Relay state monitoring and diagnosis method and system

Abstract

The invention belongs to the technical field of equipment state monitoring, and discloses a relay state monitoring and diagnosis method and system. Vibration time domain signals of all monitoring points in the working process of a relay are collected, a time delay characteristic matrix is constructed through segmented cross-correlation operation, and a near-field sound source candidate area is analyzed and recognized in combination with the spatial divergence of a time delay gradient vector field. A sound-vibration coupling propagation model is introduced, near-field sound source excitation and structure conduction vibration components are distinguished, multi-component signal decoupling is realized through spatial constraint and frequency domain sparsity constraint, and a relay body vibration signal is effectively separated. The method comprises the following steps: performing time-frequency decomposition on a body vibration signal, extracting a transient impact feature of a contact action and a steady-state vibration feature of coil excitation, constructing a multi-dimensional state feature vector, and realizing state diagnosis through similarity matching with a preset feature library. According to the invention, the early fault detection sensitivity and diagnosis accuracy of the relay are improved, and a reliable basis is provided for preventive maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment state monitoring, more particularly, the present application relates to a relay state monitoring and diagnosis method and system. BACKGROUND

[0002] The existing relay state monitoring and diagnosis technology has limitations in the face of complex industrial environments, especially in the aspects of near-field sound source isolation and structural conduction vibration decoupling. In industrial sites, relays usually work inside high-density power distribution cabinets or automation control systems, and the mechanical vibration, electromagnetic interference and environmental noise of surrounding equipment are continuously superimposed on the relay body signal, forming a complex vibration field with multiple sources mixed together. Traditional monitoring methods rely only on single-point or a small number of sensors to collect mixed vibration signals, and lack effective mathematical models to separate the near-field impact vibration caused by contact action from the structural conduction vibration caused by coil excitation, resulting in unclear signal interpretation. Especially in high-frequency relay systems, the transient impact caused by contact bounce and the sustained vibration caused by coil electromagnetic force overlap in time and frequency domains, and traditional frequency spectrum analysis methods cannot distinguish their respective characteristics. In addition, during the propagation of vibration on the relay shell, the structural conduction of different frequency bands shows obvious frequency dispersion phenomenon, that is, high-frequency components attenuate quickly and low-frequency components propagate far, but the existing technology lacks a systematic modeling of this frequency-dependent attenuation law, and cannot accurately depict the difference characteristics of the vibration propagation path. In actual applications, such as protection relays in power systems, misjudgment of faults can lead to serious safety accidents and economic losses; and in intelligent manufacturing production lines, relay failure often becomes a key bottleneck for system downtime. However, due to the inability to effectively isolate near-field sound sources and decouple structural conduction vibration, existing monitoring systems have low detection sensitivity and high misjudgment rate when facing early faults and compound faults, especially gradual faults such as increased contact roughness and slightly reduced spring elasticity, which are difficult to detect in a timely manner.

[0003] In view of this, the present application proposes a relay state monitoring and diagnosis method and system to solve the above problems. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a relay state monitoring and diagnosis method, comprising:

[0005] Collecting vibration time domain signals of each monitoring point during the working process of the relay, and recording the spatial coordinates of each monitoring point;

[0006] Performing segmented cross-correlation operation on the vibration time domain signals to extract the vibration wave arrival time delay between any two monitoring points, and constructing a time delay feature matrix between the monitoring points;

[0007] a time delay gradient vector of each monitoring point in the time delay feature matrix is calculated, and a near-field sound source candidate area on the relay shell is identified according to divergence distribution of the time delay gradient vector in space;

[0008] a structural conduction path length between the near-field sound source candidate area and each monitoring point is obtained, and a frequency-dependent attenuation coefficient of the structural conduction path under different frequency bands is calculated in combination with a spectral energy distribution of the vibration time domain signal;

[0009] a sound-vibration coupling propagation model is constructed based on the frequency-dependent attenuation coefficient and the time delay feature matrix, and the sound-vibration coupling propagation model includes a difference feature of a transfer path of a near-field sound source excitation component and a structural conduction vibration component;

[0010] a multi-component decoupling objective function is established by introducing a spatial constraint condition and a frequency domain sparsity constraint condition into the sound-vibration coupling propagation model, and a relay body vibration signal is separated by solving the multi-component decoupling objective function using an iterative optimization algorithm;

[0011] a multi-dimensional state feature vector is constructed by performing time-frequency decomposition on the relay body vibration signal and extracting transient impact features in a contact action stage and steady vibration features in a coil excitation stage;

[0012] the multi-dimensional state feature vector is matched with a preset relay state feature library, a current operating state of the relay is determined according to a matching result, and a state diagnosis conclusion is output.

[0013] A relay state monitoring and diagnosis system is used to implement a relay state monitoring and diagnosis method, and includes:

[0014] a data acquisition module configured to acquire vibration time domain signals of each monitoring point during relay operation and record spatial coordinates of each monitoring point;

[0015] a matrix construction module configured to perform segmented cross-correlation operation on the vibration time domain signals, extract vibration wave arrival time delays between any two monitoring points, and construct a time delay feature matrix between the monitoring points;

[0016] a sound source identification module configured to calculate a time delay gradient vector of each monitoring point in the time delay feature matrix, and identify a near-field sound source candidate area on the relay shell according to divergence distribution of the time delay gradient vector in space;

[0017] an attenuation coefficient calculation module configured to obtain a structural conduction path length between the near-field sound source candidate area and each monitoring point, and calculate a frequency-dependent attenuation coefficient of the structural conduction path under different frequency bands in combination with a spectral energy distribution of the vibration time domain signal;

[0018] A coupling model construction module is configured to construct a sound-vibration coupling propagation model based on the frequency-dependent attenuation coefficient and the time delay feature matrix.

[0019] A signal decoupling module is configured to introduce a spatial constraint condition and a frequency domain sparsity constraint condition into the sound-vibration coupling propagation model, establish a multi-component decoupling objective function, and solve the multi-component decoupling objective function by using an iterative optimization algorithm to separate the relay body vibration signal.

[0020] A feature extraction module is configured to perform time-frequency decomposition on the relay body vibration signal, extract transient impact features in the contact action stage and steady-state vibration features in the coil excitation stage, and construct a multi-dimensional state feature vector.

[0021] A state diagnosis module is configured to perform similarity matching between the multi-dimensional state feature vector and a preset relay state feature library, determine the current operating state of the relay according to the matching result, and output a state diagnosis conclusion.

[0022] The technical effects and advantages of the relay state monitoring and diagnosis method and system of the present application are as follows:

[0023] The present application improves the perception ability of the monitoring system to the early degradation state, so that the maintenance personnel can timely find potential problems before the fault develops to the critical stage, avoiding production line downtime and system paralysis caused by sudden failure of the relay. In the power system, the present application effectively prevents safety accidents and chain failures caused by protection failure, greatly reduces the operation risk and economic loss. The present application has very high detection sensitivity to gradual faults such as contact micro-wear and spring slight fatigue, making it possible to truly implement preventive maintenance strategy. In the intelligent manufacturing environment, the present application reduces unnecessary maintenance intervention and unplanned downtime, while avoiding premature replacement of relay components that are still in good condition, achieving optimal allocation of maintenance resources. For a large number of relay groups in large-scale automation systems, the state evaluation and life prediction functions of the present application greatly improve the overall system reliability, and the operation and maintenance team can develop a scientific replacement strategy based on the diagnosis results to balance the immediate maintenance cost and long-term operation risk. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A schematic diagram of a relay state monitoring and diagnosis method of the present application;

[0025] Figure 2 A schematic diagram of a relay state monitoring and diagnosis system of the present application. DETAILED DESCRIPTION

[0026] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0027] The examples of the present application provide a relay state monitoring and diagnosis method and system. The execution subject of the system includes but is not limited to the relay state monitoring system, the vibration analysis platform, the industrial equipment diagnosis center, the intelligent maintenance system and the like which can be regarded as the general computing nodes of the present application, and the diagnosis system includes but is not limited to at least one of the vibration-based state monitoring system, the relay life prediction system and the multi-sensor fusion diagnosis platform.

[0028] Please refer to Figure 1 In the embodiments of the present application, a specific implementation of a relay state monitoring and diagnosis method includes:

[0029] Vibration time domain signals of each monitoring point in the working process of the relay are collected, and spatial coordinates of each monitoring point are recorded. The vibration time domain signals are acquired in real time by a plurality of acceleration sensors arranged on the surface of the relay shell, including complete vibration responses of key working stages such as relay contact action, coil excitation and mechanical transmission. The arrangement of the monitoring points follows the key position coverage and signal integrity principle, and usually includes important positions such as the area near the contact, the area around the coil and the surface of the mechanical connecting components. The spatial coordinates are recorded by using a three-dimensional rectangular coordinate system, a reference coordinate system is established with a fixed point of the relay shell as the origin, and the accurate positioning of the spatial position is ensured, thereby providing a spatial basis for subsequent time delay analysis and sound source identification.

[0030] The vibration time domain signals are subjected to segmented cross-correlation operation, vibration wave arrival time delays between any two monitoring points are extracted, and a time delay feature matrix between the monitoring points is constructed. The segmented cross-correlation operation is a key step for extracting the time delay information, and the propagation time delay is determined by comparing the phase difference of the vibration signals of different monitoring points. The calculation process first divides the signals into a plurality of time periods according to the action period of the relay, then calculates the normalized cross-correlation function between each pair of monitoring points, extracts the peak position as the initial time delay estimation, and then obtains the time delay value with sub-sampling precision through interpolation fitting. The time delay feature matrix is a key data structure for describing the vibration propagation law, and each element in the matrix corresponds to the accurate time delay between a pair of monitoring points, thereby providing basic data for subsequent gradient analysis.

[0031] The time delay gradient vector of each monitoring point in the time delay feature matrix is calculated, and a near-field sound source candidate area on the relay shell is identified according to the divergence distribution of the time delay gradient vector in space. The time delay gradient vector reflects the direction and speed of the vibration wave front propagation, and is an important clue for identifying the vibration source. The calculation process first selects a reference point, extracts the time delay relationship with other points, then obtains the gradient vector through spatial gradient fitting, and further calculates the divergence distribution of the gradient field. The area with large divergence value usually corresponds to the position of the vibration source, and by setting a proper threshold, the near-field sound source candidate area can be identified, providing the basis for the sound source position for subsequent propagation model construction.

[0032] The structural conduction path length between the near-field sound source candidate area and each monitoring point is obtained, and the frequency-dependent attenuation coefficient of the structural conduction path under different frequency bands is calculated by combining the spectral energy distribution of the vibration time domain signal. The structural conduction path is the physical channel for vibration propagation in the relay shell, and the path length is closely related to signal attenuation. The acquisition process first calculates the shortest conduction path according to the relay structure geometry, then analyzes the energy attenuation law of the vibration signal at different frequency bands to establish a quantitative relationship between attenuation and distance, and extracts the frequency-dependent attenuation coefficient. This coefficient reflects the attenuation characteristics of vibration propagation in the structure at different frequencies, and is a key parameter for constructing an accurate propagation model.

[0033] Based on the frequency-dependent attenuation coefficient and the time delay feature matrix, an acoustic- vibration coupling propagation model is constructed, which includes the difference characteristics of the transfer path of the near-field sound source excitation component and the structural conduction vibration component. The acoustic-vibration coupling propagation model is a mathematical expression describing the propagation mechanism of vibration in the relay, which combines acoustic propagation and structural vibration mechanisms. The construction process first estimates the vibration wave propagation speed, establishes the frequency domain transfer function, then introduces the multi-source assumption and spatial sparsity constraint to form a complete coupling model. This model can distinguish between vibrations generated by near-field sound sources (such as contact impact) and vibrations generated by structural conduction (such as coil excitation), providing a theoretical basis for signal separation.

[0034] By introducing spatial constraints and frequency domain sparsity constraints into the acoustic-vibration coupling propagation model, a multi-component decoupling objective function is established, and an iterative optimization algorithm is used to solve the multi-component decoupling objective function to separate the relay body vibration signal. Multi-component decoupling is a key step in extracting pure body signals, and through the optimization algorithm, mixed vibrations are separated into components of different sources. The decoupling process first determines the spatial constraint range of the relay body vibration, identifies the dominant frequency to construct the frequency domain sparsity constraint, then establishes a decoupling function with the objective of minimizing reconstruction error, and solves it iteratively through the alternating direction multiplier method, finally extracts the relay body vibration signal. This signal eliminates environmental noise and propagation distortion, and directly reflects the internal mechanical state of the relay, laying a foundation for subsequent feature extraction.

[0035] The vibration signal of the relay body is decomposed in time and frequency, the transient impact feature of the contact action stage and the steady vibration feature of the coil excitation stage are extracted, and a multi-dimensional state feature vector is constructed. Time-frequency decomposition is an effective method for analyzing non-stationary vibration signals, which can obtain feature information in time domain and frequency domain at the same time. The decomposition process uses continuous wavelet transform for time-frequency analysis, identifies the time boundary of contact action, and extracts transient impact feature parameters; at the same time, the steady-state frequency band corresponding to the coil excitation is analyzed, and the frequency statistical feature is extracted; and the modal energy distribution feature is obtained through empirical mode decomposition. These multi-dimensional features together constitute the state feature vector, which comprehensively describes the working state feature of the relay, and provides a feature basis for state recognition.

[0036] The multi-dimensional state feature vector is matched with the preset relay state feature library, the current running state of the relay is determined according to the matching result, and the state diagnosis conclusion is output. Similarity matching is the last step of state judgment, and the running state is determined by comparing the difference between the current feature and the known state. The matching process first normalizes the feature vector, then calculates the Mahalanobis distance of each state reference vector in the feature library, converts it into a similarity value for sorting comparison, and selects the state corresponding to the highest similarity as the diagnosis result. The system can identify normal state and multiple fault states such as contact wear, spring fatigue and coil aging, and provide diagnosis confidence evaluation to ensure the reliability of the diagnosis result.

[0037] In the embodiment of the application, the detailed implementation steps of constructing the time delay feature matrix between the monitoring points include:

[0038] The vibration time domain signal is divided into multiple time periods according to the action period of the relay, and each time period corresponds to a complete action process of the relay. Time period division is the basic step of signal analysis, and dividing continuous monitoring data into meaningful units facilitates detailed analysis. The division process first identifies the starting point of the relay action through energy detection algorithm, and then sets the window length according to the typical action duration, ensuring that each time period contains a complete on-off process. For relays with frequent actions, adaptive window technology is used to dynamically adjust the segment length to ensure that all feature information is captured. The divided time period usually includes the precursor stage (coil excitation starts), the contact action stage (contact or separation) and the holding stage (stable connection or disconnection), which records the complete dynamic characteristics of the relay working process.

[0039] The normalized cross-correlation function of the two signals is calculated. The normalized cross-correlation is a classical method for measuring the similarity of two signals, and can effectively extract the phase difference information. The calculation process first preprocesses the signals, including detrending, filtering and normalization, to eliminate the effects of baseline drift and amplitude difference; then the cross-correlation function is calculated by time domain convolution or frequency domain multiplication method, and the cross-correlation sequence representing the phase relationship of the two signals is obtained. Normalization ensures that the result is not affected by the signal amplitude, and only reflects the phase relationship. The calculation formula is:

[0040] ;

[0041] wherein, is the normalized cross-correlation function, and are the vibration signals of the two monitoring points, is the time lag, indicates the summation operation over the entire time period.

[0042] The peak position in the normalized cross-correlation function is searched, and the time lag corresponding to the peak value is recorded as the initial time delay estimate between the two monitoring points. Peak searching is a key step in determining the initial time delay, and the peak position directly corresponds to the time difference of maximum correlation. The search process uses a peak detection algorithm, which first locates the maximum cross-correlation value in the global range to determine the rough time delay position; then verifies the rationality of the peak value within the physical constraint range, and eliminates possible false peaks; finally, the time lag corresponding to the peak value is recorded as the initial estimate. For cases with low signal-to-noise ratio, a multi-peak analysis strategy is used to judge the true peak value by combining amplitude and width information, improving the noise resistance of time delay estimation.

[0043] A search window is set around the initial time delay estimate, and a parabolic interpolation fitting is performed on the normalized cross-correlation function to obtain a sub-sampling precision time delay correction value. Interpolation fitting is an important technique for improving time delay precision, which can break through the limitation of the original sampling rate through sub-sampling processing. The fitting process sets a window of ± several sampling points around the initial peak value, usually 3-5 points, and then uses a quadratic parabolic model to perform least squares fitting on the peak value region to calculate the theoretical peak position. This method can obtain a time delay precision much better than the sampling interval, and is particularly suitable for high-precision time delay analysis. The difference between the fitted peak position and the initial integer sampling point position is the time delay correction value, usually in decimal form, representing the fine adjustment amount of sub-sampling level.

[0044] The initial time delay estimation value is added to the time delay correction value to obtain the accurate time delay between two monitoring points. The accurate time delay calculation is the last step of time delay estimation, which integrates the information of integer sample accuracy and sub-sample accuracy. The calculation process is simple and direct, and the initial time delay estimation value (integer sample number) is added to the time delay correction value (decimal part) to obtain the final accurate time delay value. For a system with a sampling rate of , the time delay resolution can reach level, wherein is the interpolation multiple, and the accuracy of the original sampling interval of 1 / 10 to 1 / 20 can be usually obtained, which provides a high-precision data basis for subsequent gradient analysis.

[0045] All monitoring point pairs are traversed, and the accurate time delay between each monitoring point pair is filled into the matrix to construct a time delay feature matrix. The time delay feature matrix is a symmetric matrix with zero diagonal elements. The time delay feature matrix is a complete data structure for describing the time delay relationship of the system, which contains the time propagation relationship between all monitoring point pairs. The construction process traverses all possible monitoring point pair combinations through double loops, calculates and records the accurate time delay between each pair, and fills it into the corresponding matrix element position. Since the time delay from point A to point B is anti-symmetric with the time delay from point B to point A (numerical value is equal and sign is opposite), the matrix has anti-symmetry, that is, ; at the same time, the time delay of a point to itself is zero, so the diagonal elements are all zero. The finally generated time delay feature matrix is an n*n matrix (n is the number of monitoring points), which completely records the time delay relationship of vibration propagation and provides a data basis for sound source positioning and propagation path analysis.

[0046] In the embodiment of the application, the detailed implementation steps for identifying the near-field sound source candidate area on the relay shell include:

[0047] A monitoring point in the time delay feature matrix is selected as a reference point, and the time delay values between the reference point and all other monitoring points are extracted to form a time delay distribution sequence. The reference point selection is the starting step of gradient analysis, which affects the accuracy and efficiency of subsequent calculation. The monitoring point with high signal quality, stable position and wide coverage is preferred as the reference point, and the monitoring point near the core component of the relay is usually selected. The time delay distribution sequence is extracted from the matrix row (or column) data corresponding to the reference point to form a one-dimensional sequence, which records the propagation time relationship of vibration from the reference point to all other points. For a system with monitoring points, when the th monitoring point is selected as the reference point, the time delay distribution sequence is , containing elements. This sequence is the basic data for spatial gradient calculation, which directly affects the accuracy of the gradient field.

[0048] Based on the time delay distribution sequence and the spatial coordinates of each monitoring point, the least squares method is used to fit the gradient field of the time delay with respect to spatial location, obtaining the time delay gradient vector at the reference point. Gradient field fitting is a key step in extracting continuous gradient information from discrete time delay data, obtaining gradient information by establishing a functional relationship between time delay and spatial location. The fitting process first establishes a spatial linear regression model of the time delay, using the time delay value as the dependent variable and the spatial coordinate increment as the independent variable; then, a design matrix is ​​constructed, with each row containing the three-dimensional coordinate increment of a monitoring point relative to the reference point. Finally, the overdetermined equations are solved using the least squares method to obtain the partial derivatives of the time delay with respect to the x, y, and z directions, which are then combined to form the gradient vector. The linear regression equation is:

[0049] ;

[0050] in, For position The time delay at the location, For reference point The time delay at that point is always zero. Let be the time delay gradient vector to be solved. The least squares solution is:

[0051] ;

[0052] in, To design the matrix, its row vectors represent the coordinate increments of each point relative to a reference point. This represents the time delay vector at each point. The resulting gradient vector... The direction of the wavefront points to the direction in which the vibration wave propagates the fastest, and its magnitude reflects the steepness of the wavefront, making it an important indicator for sound source localization.

[0053] Repeat the above steps for all monitoring points to obtain the time delay gradient vector at each monitoring point, thus constructing a time delay gradient vector field for each monitoring point. Constructing the gradient vector field is a crucial step in extending single-point gradient analysis to global analysis, forming a complete vector field by calculating the gradient at each monitoring point. The construction process iteratively sets each monitoring point as a reference point, repeatedly performing time delay sequence extraction and gradient fitting operations, calculating the gradient vector at that point, until the gradient information for all monitoring points is obtained. The final gradient vector field is a set of vectors. Each vector corresponds to a monitoring point location and points in the direction of vibration propagation at that location. This vector field comprehensively describes the spatial distribution characteristics of vibration propagation, providing complete vector data for subsequent divergence calculations.

[0054] The divergence value of the time delay gradient vector field at each monitoring point is calculated, and the divergence value is obtained by summing the spatial partial derivatives of the time delay gradient vector. The divergence calculation is the core step of extracting the sound source information from the gradient vector field, which corresponds to the source-sink distribution in the flow field in physics. The calculation process adopts a numerical differentiation method to calculate the spatial variation rate of the gradient vector along the x, y and z directions at each monitoring point, and the divergence value is obtained by summing. The divergence calculation formula is:

[0055] ;

[0056] Since the monitoring points are distributed discretely, the partial derivative is calculated by using a difference approximation in the actual calculation, and a central difference format is usually used to improve the precision. The divergence value has a clear physical meaning, a positive value indicates that the point is a wave source (divergence point), a negative value indicates that the point is a convergence point, and a zero value indicates a normal point in the propagation process. The sound source position usually corresponds to the maximum value region of the divergence field, and is a direct basis for identifying the vibration source.

[0057] The spatial distribution of the divergence value is counted, and the monitoring points and their neighborhoods whose divergence values are greater than a preset divergence threshold are marked as near-field sound source candidate regions, wherein the preset divergence threshold is determined according to the mean value and standard deviation of the statistical distribution of the divergence value. The near-field sound source candidate region marking is the last step of sound source positioning, and the possible sound source position is identified by thresholding the divergence field. The statistical process first analyzes the distribution characteristics of the divergence value, calculates the mean value and the standard deviation ; then the divergence threshold is set according to the statistical characteristics, which is usually in the form of , wherein is a multiple coefficient, which is adjusted according to the detection sensitivity requirement, and the typical value is 2-3; finally, the points whose divergence values exceed the threshold are marked as candidate sound source points, and their spatial neighborhoods (usually spherical regions) are also marked, forming a complete candidate region. For multiple sound source cases, a clustering algorithm is used to merge similar candidate points into independent candidate regions to avoid repeated marking. These candidate regions are the focus of subsequent analysis and provide sound source position constraints for the construction of the propagation model.

[0058] In the embodiment of the application, the detailed implementation steps of calculating the frequency-dependent attenuation coefficient of the structure conduction path under different frequency bands include:

[0059] According to the geometry of the relay housing and the location of the near-field sound source candidate area, the shortest structural conduction path from the near-field sound source candidate area to each monitoring point is calculated, denoted as path length value. The structural conduction path is a physical channel for vibration propagation in solid structures, and the path length directly affects the propagation attenuation. The calculation process first establishes a three-dimensional geometric model of the relay housing, including the shell, support and internal structure, etc. major components; Then map the near-field sound source candidate area and monitoring points to the model, marked as nodes; Finally, an improved Dijkstra algorithm is used to search for the shortest path on the geometric model, considering the constraints of structural continuity and material properties. For complex structures, grid processing is used to discretize continuous geometry into a node network, improving computational efficiency. The final path length value is the basic data for attenuation analysis, reflecting the actual propagation distance of vibration from the sound source to the monitoring point, usually recorded in millimeters, providing a spatial reference for subsequent attenuation law analysis.

[0060] The vibration time-domain signal is subjected to short-time Fourier transform to obtain the time-frequency spectrum of the vibration signal, and the instantaneous energy sequence of the preset frequency band is extracted from the time-frequency spectrum. Short-time Fourier transform (STFT) is a classic method for analyzing the time-frequency characteristics of non-stationary signals, which can show the variation law of signal frequency components with time. The transformation process first selects an appropriate window function (usually Hanning window) and window length for signal segmentation; Then the FFT calculation is performed on the signal segment in each window to obtain the local frequency spectrum; Finally, the frequency spectrum of all windows is arranged in time sequence to form a complete time-frequency spectrum. The selection of the preset frequency band is based on the typical frequency range of the relay vibration, usually including the low frequency band (50-500Hz, corresponding to structural resonance), the medium frequency band (500-2000Hz, corresponding to mechanical action) and the high frequency band (2000-5000Hz, corresponding to contact impact). The instantaneous energy sequence is obtained by integrating the corresponding frequency band of the time-frequency spectrum, which reflects the variation of energy in a specific frequency band with time, and is the key input data for attenuation analysis.

[0061] The monitoring points in the near-field sound source candidate area are selected as reference energy points, and the energy ratio of other monitoring points in the same frequency band to the reference energy point is calculated, which is recorded as the energy attenuation ratio. Energy attenuation ratio analysis is a direct method to study the attenuation law of vibration propagation. By comparing the energy levels at different distances, the attenuation process is quantitatively described. The analysis process first selects the point with the best signal quality in the candidate area as the reference point, which represents the energy level at the sound source. Then, the energy ratio of other points in the same frequency band to the reference point is calculated to obtain the energy attenuation ratio. Finally, the attenuation ratio and the corresponding path length are paired to form the relationship between attenuation and distance. The energy attenuation ratio is a dimensionless value, usually ranging from 0 to 1, and the smaller the value, the more serious the attenuation. This attenuation analysis method based on measured data avoids the simplification assumptions of theoretical models and can accurately reflect the complex attenuation phenomena in actual structures, providing reliable experimental data for solving the attenuation coefficient.

[0062] A logarithmic linear relationship between the energy attenuation ratio and the path length value is established, and the slope of the logarithmic linear relationship is solved by regression analysis method. The slope is recorded as the frequency-dependent attenuation coefficient at this frequency band. Logarithmic linear regression is a classic method to extract the attenuation coefficient, based on the physical law that vibration energy decays exponentially during propagation. The regression process first takes the natural logarithm of the energy attenuation ratio, converting it into a linear relationship model with distance. Then, the least squares method is used for linear regression to fit the optimal straight line. Finally, the slope of the straight line is extracted as the attenuation coefficient. The logarithmic linear relationship model is:

[0063] ;

[0064] wherein, is the energy attenuation ratio, is the path length value, is the attenuation coefficient to be solved. The value is usually positive, with a unit of 1 / mm or 1 / m, and the larger the value, the faster the attenuation. This coefficient directly reflects the propagation loss characteristics of vibration in a specific frequency band and structure, and is the core parameter of the propagation model, which is crucial for predicting long-distance vibration response.

[0065] The above operation is repeated for multiple preset frequency bands to obtain frequency-dependent attenuation coefficients in different frequency bands, and a frequency-dependent attenuation coefficient spectrum is constructed. The attenuation coefficient spectrum is a complete data set describing the variation of attenuation with frequency, revealing the frequency-selective characteristics of structural conduction. The construction process performs energy attenuation analysis and regression calculation for each preset frequency band to obtain the attenuation coefficient of the frequency band, and finally forms the corresponding relationship between frequency and attenuation coefficient. For complex structures, the attenuation coefficient usually increases with increasing frequency, showing the characteristics of fast attenuation at high frequencies and long propagation at low frequencies, which is consistent with the viscoelasticity of materials and the phonon scattering mechanism of structures. The complete attenuation coefficient spectrum is usually recorded in the form of a table or a function curve, providing a parameter basis for the propagation prediction of different frequency components, and is a key element for building an accurate propagation model.

[0066] In the embodiments of the present application, the detailed implementation steps for constructing the sound-vibration coupling propagation model include:

[0067] According to the time delay characteristic matrix and the spatial coordinates of the monitoring points, the average propagation speed of the vibration wave in the relay shell is estimated. Propagation speed estimation is a basic step for building a propagation model, and directly affects the accuracy of the time-space mapping. The estimation process is based on the proportional relationship between time delay and distance. First, extract the time delay values of multiple pairs of monitoring points from the time delay characteristic matrix; then calculate the Euclidean distance between these point pairs; finally, calculate the local propagation speed by the ratio of distance to time delay, and take the average value as the overall estimate. To improve accuracy, a weighted average method is used, giving higher weights to high signal-to-noise ratio and short distance measurements. For complex structures, there may be directional differences, and the propagation speed in different directions needs to be estimated separately. The estimation formula is:

[0068] ;

[0069] wherein, is the average propagation speed, is the distance between the pair of monitoring points, is the corresponding time delay value, is the weight coefficient. The propagation speed is usually in units of m / s, and the elastic wave speed in typical metal structures is in the range of 3000-6000 m / s, with the specific value being affected by material properties and structural form. Accurate speed estimation provides a time-space conversion benchmark for subsequent phase calculation.

[0070] Based on the average propagation velocity and frequency-dependent attenuation coefficient, a frequency domain transfer function is constructed from the near-field sound source candidate region to each monitoring point. The frequency domain transfer function includes a phase delay term and an amplitude attenuation term. The frequency domain transfer function is a mathematical expression describing the phase and amplitude transformations of vibration propagation from the sound source to the monitoring point, and is a core component of the sound-vibration propagation model. The construction process first calculates the phase delay based on the average propagation velocity and geometric distance, expressed as a linear function of frequency; then, it calculates the amplitude attenuation based on the frequency-dependent attenuation coefficient, expressed as an exponential function of distance and frequency; finally, the two parts are combined to form the complete transfer function. The transfer function is expressed in complex form, with the real and imaginary parts corresponding to the amplitude and phase information, respectively. For a distance of... , frequency is The transfer function expression is:

[0071] ;

[0072] Among them, the first item This is the amplitude attenuation term. For frequency The attenuation coefficient at the location; the second term For phase delay term, The propagation speed is represented by this transfer function. This function comprehensively describes the transformation law of vibration during propagation, providing a propagation model basis for subsequent signal separation.

[0073] Assuming multiple independent near-field sound sources and at least one structurally conducted excitation source exist on the relay housing, the vibration signal at each monitoring point is represented as a weighted superposition of all sound source points and the structurally conducted excitation source. The multi-source assumption is a key approach to handling complex vibration fields, decomposing the observed mixed signal into contributions from multiple independent sources. The model assumes that relay vibration contains two types of sources: near-field sound sources (such as local impacts from contact collisions) and structurally conducted excitation sources (such as overall vibrations generated by coil excitation). The superposition model is based on linear system theory, representing the signal at each monitoring point as a weighted sum of the source signals modulated by their corresponding transfer functions. For n near-field sound sources and 1 structurally conducted source, the... The vibration signal model for each monitoring point is as follows:

[0074] ;

[0075] in, For monitoring points The vibration signal spectrum For the first The source signal spectrum of a near-field sound source The source signal spectrum of the structural conduction excitation source. and are the corresponding transfer functions. This superposition model is the theoretical basis of signal separation, which converts the complex mixing problem into a solvable linear equation set.

[0076] The spatial sparsity assumption of sound source points is introduced, and a sparse constraint matrix of sound source positions is constructed. The sparse constraint matrix is embedded in the mathematical expression of the weighted superposition. The spatial sparsity constraint is an important technique for solving underdetermined separation problems, based on the physical characteristics of the discrete distribution of sound sources in space. The constraint construction first establishes a dense grid array on the relay shell, regarding each grid point as a potential sound source position; then the L1 norm regularization term is introduced to encourage the contribution of most grid points to be zero, and only the contribution of a small number of real sound source points is retained; finally, the constraint is embedded in the superposition model to form a sparse representation problem. In practical implementation, the space is discretized into m grid points, of which only a small number k (k << m) are real sound sources, and the model becomes:

[0077] with the constraint condition:

[0078] where, represents the L0 norm of vector S, that is, the number of non-zero elements. Since the L0 problem is difficult to solve directly, it is usually transformed into an L1 constraint for approximation. This sparse constraint greatly reduces the number of parameters to be solved, improving the solvability and stability of the solution of the separation problem.

[0079] The mathematical expression of the weighted superposition, the frequency domain transfer function, and the sparse constraint matrix are combined to form a sound-vibration coupling propagation model. The sound-vibration coupling propagation model can describe the contribution distribution of near-field sound source components and structural conduction vibration components at each monitoring point. The sound-vibration coupling propagation model is a complete mathematical expression that integrates the propagation characteristics of multiple sources, combining independent components into a unified analysis framework. The model integrates the linear superposition principle, the frequency domain transfer function, and the spatial sparsity constraint to form a compact expression in matrix form:

[0080] ;

[0081] where, is the observation signal matrix of all monitoring points, is the near-field sound source transfer function matrix, is the near-field sound source signal matrix, is the structural conduction transfer function matrix, is the structural conduction source signal matrix, is the noise term. This model completely describes the characteristic differences of different types of vibration sources in the propagation process, distinguishing between local near-field effects and global structural conduction effects, providing a theoretical framework for signal separation and source identification. It is the basis model for subsequent multi-component decoupling.

[0082] In the embodiment of the application, the detailed implementation steps for separating the relay body vibration signal include:

[0083] According to the structural characteristics of the relay, the main position of the relay body vibration is determined, and the position is set as the spatial constraint range of the structural conduction excitation source. The spatial constraint is a priori knowledge introduced to improve the separation accuracy, which is based on the physical understanding of the structure and working principle of the relay. The constraint setting first analyzes the positions and functions of the core components of the relay, identifies the positions that generate the main structural vibration, which are usually the coil assembly and the core position; then the accurate three-dimensional coordinate range is determined by combining the CAD model or physical measurement; finally, the region is defined as the spatial constraint of the structural conduction excitation source. The constraint form is usually a spherical or ellipsoidal region in three-dimensional space, with the center located at the geometric center of the coil and the radius covering the main part of the coil and the core. This spatial constraint directly limits the solution space of the decoupling algorithm, avoiding the structural vibration source being incorrectly assigned to other positions, and improving the physical reasonableness and stability of the separation.

[0084] The energy distribution of the statistical vibration time domain signal in the frequency domain is identified, the dominant frequency component with an energy ratio greater than a preset ratio threshold is identified, and a frequency domain sparsity constraint condition is constructed. The frequency domain sparsity constraint condition limits the frequency spectrum of the relay body vibration signal to have non-zero values only at the dominant frequency component. The frequency domain sparsity constraint is an important constraint condition that utilizes the spectral characteristics of the vibration signal, based on the physical fact that different vibration sources have different spectral characteristics. The constraint construction first obtains the power spectral density by performing FFT transformation on the vibration signal; then sorts the energy contribution and calculates the cumulative energy distribution; then selects a frequency point set whose cumulative energy accounts for 80-90% of the total energy, and defines it as the dominant frequency component; finally, a frequency domain mask function is constructed, which takes the value 1 at the dominant frequency and the value 0 at other frequencies. This constraint limits the frequency domain distribution of the body vibration signal, and the body vibration (such as coil excitation vibration) usually has a relatively concentrated frequency spectrum, mainly containing the fundamental frequency and its harmonic components, which are significantly different from the broadband impact vibration and random noise in frequency domain characteristics, providing an important frequency domain basis for signal separation.

[0085] A multi-component decoupling objective function is established with the minimum reconstruction error as the target, and the reconstruction error is the mean square error between the measured vibration signal of each monitoring point and the sound-vibration coupling propagation model reconstruction signal. The objective function design is the core of the optimization solution, which clearly defines the evaluation standard of the separation quality. The design process first defines the reconstruction error index, which uses the Euclidean distance of the frequency domain signal to measure the difference between the observed signal and the model reconstruction signal; then minimizes the error as the main optimization target to form the basic objective function framework; finally, the constraint term and the regularization term are added to perfect the objective function structure. The basic form is the least square error:

[0086] ;

[0087] where, denotes the Frobenius norm of a matrix, i.e., the sum of the squares of all elements. This error term directly measures the fitting degree of the model to the observed data, and is the basic measure of decoupling quality. The smaller the error, the more accurate the separation and the more complete the reconstruction.

[0088] The spatial constraint range corresponding regularization term and the L1 norm penalty term corresponding to the frequency domain sparsity constraint condition are added in the multi-component decoupling objective function. Constraint introduction is a key step to incorporate prior knowledge into the optimization process, and the constraint effect is achieved through the mathematical form of the regularization term. The introduction process first converts the spatial constraint into a quadratic regularization term to concentrate the solution in the constraint region; then converts the frequency domain sparsity constraint into an L1 norm penalty term to make the spectrum tend to zero at non-dominant frequencies; finally, adjust the weight coefficients of each term to balance the fitting accuracy and constraint strength. The complete objective function is:

[0089] where the first term is the reconstruction error, the second term is the sparsity constraint (L1 norm) of the near-field source, and the third term is the spatial constraint (weighted L2 norm) of the structural conduction source, and are weight coefficients, and is the spatial weight matrix. This composite objective function considers data fitting and prior constraints, balances model complexity and generalization ability, and provides a clear optimization direction for the solution algorithm.

[0090] The alternating direction multiplier method is used to iteratively solve the multi-component decoupling objective function, and the estimated values of the near-field sound source component and the structural conduction vibration component are updated alternately each time. The alternating direction multiplier method (ADMM) is an efficient algorithm for solving constrained optimization problems, and is particularly suitable for handling composite objective functions. The solution process first initializes the near-field sound source component and the structural conduction component to zero matrices; then enters the iteration loop, which is executed in two steps each iteration: fixing updates , and fixing updates ; the reconstruction error is continuously reduced during the iteration process until the error change is less than the preset threshold or the maximum number of iterations is reached. Each update uses convex optimization techniques to solve The update is realized by closed-form solution of quadratic programming. The ADMM algorithm can effectively deal with non-smooth constraint terms, avoid the limitations of traditional gradient methods, has fast convergence speed, is not sensitive to initial value, and is particularly suitable for large-scale separation problems. After iteration convergence, the optimal estimation of the near-field sound source component and the structure conduction component is obtained, and the effective separation of the mixed vibration signal is realized.

[0091] The structure conduction vibration component is extracted from the converged solution result, denoted as the relay body vibration signal. The body vibration signal extraction is the final step of the decoupling process, which converts the optimization solution result into an actual time-domain signal. The extraction process first obtains the frequency-domain estimation of the structure conduction component from the optimization result ; then projects it to each monitoring point through the frequency-domain transfer function to obtain the body vibration spectrum at the monitoring point; finally, it is converted back to the time domain through inverse Fourier transform to obtain the complete body vibration time-domain signal. For multiple monitoring points, the point with the highest signal-to-noise ratio is usually selected as the final output, or the robustness is improved through weighted averaging. The extracted body vibration signal eliminates near-field interference and environmental noise, accurately reflects the internal mechanical state of the relay, mainly contains the vibration components generated by coil excitation and mechanical transmission, provides a high-quality signal basis for subsequent state feature extraction, and is the key input data for state monitoring and fault diagnosis.

[0092] In the embodiment of the application, the detailed implementation steps of constructing a multi-dimensional state feature vector include:

[0093] The relay body vibration signal is subjected to continuous wavelet transform to obtain a time-frequency distribution graph, and the starting time and ending time of relay contact action are identified according to the energy ridge line of the time-frequency distribution graph. Continuous wavelet transform (CWT) is an efficient tool for analyzing non-stationary signals, providing better time-frequency resolution than STFT. The transformation process first selects a wavelet basis function suitable for transient signal analysis, usually using Morlet wavelet or Gabor wavelet; then through scale transformation and translation operation, the convolution of the signal and the wavelet basis function is calculated to obtain a time-frequency coefficient matrix; finally, the coefficient matrix is converted into an energy density graph to visually display the distribution of signal energy in the time-frequency plane. The energy ridge line is the connection line of the energy peak value in the time-frequency graph, reflecting the time-varying characteristics of the main components of the signal, and is particularly suitable for identifying the time boundaries of transient events. Contact action usually shows a significant energy mutation in the time-frequency graph, the starting time corresponds to the point where the energy rises rapidly, and the ending time corresponds to the point where the energy falls back to the background level. The accurate identification of this time boundary provides a time window positioning for subsequent feature extraction, ensuring that the feature calculation focuses on the key action stage.

[0094] The vibration signal segment between the start time and the end time is extracted, denoted as a contact action stage signal, and a peak factor and a pulse factor of the contact action stage signal are calculated as transient impact characteristic parameters. The transient impact characteristic parameters are important indexes for describing the transient characteristics of the contact action and reflect the impact strength and waveform characteristics. The parameters are calculated as follows. First, the contact action stage signal is extracted according to the identified time boundary. Then, the peak factor, that is, the ratio of the signal peak value to the root mean square value, is calculated to reflect the steepness of the signal peak. At the same time, the pulse factor, that is, the ratio of the signal peak value to the average absolute value, is calculated to reflect the impact strength of the signal. The two dimensionless parameters together describe the transient characteristics of the contact action and are highly sensitive to the contact state and the change in the action speed. Normal contact action is characterized by moderate peak factor and pulse factor, and contact wear or spring fatigue failure usually leads to significant changes in these parameters, for example, poor contact increases the peak factor, and spring weakening reduces the pulse factor. These transient characteristic parameters are important indexes of the mechanical state of the relay and provide direct characteristic basis for fault diagnosis.

[0095] In the time-frequency distribution diagram, a steady-state vibration frequency band corresponding to the coil excitation is identified, time-frequency coefficients in the frequency band are extracted, and statistical moment features of the time-frequency coefficients, including frequency center, frequency variance and frequency skewness, are calculated as steady-state vibration characteristic parameters. The steady-state vibration characteristic parameters are key indexes for describing the coil excitation response and reflect the frequency spectrum characteristics of the electromagnetic excitation. The parameters are extracted as follows. First, the coil excitation frequency band is identified in the time-frequency diagram, which is usually a continuous and stable energy distribution in the low-frequency region (50-500 Hz). Then, the time-frequency coefficient matrix of the frequency band is extracted. Finally, the statistical moment features of the frequency spectrum are calculated. The frequency center reflects the position of the main frequency where the energy is concentrated, the frequency variance describes the dispersion degree of the frequency spectrum, and the frequency skewness represents the asymmetry of the frequency spectrum distribution. These features together constitute the frequency spectrum image of the coil excitation response and are highly sensitive to the change in the coil state. Normal coil is characterized by stable frequency center and appropriate frequency spectrum width, and coil aging or insufficient excitation leads to frequency center shift, frequency spectrum widening or skewness change. These steady-state features are important indexes of the electrical state of the coil and provide frequency domain feature support for the diagnosis of coil-related faults.

[0096] The relay body vibration signal is subjected to empirical mode decomposition, a plurality of intrinsic mode function components are obtained, energy proportions of the intrinsic mode function components are calculated, and instantaneous frequency averages of the first three intrinsic mode function components with the largest energy proportions are selected as the modal characteristic parameters. The modal characteristic parameters are deep features describing the internal structure of the signal and reflect the inherent characteristics of the vibration system. The parameter extraction first subjects the signal to empirical mode decomposition (EMD), adaptively decomposes the complex signal into a plurality of intrinsic mode functions (IMFs), and each IMF represents a vibration mode of different scale; then the energy proportions of the IMF components are calculated, and the dominant mode is identified; and finally, the instantaneous frequency average of the main IMF is calculated as the modal characteristic parameter. The instantaneous frequency is obtained through Hilbert transform and reflects the frequency variation characteristics of the mode with time. These modal characteristics are directly related to the mechanical structure characteristics of the relay and are highly sensitive to structural changes and mechanical looseness. The normal relay shows stable modal energy distribution and instantaneous frequency, and mechanical failure will cause energy distribution change or instantaneous frequency drift. These modal characteristics provide a structural dynamics perspective for fault diagnosis and are particularly suitable for early detection and identification of mechanical faults.

[0097] The transient impact characteristic parameters, the steady-state vibration characteristic parameters and the modal characteristic parameters are arranged in order to form a multi-dimensional state characteristic vector. The multi-dimensional state characteristic vector is a feature set describing the comprehensive state of the relay, which integrates multi-angle information in the time domain, the frequency domain and the modal domain. The vector construction organizes the features in the order of functional modules, including transient impact characteristics (such as peak factor, pulse factor), steady-state vibration characteristics (such as frequency center, frequency variance, frequency skewness) and modal characteristics (such as instantaneous frequency average of the main IMF). This multi-dimensional feature structure realizes comprehensive representation of different aspects of the relay state, the transient characteristics reflect the contact action state, the steady-state characteristics reflect the coil electrical state, and the modal characteristics reflect the mechanical structure state. Through this comprehensive characteristic vector, the characteristic changes caused by different types of faults can be captured, providing a rich feature basis for subsequent state recognition and fault diagnosis, and greatly improving the accuracy and comprehensiveness of state monitoring.

[0098] In the embodiment of the present application, the detailed implementation steps for determining the current operating state of the relay include:

[0099] The multi-dimensional state feature vector is normalized to eliminate the influence of different dimensional differences. Normalization is the basic step of feature processing, which ensures that features of different dimensions have equal weights in subsequent analysis. The processing process first calculates the mean and standard deviation of the statistical distribution of each feature in the training set; then the Z-score standardization method is used to subtract the mean and divide by the standard deviation, converting it to a standard distribution with a mean of 0 and a standard deviation of 1. For some non-normal distribution features, the quantile normalization method is used to convert based on the cumulative distribution function, ensuring that the normalized feature distribution is more uniform. Special attention is paid to abnormal value processing during normalization, and the extreme value influence is reduced through quantile truncation or logarithmic transformation techniques. The normalized feature vector has similar numerical ranges in each dimension, eliminating the original dimensional differences (such as peak factor usually 3-5, frequency center may be hundreds of hertz), providing a fair basis for subsequent distance calculation and improving the accuracy and robustness of feature matching.

[0100] The reference feature vectors of each known state category are extracted from the relay state feature library, which contains reference feature vectors corresponding to normal state, contact wear state, spring fatigue state and coil aging state. Reference feature vector extraction is the premise of state matching, providing standard feature templates for known states. The feature library is constructed based on a large amount of experimental and field data, and the feature vectors of typical states are extracted through vibration signal analysis of different state relays. After verification, they are stored in the feature library. The feature library contains multiple state categories: normal state represents the ideal working state of the relay; contact wear state reflects the contact abnormality caused by the increase in surface erosion and roughness; spring fatigue state describes the mechanical force weakening caused by elastic element failure; coil aging state shows the decline in electrical performance such as insulation degradation and inter-turn short circuit. Each state usually contains multiple feature vector samples, reflecting different degrees and different forms of the same state, enhancing the coverage and adaptability of the feature library. These reference feature vectors are the benchmark for state recognition, providing a comparison standard for the classification and judgment of unknown states, and are the knowledge base for state monitoring.

[0101] The Mahalanobis distance between the multi-dimensional state feature vector and each reference feature vector is calculated, and the calculation of the Mahalanobis distance considers the covariance relationship between the dimensions of the feature vectors. The Mahalanobis distance is a high-level indicator of measuring the similarity of feature vectors, considering the correlation and scale difference between features. The calculation process first estimates the covariance matrix of the feature space, describing the statistical association between the feature dimensions; then calculates the Mahalanobis distance between the feature vectors based on the covariance matrix, realizing the distance measurement considering the correlation. Compared with the Euclidean distance, the Mahalanobis distance realizes the rotation and scaling transformation of the feature space through the covariance matrix, eliminating the influence of feature correlation and providing a more reasonable similarity evaluation. For highly correlated feature dimensions, the Mahalanobis distance automatically reduces their influence weight, avoiding the amplification effect of redundant information; for high-recognizability feature dimensions, it automatically increases their contribution proportion, enhancing the discrimination ability. This intelligent weighted distance measurement is particularly suitable for similarity evaluation of multi-dimensional features, providing an accurate mathematical basis for state recognition.

[0102] The Mahalanobis distance is converted into a similarity value, which is obtained by taking the negative exponential function of the Mahalanobis distance. Similarity conversion is a processing step that converts distance indicators into more intuitive similarity indicators, facilitating comparison and threshold judgment. The conversion uses a negative exponential function mapping to convert the unbounded distance value into a similarity value in the [0, 1] interval, and the conversion formula is:

[0103] ;

[0104] wherein, is the similarity value, is the Mahalanobis distance, is a scaling parameter for adjusting the sensitivity of the conversion. A similarity value of 1 indicates complete similarity (distance is 0), and a similarity value close to 0 indicates extreme dissimilarity (distance is very large). This conversion maintains the monotonic relationship of the original distance while providing a more standardized and intuitive similarity measurement, facilitating the setting of a unified judgment threshold and simplifying the subsequent decision logic. The scaling parameter σ is usually determined through cross-validation, and the value that maximizes the inter-class discrimination is selected, typically the median of the intra-class distance in the training set. This similarity indicator provides a probabilistic decision basis for the final judgment of the state, improving the interpretability and confidence assessment ability of the diagnosis.

[0105] The state category corresponding to the reference feature vector with the maximum similarity value is selected as the preliminary diagnosis result, and the similarity value is recorded as the diagnosis confidence. The state selection is the core decision step of the diagnosis process, and the most possible state category is determined based on the maximum similarity principle. The selection process first compares the similarity values of the current feature vector and all reference feature vectors, identifies the maximum similarity and the corresponding state category, and then takes the category as the preliminary diagnosis result and records the maximum similarity value as the diagnosis confidence, which quantifies the reliability level of the result. To improve robustness, for multiple states with similar similarity, an integrated decision method can be used to consider the weighted voting results of multiple high similarity states. The diagnosis confidence is an important indicator of result evaluation, which directly reflects the matching degree of the current state and the known mode, and provides a reference basis for the subsequent decision process, especially providing a risk assessment basis at the fuzzy state boundary. This state selection strategy based on maximum similarity not only inherits the simplicity and efficiency of traditional pattern matching, but also increases the flexibility of decision-making through confidence evaluation, which adapts to the diagnosis needs of different scenes.

[0106] If the diagnosis confidence is lower than the preset confidence threshold, it is marked as a suspected abnormal state, triggering the manual review process; if the diagnosis confidence is higher than the preset confidence threshold, the preliminary diagnosis result is output as the final state diagnosis conclusion. Confidence judgment is a safety mechanism of the diagnosis process, which identifies uncertain conditions through threshold judgment to avoid false diagnosis. The judgment process first sets an appropriate confidence threshold, which is usually determined based on the accuracy statistics of historical diagnosis; then compares the confidence of the current diagnosis with the threshold to make a corresponding decision. For high confidence results, the preliminary diagnosis is directly adopted as the final conclusion to improve processing efficiency; for low confidence results, it is marked as a suspected state and triggers manual review to make a judgment combined with expert experience and more evidence to avoid risks caused by system misjudgment. The setting of the confidence threshold balances the automation efficiency and safety reliability, which is usually between 0.7-0.85, and is adjusted according to the risk tolerance of the application scene. This double-threshold decision mechanism not only guarantees the automatic and efficient diagnosis in most cases, but also provides a safety channel for expert intervention for special and complex cases, improving the overall reliability and adaptability of the system, which is especially suitable for state monitoring applications of critical equipment.

[0107] In the embodiment of the application, according to the time delay distribution sequence and the spatial coordinates of each monitoring point, the least square method is used to fit the gradient field of the time delay with respect to the spatial position, and the detailed implementation steps of obtaining the time delay gradient vector at the reference point include:

[0108] The spatial linear regression model of the time delay is established by taking the time delay values in the time delay distribution sequence as the dependent variable and the spatial coordinate increment of each monitoring point relative to the reference point as the independent variable. The spatial linear regression model is a basic mathematical tool for mapping from discrete time delay data to continuous gradient field. The model is established by first determining the reference coordinate system, usually establishing a local rectangular coordinate system with the reference point as the origin; then calculating the coordinate increment vector of each monitoring point relative to the reference point , that is, the difference value obtained by subtracting the reference point coordinates from the monitoring point coordinates; and finally establishing a linear mapping relationship between the time delay and the spatial position. The basic form of the linear regression model expresses the linear relationship between the time delay and the coordinate increment, where the coefficients to be solved are the partial derivatives of the time delay in each direction. This linear model represents the rate of change of the time delay in three-dimensional space as the partial derivative in each direction, which intuitively reflects the variation of the time delay in space and lays a mathematical foundation for the gradient vector solution.

[0109] The design matrix of the spatial linear regression model is constructed, and each row of the design matrix contains the components of the spatial coordinate increment of a monitoring point. The design matrix is a key data structure for least squares solution, which converts the problem into a standard matrix form for easy calculation. The matrix construction process first determines the matrix dimension, for n monitoring points (excluding the reference point itself), the design matrix is an n x 3 matrix, where 3 represents the coordinate dimension of three-dimensional space; then fill in the matrix elements, the i-th row corresponds to the coordinate increment of the i-th monitoring point. At the same time, the dependent variable vector is constructed, which contains the time delay values of all monitoring points. Each column of the design matrix corresponds to a spatial direction, and each row corresponds to an observation point, which organizes the discrete spatial-time delay correspondence into a standard matrix form, facilitating subsequent least squares solution. This matrix design makes full use of the redundant information of multi-point measurement, improving the noise resistance and accuracy of gradient estimation.

[0110] The least squares method is used to solve the spatial linear regression model to obtain the partial derivatives of the time delay in each direction of the spatial coordinate. The least squares method is a classic method for solving overdetermined equations, which finds the optimal fitting parameters by minimizing the sum of squared residuals. The solution process expresses the linear regression problem as a matrix equation, where the parameter vector to be solved contains the partial derivatives of the time delay in x, y, and z directions. When the number of monitoring points is greater than 3, the equation system is usually overdetermined, and there is no exact solution. In this case, the least squares criterion is used to find the solution that minimizes the sum of squared residuals. To improve the stability of the calculation, numerical methods such as QR decomposition and SVD decomposition are usually used in practical implementation, especially when the monitoring points are unevenly distributed, which may lead to ill-conditioned matrix problems. The least squares method not only provides an estimate of the partial derivative, but also evaluates the fitting quality through residual analysis, providing a reference for the credibility of the gradient result. The final obtained partial derivative directly reflects the rate of change of the time delay in three orthogonal directions and is the basic component for constructing the gradient vector.

[0111] The partial derivatives are combined into a vector form to obtain a time delay gradient vector at the reference point, the direction of the time delay gradient vector indicates the fastest direction of the vibration wave propagation, and the modulus value of the time delay gradient vector reflects the steepness of the vibration wave front. The gradient vector combination is the last step of the analysis process, which integrates the partial derivatives in three directions into a unified vector expression. The combination process directly arranges the solved partial derivatives according to the spatial direction to form a three-dimensional vector. This vector has a clear physical meaning: the vector direction is perpendicular to the time delay contour, and points to the direction in which the time delay increases the fastest, that is, the direction of the vibration wave propagation; the vector modulus value represents the time delay change rate per unit distance, reflecting the steepness of the wave front, and is inversely proportional to the wave speed. For a point sound source, the gradient direction should theoretically point to the sound source position; for a plane wave, the gradient direction is perpendicular to the wave front. The modulus value of the gradient vector is related to the intensity and distance of the vibration source, and generally the closer to the vibration source, the larger the gradient modulus value. In practical applications, by analyzing the gradient vector field at multiple reference points, the vibration propagation path can be tracked, and the vibration source position can be identified, providing key information for subsequent sound source positioning and propagation analysis. This analysis method based on spatial gradient fully utilizes the geometric characteristics of vibration propagation, realizes effective mapping from time delay distribution to propagation direction, and is the mathematical basis for sound source identification.

[0112] The above describes a relay state monitoring and diagnosis method in an embodiment of the application. Next, a relay state monitoring and diagnosis system in an embodiment of the application is described. Please refer to Figure 2 An embodiment of the relay state monitoring and diagnosis system in an embodiment of the application includes:

[0113] A data acquisition module is configured to acquire vibration time domain signals of each monitoring point in a relay working process and record spatial coordinates of each monitoring point.

[0114] A matrix construction module is configured to perform a segmented cross-correlation operation on the vibration time domain signals, extract a vibration wave arrival time delay between any two monitoring points, and construct a time delay feature matrix between the monitoring points.

[0115] A sound source identification module is configured to calculate a time delay gradient vector of each monitoring point in the time delay feature matrix, and identify a near-field sound source candidate area on a relay shell according to a divergence distribution of the time delay gradient vector in space.

[0116] An attenuation coefficient calculation module is configured to obtain a structural conduction path length between the near-field sound source candidate area and each monitoring point, and calculate a frequency-dependent attenuation coefficient of the structural conduction path under different frequency bands in combination with a frequency spectrum energy distribution of the vibration time domain signals.

[0117] A coupling model construction module is configured to construct a sound-vibration coupling propagation model based on the frequency-dependent attenuation coefficient and the time delay feature matrix.

[0118] The signal decoupling module is configured to establish a multi-component decoupling objective function by introducing a spatial constraint condition and a frequency domain sparsity constraint condition in a sound-vibration coupling propagation model, and to separate the relay body vibration signal by solving the multi-component decoupling objective function using an iterative optimization algorithm.

[0119] The feature extraction module is configured to perform time-frequency decomposition on the relay body vibration signal, extract transient impact features in the contact action stage and steady-state vibration features in the coil excitation stage, and construct a multi-dimensional state feature vector.

[0120] The state diagnosis module is configured to perform similarity matching between the multi-dimensional state feature vector and a preset relay state feature library, determine the current operating state of the relay according to the matching result, and output a state diagnosis conclusion.

[0121] The various modules are connected through wired and / or wireless means to realize data transmission between the modules.

[0122] The present application realizes comprehensive monitoring and accurate diagnosis of the relay state through vibration signal acquisition, time delay feature extraction, near-field sound source identification, sound-vibration coupling model construction and state feature matching. The sound-vibration coupling analysis method of the present application can accurately separate the relay body vibration signal and effectively identify various fault states.

[0123] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or replace some of the technical features with equivalent ones without departing from the principles and spirit of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0124] It should be noted that the formulas in the present specification are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the real situation. The preset parameters and threshold values in the formula are set by those skilled in the art according to the actual situation.

[0125] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for relay condition monitoring and diagnosis, characterized in that, include: The vibration time-domain signals of each monitoring point during the operation of the relay are collected, and the spatial coordinates of each monitoring point are recorded. Perform piecewise cross-correlation on the vibration time-domain signal, extract the arrival time delay of vibration waves between any two monitoring points, and construct the time delay feature matrix between monitoring points; Calculate the time delay gradient vector for each monitoring point in the time delay feature matrix, and identify near-field sound source candidate regions on the relay housing based on the spatial divergence distribution of the time delay gradient vector; Obtain the structural conduction path length between the near-field sound source candidate region and each monitoring point, and calculate the frequency dependence attenuation coefficient of the structural conduction path under different frequency bands by combining the spectral energy distribution of the vibration time-domain signal. Based on the frequency-dependent attenuation coefficient and the time delay feature matrix, an acoustic-vibration coupled propagation model is constructed. By introducing spatial constraints and frequency domain sparsity constraints into the acoustic-vibration coupling propagation model, a multi-component decoupling objective function is established. An iterative optimization algorithm is used to solve the multi-component decoupling objective function, and the vibration signal of the relay body is separated. The vibration signal of the relay body is decomposed into time and frequency, and the transient impact characteristics of the contact action stage and the steady-state vibration characteristics of the coil excitation stage are extracted to construct a multi-dimensional state feature vector. The multidimensional state feature vector is matched with a preset relay state feature library for similarity. Based on the matching result, the current operating state of the relay is determined and a state diagnosis conclusion is output.

2. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The step of performing piecewise cross-correlation on the vibration time-domain signal, extracting the arrival time delay of the vibration wave between any two monitoring points, and constructing a time delay feature matrix between monitoring points includes: The vibration time-domain signal is divided into multiple time periods according to the relay's operating cycle; Select vibration signals from any two monitoring points within the same time period and calculate their normalized cross-correlation function. Search for the peak position in the normalized cross-correlation function, and record the time lag corresponding to the peak as the initial time delay estimate between the two monitoring points; A search window is set near the initial time delay estimate, and parabolic interpolation fitting is performed on the normalized cross-correlation function to obtain the time delay correction value with subsampling accuracy. The initial delay estimate is added to the delay correction value to obtain the precise delay between the two monitoring points; Traverse all monitoring point pairs and fill the matrix with the precise time delay between each monitoring point pair to construct the time delay feature matrix.

3. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The step of calculating the time delay gradient vector for each monitoring point in the time delay feature matrix, and identifying near-field sound source candidate regions on the relay housing based on the spatial divergence distribution of the time delay gradient vector, includes: Select one monitoring point from the time delay feature matrix as a reference point, and extract the time delay values ​​between the reference point and all other monitoring points to form a time delay distribution sequence; Based on the time delay distribution sequence and the spatial coordinates of each monitoring point, the gradient field of the time delay with respect to the spatial position is fitted using the least squares method to obtain the time delay gradient vector at the reference point. Repeat the above operation for all monitoring points to obtain the time delay gradient vector at each monitoring point and construct the time delay gradient vector field of the monitoring points. Calculate the divergence value of the time-delay gradient vector field at each monitoring point; The spatial distribution of the divergence values ​​is statistically analyzed, and monitoring points with divergence values ​​greater than a preset divergence threshold and their neighborhoods are marked as candidate regions for near-field sound sources.

4. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The step of obtaining the structural conduction path length between the near-field sound source candidate region and each monitoring point, and calculating the frequency-dependent attenuation coefficient of the structural conduction path at different frequency bands in combination with the spectral energy distribution of the vibration time-domain signal, includes: Based on the geometry of the relay housing and the location of the near-field sound source candidate region, the shortest structural conduction path from the near-field sound source candidate region to each monitoring point is calculated and denoted as the path length value. Perform a short-time Fourier transform on the vibration time-domain signal to obtain the time spectrum of the vibration signal, and extract the instantaneous energy sequence of the preset frequency band from the time spectrum; A monitoring point located within the near-field sound source candidate region is selected as a reference energy point. The energy ratio of other monitoring points in the same frequency band to that of the reference energy point is calculated and denoted as the energy attenuation ratio. Establish a logarithmic linear relationship between the energy attenuation ratio and the path length value, and use regression analysis to solve for the slope of the logarithmic linear relationship. Record the slope as the frequency-dependent attenuation coefficient in this frequency band.

5. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The construction of the acoustic-vibration coupled propagation model based on the frequency-dependent attenuation coefficient and the time delay feature matrix includes: Based on the time delay feature matrix and the spatial coordinates of the monitoring point, the average propagation speed of the vibration wave in the relay housing is estimated. Based on the average propagation speed and the frequency-dependent attenuation coefficient, a frequency domain transfer function is constructed from the near-field sound source candidate region to each monitoring point; Assuming there are multiple independent near-field sound sources and at least one structural conduction excitation source on the relay housing, the vibration signal of each monitoring point is represented as a weighted superposition of all sound sources and structural conduction excitation sources; By introducing the spatial sparsity assumption of sound source points, a sparse constraint matrix of sound source locations is constructed, and the sparse constraint matrix is ​​embedded in the weighted superposition mathematical expression. The weighted superposition mathematical expression, the frequency domain transfer function, and the sparse constraint matrix are combined to form the acoustic-vibration coupled propagation model.

6. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The process involves introducing spatial constraints and frequency domain sparsity constraints into the acoustic-vibration coupling propagation model to establish a multi-component decoupling objective function. An iterative optimization algorithm is then used to solve this objective function, resulting in the separation of the relay body vibration signal, including: Based on the structural characteristics of the relay, the main location where the relay body vibration is generated is determined, and this location is set as the spatial constraint range of the structural conduction excitation source; The energy distribution of the vibration time-domain signal in the frequency domain is statistically analyzed, the dominant frequency components with an energy proportion greater than a preset proportion threshold are identified, and frequency domain sparsity constraints are constructed. Establish a multi-component decoupling objective function with the goal of minimizing reconstruction error; Add the regularization term corresponding to the spatial constraint range and the L1 norm penalty term corresponding to the frequency domain sparsity constraint to the multi-component decoupling objective function; The multi-component decoupling objective function is solved iteratively using the alternating direction multiplier method. In each iteration, the estimated values ​​of the near-field sound source component and the structure-transmitted vibration component are updated alternately until the reconstruction error converges to a preset error threshold. The structural vibration component is extracted from the converged solution and denoted as the vibration signal of the relay body.

7. The relay condition monitoring and diagnosis method according to claim 1, characterized in that, The time-frequency decomposition of the relay body vibration signal extracts the transient impact characteristics during the contact operation phase and the steady-state vibration characteristics during the coil excitation phase, constructing a multi-dimensional state feature vector, including: A continuous wavelet transform is performed on the vibration signal of the relay body to obtain a time-frequency distribution map. The start and end times of the relay contact action are identified based on the energy ridge line of the time-frequency distribution map. The vibration signal segment between the start time and the end time is extracted and denoted as the contact action stage signal. The peak factor and impulse factor of the contact action stage signal are calculated as transient impact characteristic parameters. In the time-frequency distribution diagram, identify the steady-state vibration frequency band corresponding to the coil excitation, extract the time-frequency coefficients in the frequency band, and calculate the statistical moment characteristics of the time-frequency coefficients, including frequency center, frequency variance and frequency skewness, as steady-state vibration characteristic parameters; Empirical mode decomposition is performed on the vibration signal of the relay body to obtain multiple intrinsic mode function components. The energy ratio of each intrinsic mode function component is calculated, and the average instantaneous frequency of the three intrinsic mode function components with the largest energy ratio is selected as the modal characteristic parameter. The transient impact characteristic parameters, the steady-state vibration characteristic parameters, and the modal characteristic parameters are arranged in order to form the multidimensional state characteristic vector.

8. The relay status monitoring and diagnosis method according to claim 1, characterized in that, The step of performing similarity matching between the multidimensional state feature vector and a preset relay state feature library, and determining the current operating state of the relay based on the matching result, includes: The multidimensional state feature vector is normalized to eliminate the influence of differences in the dimensions of different feature parameters; Reference feature vectors for each known state category are extracted from the relay state feature library, which includes reference feature vectors corresponding to normal state, contact wear state, spring fatigue state, and coil aging state. Calculate the Mahalanobis distance between the multidimensional state feature vector and each reference feature vector; Convert the Mahalanobis distance into similarity values; The state category corresponding to the reference feature vector with the highest similarity value is selected as the preliminary diagnosis result, and the similarity value is recorded as the diagnosis confidence. If the diagnostic confidence level is lower than the preset confidence threshold, it is marked as a suspected abnormal state, triggering a manual review process; if the diagnostic confidence level is higher than the preset confidence threshold, the preliminary diagnostic result is output as the final state diagnostic conclusion.

9. The relay condition monitoring and diagnosis method according to claim 3, characterized in that, The step of fitting the gradient field of the time delay with respect to spatial location using the least squares method based on the time delay distribution sequence and the spatial coordinates of each monitoring point to obtain the time delay gradient vector at the reference point includes: Using the time delay values ​​in the time delay distribution sequence as the dependent variable and the spatial coordinate increments of each monitoring point relative to the reference point as the independent variable, a spatial linear regression model of time delay is established. Construct the design matrix of the spatial linear regression model, wherein each row of the design matrix contains the components of the spatial coordinate increment of a monitoring point; The spatial linear regression model is solved using the least squares method to obtain the partial derivatives of the time delay with respect to each direction of the spatial coordinates. The partial derivatives are combined into a vector form to obtain the time delay gradient vector at the reference point.

10. A relay condition monitoring and diagnostic system, used to implement the relay condition monitoring and diagnostic method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire the vibration time-domain signals of each monitoring point during the operation of the relay and record the spatial coordinates of each monitoring point; The matrix construction module is used to perform piecewise cross-correlation calculation on the vibration time-domain signal, extract the arrival time delay of vibration waves between any two monitoring points, and construct the time delay feature matrix between monitoring points. The sound source identification module is used to calculate the time delay gradient vector of each monitoring point in the time delay feature matrix, and to identify near-field sound source candidate regions on the relay housing based on the spatial divergence distribution of the time delay gradient vector. The attenuation coefficient calculation module is used to obtain the structural conduction path length between the near-field sound source candidate region and each monitoring point, and calculate the frequency-dependent attenuation coefficient of the structural conduction path under different frequency bands by combining the spectral energy distribution of the vibration time-domain signal. The coupling model construction module is used to construct an acoustic-vibration coupling propagation model based on the frequency-dependent attenuation coefficient and the time delay feature matrix. The signal decoupling module is used to establish a multi-component decoupling objective function by introducing spatial constraints and frequency domain sparsity constraints into the acoustic-vibration coupling propagation model, and to solve the multi-component decoupling objective function by using an iterative optimization algorithm to separate the vibration signal of the relay body. The feature extraction module is used to perform time-frequency decomposition on the vibration signal of the relay body, extract the transient impact features of the contact action stage and the steady-state vibration features of the coil excitation stage, and construct a multi-dimensional state feature vector. The status diagnosis module is used to perform similarity matching between the multidimensional status feature vector and a preset relay status feature library, determine the current operating status of the relay based on the matching result, and output the status diagnosis conclusion.

Images