band, which reflects the space-time correlation strength of the mechanical vibration energy in the frequency domain and the electrical parameter abnormalities, which needs to be analyzed in the subsequent fault diagnosis, the space-time correlation matrix fuses the correlation feature and the energy entropy weight, and the frequency domain features of mechanical vibration and the time domain features of electrical parameters are cross-domain correlated, compared with single correlation coefficient analysis, the representation dimension of mechanical-electrical coupling faults is more comprehensive, and structured input data is provided for subsequent bidirectional long short-term memory network feature fusion.

[0024] The feature fusion and fault quantification unit 3 inputs the space-time correlation matrix and the odd-even harmonic energy ratio into a bidirectional long short-term memory network, dynamically allocates feature weights through an attention mechanism, outputs a coupling factor, performs time-frequency domain joint analysis on the coupling factor to extract an intrinsic mode function, calculates the sample entropy of each intrinsic mode function, determines whether there is an electrical parameter anomaly caused by mechanical vibration according to the sample entropy, and outputs a fault probability value according to the determination result using a dynamic time warping matching algorithm.

[0025] In order to extract the time sequence features in the space-time correlation matrix and the odd-even harmonic energy ratio, the feature fusion and fault quantification unit 3 uses a bidirectional long short-term memory network for time sequence modeling, and the specific implementation is as follows: Since the bidirectional long short-term memory network needs a one-dimensional continuous vector as input, the bidirectional long short-term memory network first expands the 32-dimensional space-time correlation matrix into a one-dimensional continuous vector by row, each row of the matrix corresponds to a weighted correlation coefficient of a reserved sub-band, and after expansion, a feature sequence reflecting the correlation strength in the frequency domain is formed, then the sequence is spliced with the odd-even harmonic energy ratio to form a 33-dimensional input vector, the odd-even harmonic energy ratio is a one-dimensional feature, which is the ratio of odd harmonic energy to even harmonic energy, and the odd-even harmonic energy ratio reflects the distortion characteristics of the current waveform (such as insulation aging which will cause the proportion of odd harmonics to rise), and the mechanical vibration frequency domain features of the space-time correlation matrix form a complement, after splicing, the feature fusion of “mechanical frequency domain-electrical time domain” is realized, the dimension design of the 33-dimensional vector takes into account the feature integrity and computational efficiency, compared with directly inputting the matrix (32x1) and separately processing the odd-even harmonic, the feature interaction is improved, the 33-dimensional input vector sequence is processed in time sequence, and the forward hidden state calculation logic of each time step is as follows: The forward hidden state is a zero vector with a dimension of 64, and the forward hidden state is calculated step by step in time sequence, and the state of each time step is determined by the hidden output of the previous time step and the current input vector. For the input vector of the tth time step, the forward hidden state is determined by the forward hidden output of the previous time step (t-1) and the current input vector. Specifically, the information is filtered and updated through the forget gate, the input gate and the output gate in the network, the historical information related to the fault feature is retained, and the irrelevant noise is forgotten. The output of the forward hidden state of each time step is a 64-dimensional vector, which ensures the consistency of the dimension of the time sequence feature. The input sequence is processed in time reverse order, and the influence of the future time step on the current state is captured. The specific steps are as follows: starting from the last time step, the initial backward hidden state is also a zero vector with a dimension of 64, and the backward hidden state of the tth time step is determined by the backward hidden output of the next time step (t+1) and the current input vector. The output of the backward hidden state of each time step is a 64-dimensional vector, which matches the dimension of the forward hidden state. The backward hidden state is calculated step by step in time reverse order, and the state of each time step is determined by the hidden output of the next time step and the current input vector. Finally, the forward and backward hidden states of the same time step are spliced into a 128-dimensional fusion vector. The forward hidden state and the backward hidden state of the same time step are spliced along the dimension direction to generate a 128-dimensional fusion vector. The vector contains the mechanical-electrical correlation features of the past time steps and the trend features of the future time steps, for example: the track deformation frequency band feature in the forward hidden state and the odd-even harmonic rising trend in the backward hidden state can be combined to predict the development direction of the contact failure. The splicing operation retains the independence of the bidirectional time sequence feature, and compared with the simple weighted sum, the retention rate of the time sequence evolution details of the fault feature is improved. The bidirectional long short-term memory network processing process forms a technical closed loop with the time-space correlation matrix construction technology described above: through bidirectional time sequence feature extraction, the cross-domain correlation features of mechanical vibration and electrical parameters are converted into fusion vectors containing time context, which provides rich time sequence information for the subsequent attention mechanism feature weight allocation and lays a foundation for the accurate output of the coupling factor.

[0026] In order to filter the key historical features from the 128-dimensional fusion vector output by the bidirectional long short-term memory network, the feature fusion and fault quantization unit 3 adopts an attention mechanism to dynamically allocate feature weights, and generates a coupling factor representing the degree of mechanical-electrical coupling. The specific implementation is as follows: Firstly, the length of the historical time window is determined, and the attention mechanism extracts the 128-dimensional fusion vectors of all time steps in the historical time window to form a historical feature set. The historical feature set serves as the memory bank of the attention mechanism, storing historical fault clues related to the current state, providing a data basis for the correlation strength calculation. Through a trainable weight matrix, the current fusion vector is mapped with each historical fusion vector. Specifically, the current vector and the historical vector are concatenated into a 256-dimensional vector by column, and the weight matrix is multiplied to obtain the original score of the correlation strength between each historical vector and the current vector. The original score is input into the hyperbolic tangent function for activation. The output range of the function is -1 to 1, which can map the score to a symmetric interval, enhancing the ability to distinguish weakly correlated features., the physical meaning of the activation process is that when the original score is positive, the larger the activated value indicates that the positive correlation between the current feature and the historical feature is stronger, when the original score is negative, the smaller the activated value indicates that the negative correlation is stronger, the non-linear characteristics of the function can amplify the difference of strong correlation scores, compress the range of weak correlation scores, and make the weight distribution of key features more prominent, the score is converted into a probability distribution weight through the exponential function and summation operation, and then the historical vector is weighted and summed according to the weight to generate a context vector representing the key historical features, the correlation strength score after activation is converted into a non-negative value through an exponential function, and then the exponential scores of all historical vectors are summed, and each exponential score is divided by the sum to obtain a probability distribution weight, which reflects the contribution ratio of each historical vector to the current state, the value range is 0 to 1, and the sum of all weights is 1, the historical fusion vector is weighted and summed according to the probability distribution weight to obtain a 64-dimensional context vector, which aggregates the key features in the historical time window, for example: if the 100-200Hz frequency band strongly correlated feature corresponding to the track deformation appears multiple times in the historical time window, the context vector will strengthen the feature representation of the frequency band, and the weighted summation operation is equivalent to selective memory of historical features, features with high weights are highlighted, and features with low weights are weakened, finally the current fusion vector and the context vector are spliced through an S-shaped growth curve function to output a coupling factor in the range of 0 to 1, the current fusion vector (128-dimensional) and the context vector (64-dimensional) are spliced into a 192-dimensional vector along the dimension direction, mapped to 1-dimensional through a fully connected layer, and then input into an S-shaped growth curve function (logical sigmoid function), the output value is compressed to the range of 0 to 1 to generate the final coupling factor, the physical meaning of the factor is that the closer the value is to 1, the higher the coupling degree of mechanical vibration and electrical parameters, and the more likely there is an associated fault, the closer the value is to 0, the lower the coupling degree, and the system is in normal operation, the processing flow of the attention mechanism is linked with the bidirectional long short-term memory network described above: by dynamically weighting the historical time series features, the key clues of mechanical-electrical coupling faults are highlighted, the coupling factor is more sensitive to the coupling faults in the 100-200Hz frequency band, and more accurate input features are provided for subsequent time-frequency domain analysis and fault probability calculation.

[0027] In order to extract the essential features reflecting the mechanical-electrical coupling faults from the coupling factor, the feature fusion and fault quantization unit 3 needs to perform time-frequency domain joint analysis on the coupling factor, extract the eigenmode function and generate the time-frequency energy distribution diagram, and the specific implementation is as follows: Since the extreme point distribution of the coupling factor reflects the fluctuation characteristics of the signal, the time-frequency domain joint analysis first identifies all the maximum and minimum points of the time series of the coupling factor. Then, the maximum points are connected to form the upper envelope line and the minimum points are connected to form the lower envelope line by using cubic spline interpolation. The maximum points are fitted by a cubic spline interpolation function to form a smooth upper envelope line, which reflects the upper boundary fluctuation trend of the signal. Similarly, the minimum points are fitted by a cubic spline interpolation to generate a lower envelope line, which represents the lower boundary fluctuation characteristics of the signal. The smoothness of the cubic spline interpolation can avoid the jagged fluctuation of the envelope line. Compared with linear interpolation, the fitting error of the envelope line to the true boundary of the signal is reduced, providing an accurate basis for subsequent mean envelope calculation. The arithmetic mean of the upper and lower envelope lines is calculated as the mean envelope, and the signal after the initial screening is obtained by subtracting the mean envelope from the original coupling factor. To ensure that the screened signal meets the conditions of the intrinsic mode function (the difference between the number of extreme points and the number of zero-crossing points is not more than 1, and the envelope mean of any point is zero), repeated screening is required until the conditions of the difference between the number of extreme points and the number of zero-crossing points not more than 1 and the envelope mean of any point being zero are met to extract the intrinsic mode function. This iterative process can effectively separate different frequency components in the signal, making the physical meaning of the intrinsic mode function more explicit. The extraction of the intrinsic mode function uses an improved empirical mode decomposition method, which ensures that each mode function reflects the vibration characteristics of a single frequency component. Unlike the traditional empirical mode decomposition method, this method introduces an energy threshold constraint based on the mechanical-electrical coupling characteristics of the coupling factor during the screening process: when the energy decay of the signal in a certain screening step exceeds 30% of the current total energy, the screening is forced to stop to avoid mode aliasing caused by excessive decomposition. Finally, each intrinsic mode function is subjected to an orthogonal transformation and the instantaneous phase angle change rate is calculated as the instantaneous frequency. To eliminate the correlation between the intrinsic mode functions and improve the accuracy of time-frequency analysis, each intrinsic mode function is subjected to an orthogonal transformation. The specific steps are as follows: sort the intrinsic mode functions by energy proportion from high to low, and process the mode with high energy first, which corresponds to the main fault characteristics. For the i-th mode function, calculate its inner product with the first i-1 orthogonalized mode functions, and eliminate the correlation by linear combination. After orthogonalization, each mode function is amplitude-normalized to ensure that the energy scales of the modes are consistent. Compared with the traditional orthogonalization, this method introduces an adaptive step adjustment (the step changes dynamically with the energy proportion of the mode) to address the non-stationary characteristics of the mechanical-electrical coupling signal, reducing the correlation coefficient between the orthogonalized mode functions from 0.3 to 0.1Below, the independence of time-frequency characteristics is significantly improved, and the time-frequency energy distribution diagram is generated by superimposing the instantaneous frequency of all modal functions. For each intrinsic modal function after orthogonalization, the rate of change of its instantaneous phase angle is calculated as the instantaneous frequency. Specifically, the instantaneous phase of the modal function is obtained by Hilbert transform, and the derivative of the phase is obtained to obtain the instantaneous frequency. This frequency value reflects the local vibration frequency of the modal function at each time point. Superimpose the instantaneous frequencies of all intrinsic modal functions in time sequence to generate a time-frequency energy distribution diagram. The horizontal axis of the diagram is time, the vertical axis is frequency, and the color depth represents energy intensity. The energy is determined by the square of the amplitude of the modal function. For example, in the 100-200Hz frequency band of a certain time period, there is an energy concentration area, the color is darker, indicating that there is a characteristic frequency component of mechanical-electrical coupling fault in this period. This distribution diagram visualizes the time-frequency characteristics of the coupling factor. Compared with single spectrum analysis, it has improved ability to present the time-frequency evolution process of non-steady-state faults, and provides intuitive feature basis for subsequent sample entropy determination and fault probability calculation.

[0028] In order to accurately determine whether mechanical vibration causes electrical parameter anomaly, the feature fusion and fault quantification unit 3 needs to make joint determination based on the sample entropy value of the third intrinsic modal function, the time-frequency energy integral and the kurtosis statistic quantity. The specific implementation is as follows: Since the third eigenmodal function usually corresponds to the characteristic frequency component of the mechanical-electrical coupling fault, the sample entropy determination first constructs a vector sequence from the third eigenmodal function by taking a length of 3 continuous subsequences with a time delay of 5 sampling points, this delay sampling method avoids the high correlation of adjacent subsequences, compared with zero delay sampling, the independence of sample entropy calculation is improved, the distance between all vector pairs in the calculation vector sequence is calculated, the number of matches in which the distance between all vector pairs is less than 0.2 times the standard deviation of the sequence is counted, the vector distance is calculated by the Euclidean distance, that is, the square sum of each dimension difference is taken, and the standard deviation is the standard deviation of the entire vector sequence, then, based on the matching ratio (the ratio of the number of matches to the total number of vector pairs), the natural logarithm is operated, and the sample entropy value is output, the larger the sample entropy value, the higher the complexity and irregularity of the signal, the more likely there is a fault, based on the matching ratio, the sample entropy value is output by natural logarithm operation, when the sample entropy value is greater than 1.2 for 3 seconds, it indicates that the signal complexity of the third eigenmodal function abnormally increases, the duration of 3 seconds is set based on the development cycle of the typical fault of the crane, for example, the evolution from slight to serious of the track deformation usually exceeds 3 seconds, to avoid misjudgment caused by transient interference, 1.2 threshold value is determined by analyzing 100 groups of normal working condition and fault working condition data, the energy integral value of the frequency band is obtained by integrating the time-frequency energy of the 100-200 Hz frequency band in the time-frequency energy distribution diagram, the integral method is: summing the energy value of the 100-200 Hz frequency band at each time point on the time axis, and then dividing by the time length to obtain the average energy integral value, when the value exceeds 0.35, it indicates that the mechanical vibration energy is abnormally concentrated in the fault characteristic frequency band, the threshold value of 0.35 is determined by calibration experiment, the kurtosis statistic is used to measure the steepness of the signal amplitude distribution, the ratio of the fourth central moment to the square of the variance of all data points in the third eigenmodal function is calculated, when the kurtosis statistic of the third eigenmodal function is greater than 4.5, it indicates that the signal has a significant impact component (the kurtosis of the normal signal is close to 3), it is determined that there is a mechanical vibration causing electrical abnormality, when the sample entropy value is greater than 1.2 for 3 seconds, the time-frequency energy integral of the 100-200 Hz frequency band is greater than 0.35, and the kurtosis of the third eigenmodal function is greater than 4.5, the three conditions are met at the same time, it is determined that there is an electrical parameter abnormality caused by mechanical vibration, this joint determination mechanism avoids the misjudgment risk of a single feature, the sample entropy value reflects the signal complexity, which excludes the interference of stationary noise, the time-frequency energy integral locates the fault characteristic frequency band, which ensures that the abnormality is related to mechanical vibration, the kurtosis statistic captures the impact component, which verifies the mutation characteristics of the electrical parameter, the sample entropy joint determination process forms a technical closed loop with the time-frequency domain analysis in the previous section: through the multi-dimensional analysis of the complexity, frequency energy distribution and amplitude characteristics of the third eigenmodal function, the coupling relationship between mechanical vibration and electrical abnormality is quantified, which provides reliable abnormality determination basis for subsequent fault probability calculation.

[0029] In order to convert the mechanical and electrical coupling features into specific failure probability values, the feature fusion and failure quantification unit 3 needs to calculate and output the failure probability based on the preset physical constraint template through a dynamic programming algorithm and an exponential decay function. The specific implementation is as follows: The failure probability value generation first presets a physical constraint template, which is as follows: The track deformation template is that the coupling factor monotonically increases and the change rate is continuously higher than 0.3. The monotonically increasing coupling factor reflects the development process of the track deformation from slight to severe, and the change rate higher than 0.3 indicates that the deformation speed is relatively fast. The physical basis of this template is that the track deformation will cause the vibration energy of the crane during operation to gradually increase, and then the mechanical and electrical coupling degree will continuously increase. The monotonically increasing characteristics of the coupling factor are highly consistent with the development law of the track deformation. The insulation aging template is that the sample entropy is less than 0.8 and the total harmonic distortion rate is higher than 18%. The sample entropy less than 0.8 indicates that the signal complexity is low (i.e. the current waveform tends to be regularized), and the total harmonic distortion rate higher than 18% reflects the abnormal increase of the current harmonic content. The physical meaning of this template is that insulation aging will cause the line impedance to change, which will increase the odd harmonic components in the current waveform and reduce the irregularity of the signal. This template has been verified by 100 groups of insulation aging failure data. The poor contact template is that the peak value of the energy integral in the 50-150Hz frequency band is more than 0.4. This frequency band corresponds to the characteristic frequency of the crane contact resistance fluctuation (such as the mechanical vibration frequency caused by the intermittent disconnection of the contact point). The energy integral peak value more than 0.4 indicates that there is significant energy concentration in this frequency band. Physically, poor contact will cause mechanical impact during current on-off, which will cause a sudden change in vibration energy in the 50-150Hz frequency band. The dynamic programming algorithm is used to calculate the minimum cumulative alignment distance of the to-be-tested sequence and each template, the distance is mapped to a fault probability value of 0 to 1 through an exponential decay function, when the highest fault probability exceeds 0.8 and the abnormal judgment condition of claim 8 is met, the corresponding fault type and probability are output, the to-be-tested characteristic sequence (such as the coupling factor sequence of the current 30 seconds) is unified with the template sequence (the preset fault characteristic sequence) in the time dimension, the number of time points of the to-be-tested sequence and the template sequence is adjusted to be consistent through linear interpolation, ensuring the feasibility of subsequent alignment calculation, a two-dimensional dynamic programming matrix is constructed, the rows of the matrix correspond to the time points of the template sequence, the columns correspond to the time points of the to-be-tested sequence, and the matrix elements represent the cumulative alignment distance of the first i time points of the template sequence and the first j time points of the to-be-tested sequence. At the beginning, the left upper corner element of the matrix is zero, and the rest of the boundary elements are infinite. In the iterative calculation process, for each matrix element, the distance between the current template point and the to-be-tested point is calculated, and the minimum value of the left, upper or upper-left element is added to the cumulative distance of the current element. After iteration, the right lower corner element of the matrix is the minimum cumulative alignment distance of the template sequence and the to-be-tested sequence. The distance reflects the matching degree of the to-be-tested characteristic and the template characteristic. The smaller the distance, the higher the matching degree. The minimum cumulative distance is mapped to a fault probability value of 0 to 1 through an exponential decay function, and the formula is , wherein is the decay coefficient. The exponential decay characteristic of the function makes the small change of the distance significantly affect the probability value, improving the fault discrimination. The fault probability values of the to-be-tested sequence are calculated for the three preset templates, and the track deformation probability, insulation aging probability and contact poor probability are obtained. When the highest fault probability exceeds 0.8 and the abnormal condition of the sample entropy determination (the sample entropy is greater than 1.2 for 3 seconds, the 100-200Hz energy integral exceeds 0.35, and the kurtosis is greater than 4.5) is met, the corresponding fault type and probability are output. When the effective diagnosis times of a certain fault type exceed 50 times, the system automatically updates the corresponding template with new typical fault data. The update method is: calculate the mean of the characteristic sequences of the fault type in the 50 diagnoses as the new template sequence. The fault probability value generation process forms a technical closed loop with the sample entropy determination in the foregoing. Through the physical constraint template, the abstract mechanical-electrical characteristics are converted into specific fault probabilities. Combined with the abnormal determination condition of the sample entropy, the diagnosis result of the crane circuit fault has clear engineering guiding significance. Compared with the traditional threshold alarm, the quantitative accuracy of the fault probability output is improved, providing a more accurate basis for equipment maintenance decision.

[0030] The application obtains track three-axis vibration acceleration and three-phase current waveform through a data acquisition processing unit 1, performs three-dimensional vector synthesis and wavelet packet decomposition on vibration data through a mechanical and electrical coupling characteristic unit 2, constructs a space-time correlation matrix of harmonic distortion and vibration, and outputs a coupling factor through a feature fusion and fault quantization unit 3 using a bidirectional long short-term memory network and an attention mechanism, and outputs a fault probability value through a dynamic time warping matching algorithm after time-frequency domain analysis and sample entropy determination, thereby realizing space-time correlation modeling and dynamic fault matching of multi-source data and improving fault positioning accuracy and diagnosis accuracy.

[0031] Please refer to Figure 2 The second object of the application is to provide a method for realizing a crane line fault diagnosis system based on multi-source data fusion, comprising the following steps: S1, collect track three-axis vibration acceleration and three-phase current waveform, calculate the square value of axial acceleration of vibration data and generate three-axis energy components by integrating 100ms time window and 50% overlap rate, and extract harmonic components of current data to calculate total harmonic distortion rate and odd-even harmonic energy ratio; S2, superimpose three-axis energy components to generate track deformation energy distribution matrix, perform 5-layer decomposition using wavelet basis function to obtain 32 sub-bands, remove outlier bands with natural logarithm absolute value of adjacent band energy ratio greater than 1.5, calculate the Pearson correlation coefficient vector of the energy proportion of the reserved sub-bands and the total harmonic distortion rate, and generate a space-time correlation matrix by weighting the band energy entropy; S3, expand the space-time correlation matrix and splice it with the odd-even harmonic energy ratio to form a 33-dimensional input vector, extract forward / backward time sequence features through a bidirectional long short-term memory network, output a 128-dimensional fusion vector, and generate a coupling factor based on an attention mechanism dynamic weighting history vector, extract the first five intrinsic mode functions of the coupling factor empirical mode decomposition, calculate the sample entropy value of the third intrinsic mode function, and determine the abnormality by combining the time-frequency energy integral greater than 0.35 and the kurtosis greater than 4.5 in the 100-200Hz frequency band; S4, define a physical template, and calculate the minimum cumulative alignment distance using dynamic time warping, map it to a fault probability through an exponential decay function, and output the fault type and update the template library when the highest probability is greater than 0.8.

[0032] The above shows and describes the basic principles, main features and advantages of the application. Those skilled in the art should understand that the application is not limited to the above examples, and the above examples and descriptions in the specification are only preferred examples of the application and are not intended to limit the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection of the application is defined by the appended claims and their equivalents.

Claims

1. The crane line fault diagnosis system based on multi-source data fusion is characterized by: The following units are included: The data acquisition and processing unit (1) processes the vibration acceleration data of the three axes of the track collected by the three-axis acceleration sensor to generate a vibration vector sequence, collects the three-phase current waveform of the power supply line through the current sensor, synchronously extracts its harmonic components, and calculates the total harmonic distortion rate and the odd-even harmonic energy ratio based on the harmonic components; The mechanical and electrical coupling characteristic unit (2) performs three-dimensional vector synthesis on the vibration vector sequence to generate the track deformation energy distribution matrix, and maps it to 32 sub-bands through wavelet packet decomposition. The energy proportion of each sub-band is then calculated, and the spatiotemporal correlation matrix of harmonic distortion and vibration is constructed based on the energy proportion and total harmonic distortion rate. The feature fusion and fault quantification unit (3) inputs the spatiotemporal correlation matrix and the odd-even harmonic energy ratio into the bidirectional long short-term memory network, dynamically allocates feature weights through the attention mechanism, outputs the coupling factor, performs a joint time-frequency domain analysis on the coupling factor to extract the intrinsic mode function, and calculates the sample entropy of each intrinsic mode function. Based on the sample entropy, it is determined whether there is an electrical parameter anomaly caused by mechanical vibration, and the dynamic time warping matching algorithm is used to output the fault probability value based on the determination result.

2. The crane line fault diagnosis system based on multi-source data fusion according to claim 1 is characterized in that: The method for generating a track deformation energy distribution matrix by synthesizing three-dimensional vectors includes: The X, Y, and Z axial vibration acceleration data collected by the three-axis accelerometer are used to calculate the square value of each axial acceleration respectively. The square values ​​are integrated and accumulated according to the time window length of 100ms and the overlap rate of 50% to obtain the three-axial energy components. The three-axial energy components are superimposed in time series to generate the track deformation energy distribution matrix, and the matrix is ​​normalized.

3. The crane line fault diagnosis system based on multi-source data fusion according to claim 1 is characterized in that: The wavelet packet decomposition method comprises: Wavelet basis functions were introduced to perform a five-layer decomposition of the orbital deformation energy distribution matrix to generate 32 sub-bands. The energy proportion of each sub-band was calculated and the sub-bands were sorted from high to low according to energy value. When the absolute value of the natural logarithm of the ratio of the energy of a sub-band to the energy of the adjacent band was greater than 1.5, it was determined to be an outlier band and was removed.

4. The crane line fault diagnosis system based on multi-source data fusion according to claim 3 is characterized in that: The method for constructing the spatiotemporal correlation matrix includes: The Pearson correlation coefficient is calculated for the energy proportion of the retained sub-bands and the total harmonic distortion rate respectively to generate a 32×1 dimensional correlation coefficient vector. The weight vector is calculated according to the energy entropy of each sub-band. The correlation coefficient vector and the weight vector are multiplied to generate a spatiotemporal correlation matrix.

5. The crane line fault diagnosis system based on multi-source data fusion according to claim 4 is characterized in that: The bidirectional long short-term memory network first expands the 32-dimensional spatiotemporal correlation matrix into continuous vectors by row and splices it into a 33-dimensional input vector with the odd-even harmonic energy ratio, and then calculates the forward hidden state step by step in chronological order, where the state of each time step is determined by the hidden output of the previous time step and the current input vector. At the same time, the backward hidden state is calculated step by step in reverse chronological order, where the state of each time step is determined by the hidden output of the next time step and the current input vector, and finally the forward and backward hidden states of the same time step are spliced ​​into a 128-dimensional fusion vector.

6. The crane line fault diagnosis system based on multi-source data fusion according to claim 5 is characterized in that: The attention mechanism first extracts the 128-dimensional fusion vector of all time steps in the historical time window, calculates the correlation strength score between the current fusion vector and each historical vector through the weight matrix and activates it with the hyperbolic tangent function, converts the score into a probability distribution weight through the exponential function and summation operation, and then generates a context vector representing the key historical features by weighted summation of the historical vectors according to the weight. Finally, the current fusion vector is spliced ​​with the context vector and the coupling factor in the range of 0 to 1 is output through the S-shaped growth curve function.

7. The crane line fault diagnosis system based on multi-source data fusion according to claim 6 is characterized in that: The time-frequency domain joint analysis first identifies all maximum and minimum points of the coupling factor time series, uses cubic spline interpolation to connect the maximum points to form an upper envelope and the minimum points to form a lower envelope, calculates the point-by-point arithmetic mean of the upper and lower envelopes as the mean envelope, subtracts the mean envelope from the original coupling factor, and repeatedly screens until the difference between the number of extreme points and the number of zero-crossing points does not exceed 1 and the mean envelope of any point is zero, thereby extracting the intrinsic mode function. Finally, each intrinsic mode function is orthogonalized and the instantaneous phase angle change rate is calculated as the instantaneous frequency. The instantaneous frequencies of all mode functions are superimposed to generate a time-frequency energy distribution diagram.

8. The crane line fault diagnosis system based on multi-source data fusion according to claim 7 is characterized in that: The sample entropy determination first constructs a vector sequence by extracting a continuous subsequence of length 3 from the third intrinsic mode function according to the time delay of 5 sampling points, counts the number of matches among all vector pairs whose distance is less than 0.2 times the sequence standard deviation, and outputs the sample entropy value based on the matching ratio through natural logarithm operation; When the sample entropy value is greater than 1.2 for 3 seconds, the time-frequency energy integral value in the 100-200 Hz frequency band exceeds 0.35, and the kurtosis statistic of the third eigenmode function is greater than 4.5, it is determined that mechanical vibration has caused electrical anomalies.

9. The crane line fault diagnosis system based on multi-source data fusion according to claim 7 is characterized in that: The generation of the fault probability value first presets a physical constraint template, which is as follows: The track deformation template is: the coupling factor increases monotonically and the rate of change is continuously higher than 0.3; The insulation aging template is: sample entropy is less than 0.8 and total harmonic distortion is higher than 18%; The poor contact template is: the peak value of the time-frequency energy integral in the 50-150Hz frequency band exceeds 0.4; A dynamic programming algorithm is used to calculate the minimum cumulative alignment distance between the sequence to be tested and each template, and the distance is mapped to a fault probability value from 0 to 1 through an exponential decay function. When the highest fault probability exceeds 0.8 and meets the abnormality judgment condition of claim 8, the corresponding fault type and probability are output.

10. A method for implementing a crane line fault diagnosis system based on multi-source data fusion according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Collect the three-axis vibration acceleration and three-phase current waveforms of the track. Calculate the square value of the axial acceleration from the vibration data and integrate it with a 100ms time window and 50% overlap rate to generate the three-axial energy components. Extract the harmonic components from the current data to calculate the total harmonic distortion rate and the odd-even harmonic energy ratio. S2. Superimpose the three axial energy components to generate the orbital deformation energy distribution matrix. Use wavelet basis functions to perform a five-level decomposition to obtain 32 sub-bands. Remove outlier bands whose energy ratios of adjacent bands are greater than 1.

5. Calculate the Pearson correlation coefficient vector between the energy proportions of the retained sub-bands and the total harmonic distortion rate. Generate a spatiotemporal correlation matrix using the band energy entropy weighting. S3. Expand the spatiotemporal correlation matrix and concatenate it with the odd-even harmonic energy ratio to form a 33-dimensional input vector. Use a bidirectional long short-term memory network to extract forward and backward temporal features, outputting a 128-dimensional fusion vector. Dynamically weight the history vector based on the attention mechanism to generate a coupling factor. Perform empirical mode decomposition on the coupling factor to extract the first five intrinsic mode functions (IMFs). Calculate the sample entropy of the third IMF. Determine an anomaly based on a time-frequency energy integral greater than 0.35 and a kurtosis greater than 4.5 in the 100-200 Hz frequency band. S4. Define the physical template and use dynamic time warping to calculate the minimum cumulative alignment distance, which is mapped to the fault probability through an exponential decay function. When the highest probability is greater than 0.8, the fault type is output and the template library is updated.

Images