Waveform self-similarity fault type identification method and system based on tanimoto coefficient

Abstract

The application relates to the technical field of power distribution network protection, and provides a waveform self-similarity fault type identification method and system based on a Tanimoto coefficient. The method comprises the following steps: acquiring multi-channel time sequence waveform data, pre-processing the waveform data and performing multi-scale division, and generating a multi-scale waveform feature primitive set containing time domain, frequency domain and time-frequency domain feature vectors; based on the multi-scale waveform feature primitive set, a generalized Tanimoto coefficient between different physical channels and different analysis scales is calculated, and a cross-scale and cross-channel self-similarity tensor is constructed; the self-similarity tensor is subjected to tensor decomposition, a time intensity vector of an atomic self-similarity mode corresponding to a fault mode is extracted, and is fused into a decoupled self-similarity feature vector; the decoupled self-similarity feature vector and a pre-stored standard fault feature template are subjected to weighted distance calculation and hierarchical comparison, and a fault type identification conclusion is output according to a comparison result. The method is helpful to improve the accuracy of fault type identification under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of equipment fault monitoring technology, and in particular to a method and system for identifying fault types based on waveform self-similarity coefficients. Background Technology

[0002] In the field of modern industrial equipment condition monitoring and fault diagnosis, the automated analysis and intelligent identification of multi-physical field time-series waveform signals such as vibration and current collected by sensors has become a key technology for ensuring safe equipment operation and implementing predictive maintenance. Existing methods mostly focus on feature extraction in a single signal domain (such as the time domain or frequency domain) or rely on simple direct matching of preset fault templates. When faced with early weak composite faults with low signal-to-noise ratios under complex operating conditions, multiple coupled fault sources, these methods often fail to effectively decouple different fault modes mixed in the signal, resulting in insufficient differentiation of fault types and an increased misjudgment rate. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for identifying fault types based on waveform self-similarity coefficients, in order to solve the problems mentioned in the background art.

[0004] Firstly, this application provides a waveform self-similarity fault type identification method based on Tanimoto coefficients, including: Acquire multi-channel time-series waveform data, preprocess and multi-scale partition the waveform data to generate a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors; based on the multi-scale waveform feature primitive set, calculate the generalized Tanimoto coefficients between different physical channels and different analysis scales to construct cross-scale and cross-channel self-similarity tensors; perform tensor decomposition on the self-similarity tensors to extract the time intensity vectors of atomic self-similarity patterns corresponding to fault modes, and fuse them into decoupled self-similarity feature vectors; perform weighted distance calculation and hierarchical comparison between the decoupled self-similarity feature vectors and pre-stored standard fault feature templates, and output fault type identification conclusions based on the comparison results.

[0005] Secondly, this application provides a waveform self-similarity fault type identification system based on Tanimoto coefficients, comprising: The acquisition module acquires multi-channel time-series waveform data, preprocesses and divides the waveform data into multiple scales, and generates a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors. The calculation module calculates the generalized Tanimoto coefficients between different physical channels and different analysis scales based on the multi-scale waveform feature primitive set, and constructs cross-scale and cross-channel self-similarity tensors. The decomposition module performs tensor decomposition on the self-similarity tensor, extracts the time intensity vectors of atomic self-similarity patterns corresponding to fault modes, and fuses them into decoupled self-similarity feature vectors. The comparison module performs weighted distance calculation and hierarchical comparison between the decoupled self-similarity feature vectors and pre-stored standard fault feature templates, and outputs a fault type identification conclusion based on the comparison results.

[0006] This application provides a method and system for identifying fault types based on waveform self-similarity coefficients. Firstly, this method extracts multi-scale feature primitives in the time, frequency, and time-frequency domains from multi-physics field signals, constructing a feature set that comprehensively describes the inherent patterns of the signals, thus providing a rich and structured data foundation for subsequent in-depth analysis. Secondly, by calculating generalized Tanimoto coefficients across channels and scales and constructing a self-similarity tensor, it performs three-dimensional and quantitative modeling of the correlation patterns between features from different sources and of different properties, thereby capturing the collaborative changes and consistency patterns caused by faults in the multi-dimensional feature space. Thirdly, by decomposing the self-similarity tensor to extract atomic patterns and fusing the temporal intensity statistics of the patterns into a decoupled feature vector, it achieves the extraction and focusing of core identification information directly related to specific fault mechanisms from mixed observation signals. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0008] Figure 1 A flowchart illustrating the waveform self-similarity fault type identification method based on Tanimoto coefficients provided in this application embodiment; Figure 2 A schematic block diagram of the waveform self-similarity fault type identification system based on Tanimoto coefficients provided in the embodiments of this application. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it require execution in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0011] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0012] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the relevant listed items and all possible combinations, and includes such combinations.

[0013] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0014] Please see Figure 1 , Figure 1 This is a flowchart illustrating the waveform self-similarity fault type identification method based on Tanimoto coefficients provided in the embodiments of this application, as shown below. Figure 1 As shown, the waveform self-similarity fault type identification method based on Tanimoto coefficient provided in this application includes steps S1 to S4.

[0015] Step S1: Acquire multi-channel time-series waveform data, preprocess and multi-scale divide the waveform data to generate a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors. Specifically, firstly, the raw waveform signals collected by vibration sensors and current sensors installed on the mechanical equipment are received synchronously. The raw signals of these two channels are jointly preprocessed, that is, filters are applied simultaneously to suppress high-frequency noise and power frequency interference, and amplitude normalization is performed on the filtered signals to ensure that all data fall into a uniform numerical range to eliminate the influence of dimensions. Subsequently, the preprocessed vibration signal and current signal are divided into multi-scale segments respectively. In the time domain, a set of time windows with successively increasing lengths and no overlap are used to sequentially extract the signal of each channel, and all sampling points in each window are arranged in order to form a time-domain waveform segment vector. In the frequency domain, a windowed Fast Fourier Transform is performed on each time-domain waveform segment vector obtained in the previous step to obtain its spectrum. The highest-energy frequency components are identified from this spectrum, and only the amplitudes of these main components are arranged in frequency order to form a concise frequency-domain feature vector. In the time-frequency domain, a wavelet packet transform is performed on the complete, preprocessed long-sequence signal to decompose it into preset frequency bands related to the fault characteristic frequencies. The energy of each frequency band at different time segments is calculated and divided according to the aforementioned time-domain window alignment, forming a time-frequency energy vector characterizing the energy distribution over time and frequency. Finally, all vectors generated from the vibration and current channels at the time, frequency, and time-frequency domains are uniformly collected and a structured dataset is established. Each vector in this dataset is a "primitive" describing the local characteristics of the original signal from a specific perspective, collectively forming a multi-scale waveform feature primitive set.

[0016] Step S2: Based on the multi-scale waveform feature primitive set, calculate the generalized Tanimoto coefficients between different physical channels and different analysis scales, and construct cross-scale and cross-channel self-similarity tensors. Specifically, using the multi-scale waveform feature primitive set generated in step S1 as input, systematically calculate the similarity between any two feature vectors. The selected vector pairs have clear physical and mathematical meanings. For example, it can be comparing the time-domain segment of a vibration signal and the frequency-domain features of a current signal at the same moment, or comparing the time-frequency energy vectors of a current signal at different moments. The index used to measure similarity is the generalized Tanimoto coefficient, which is calculated based on the inner product of two vectors and their respective moduli. For each selected pair of feature vectors, a similarity value between zero and one is calculated according to the formula. The closer the value is to one, the higher the overlap and morphological consistency of the two vectors in the measured feature space. After completing the similarity calculation for all preset vector pairs, these values ​​need to be organized into a three-dimensional data structure. The first and second dimensions of this structure correspond to the indices of the two time windows associated with the vector pair, while the third dimension uniquely identifies the specific "relationship type" represented by the vector pair, such as "vibration time domain - current frequency domain" or "vibration time-frequency domain - vibration time-frequency domain". By filling all the calculated similarity values ​​into the corresponding positions of this three-dimensional structure according to their time window indices and relationship types, the cross-scale and cross-channel self-similarity tensor is constructed. This tensor is no longer a simple two-dimensional matrix, but a three-dimensional data model that can simultaneously characterize the complex correlation patterns between multi-physical fields and multi-dimensional features of the signal during its time evolution.

[0017] Step S3: Perform tensor decomposition on the self-similarity tensor, extract the time intensity vectors of atomic self-similar patterns corresponding to the fault modes, and fuse them into decoupled self-similar feature vectors. Specifically, the self-similarity tensor constructed in step S2 is used as input, and parallel factorization, a multilinear algebra tool, is used to decompose it. This decomposition process approximates the complex high-dimensional tensor as a product of several factor matrices and a core tensor. In this decomposition result, the factor matrix associated with the third dimension of the tensor (i.e., the relation type dimension) is of key significance: each column vector of this matrix represents an atomic, purely self-similar pattern. This pattern describes the intensity change trajectory of a specific cross-channel or cross-scale correlation (e.g., the synchronicity of vibration shock and a specific current modulation sideband) over the entire analysis time range. Based on prior knowledge of the fault mechanism of the target equipment, several columns of atomic pattern vectors most relevant to the fault type to be identified are pre-selected from this factor matrix. Then, statistical analysis is performed on the time intensity vectors of each selected atomic pattern, calculating statistics such as peak intensity, sequence arrangement complexity, and periodic intensity. These statistics quantify the performance characteristics of the atomic pattern in the signal from different perspectives. Finally, the various statistics calculated for all selected atomic patterns are concatenated in a fixed order to form a new one-dimensional numerical vector. This vector is called the decoupled self-similar feature vector. It is no longer a direct observation of the original signal, nor a mixed similarity tensor, but a set of core discriminative features extracted and condensed from the tensor and directly related to the potential root causes of faults.

[0018] Step S4: Perform weighted distance calculation and hierarchical comparison between the decoupled self-similar feature vector and the pre-stored standard fault feature template, and output the fault type identification conclusion based on the comparison result.

[0019] Specifically, the decoupled self-similar feature vector obtained in step S3 is compared with a pre-established and stored standard fault feature template library. The template library stores two key data points for each known fault type (such as bearing inner ring damage, rotor imbalance, etc.): a typical decoupled self-similar feature vector (standard feature vector) under that fault condition, and a corresponding weight vector indicating the difference in importance of each feature component in the standard feature vector for identifying this type of fault. The comparison process first calculates the weighted Mahalanobis distance between the vector to be diagnosed and each template in the library. This is a distance metric that considers the correlation between features and the identification weight of each component. After calculation, templates with a weighted Mahalanobis distance less than a preset first-level threshold are selected; the fault types corresponding to these templates constitute a candidate fault set. Then, a hierarchical decision-making process is performed: if the candidate fault set contains only one fault type, that type is directly output as the diagnostic conclusion; if the candidate set contains multiple fault types, a second-level arbitration is initiated, calculating the sum of the absolute deviations between the vector to be diagnosed and each candidate fault on its highest-weighted key feature components, and selecting the candidate fault with the smallest total deviation as the final diagnostic conclusion; if the candidate set is empty after the first-level screening, meaning no template is sufficiently close to the vector to be diagnosed, a conclusion labeled "unknown fault state" is output. The entire process is based entirely on numerical calculation and logical judgment, ultimately generating a clear fault type label or status indication.

[0020] The method provided in this embodiment, on the one hand, constructs a feature set that comprehensively describes the intrinsic patterns of signals by extracting multi-scale feature primitives in the time domain, frequency domain, and time-frequency domain from multi-physics field signals, thus providing a rich and structured data foundation for subsequent in-depth analysis; on the other hand, by calculating the generalized Tanimoto coefficients across channels and scales and constructing a self-similarity tensor, it performs three-dimensional and quantitative modeling of the correlation patterns between features from different sources and of different properties, thereby capturing the collaborative changes and consistency patterns caused by faults in the multi-dimensional feature space; furthermore, by decomposing the self-similarity tensor to extract atomic patterns and fusing the temporal intensity statistics of the patterns into decoupled feature vectors, it achieves the extraction and focusing of core identification information directly related to specific fault mechanisms from mixed observation signals.

[0021] In some embodiments, acquiring multi-channel time-series waveform data, preprocessing and multi-scale partitioning the waveform data, and generating a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors includes: Step S11: Simultaneously acquire the raw data of vibration signal and current signal, and perform joint noise reduction and amplitude normalization processing on the raw data.

[0022] Specifically, the raw waveform data sequences uploaded by the vibration acceleration sensor and current transformer are synchronously read from the data acquisition system. Joint noise reduction processing is performed simultaneously on the signals from both channels. First, a low-pass digital filter with an adjustable cutoff frequency is applied to filter out high-frequency random noise much higher than the characteristic frequencies of major mechanical equipment faults. Then, a notch filter is designed specifically to suppress strong interference from power system frequencies and their harmonics on the current signal. Amplitude normalization processing is performed separately on the filtered signals of each channel. Its purpose is to eliminate absolute amplitude differences caused by variations in sensor sensitivity and gain settings. Specifically, the absolute maximum value of each channel signal is calculated within a sufficiently long analysis segment. Then, every data point within the entire analysis segment of that channel is divided by this maximum value, ensuring that the values ​​of all data points fall between -1 and +1, thus achieving a unified amplitude scale.

[0023] Step S12: Use multiple non-overlapping windows of different lengths to segment the processed channel signals, and use the data point sequence in each window as a time-domain waveform segment vector.

[0024] Specifically, the preprocessed vibration and current signals are segmented using a pre-defined window length scheme. This scheme includes several different window durations; for example, short windows are used to capture transient impacts, while long windows are used to analyze gradual trends. During segmentation, starting from the signal's origin, segments are continuously extracted according to the first window length, with adjacent windows connected end-to-end without overlap, until the entire analysis period is covered, resulting in a set of time-domain waveform segments of that length. Then, the second window length is used, and the same non-overlapping segmentation process is repeated. For each channel and each window length, all segments are processed independently. For each specific segment, all data points sampled in chronological order within it are arranged into an ordered list, which is defined as a time-domain waveform segment vector. This vector completely preserves the waveform fluctuation information of the original signal within that time period.

[0025] Step S13: Perform a fast Fourier transform on each time-domain waveform segment vector to extract its first N main spectral amplitudes to form a frequency domain feature vector.

[0026] Specifically, for each time-domain waveform segment vector obtained in step S12, it is first multiplied by a window function to reduce spectral leakage, and then the Fast Fourier Transform (FFT) algorithm is applied. This algorithm converts the time-domain vector into a complex sequence, i.e., a complex spectrum, where each element of the sequence corresponds to a complex representation of a specific frequency component. Next, the magnitude of each element in this complex sequence is calculated, which is the amplitude of the corresponding frequency component, thus obtaining a spectral amplitude sequence composed of various frequency amplitudes. Subsequently, this amplitude sequence is sorted, and its elements are rearranged in descending order. After sorting, the N largest amplitudes at the top are selected, and their corresponding frequency indices in the original spectrum are recorded. Finally, these N amplitudes are rearranged into a new one-dimensional real number array of length N according to their corresponding frequency indices in ascending order. This array is the frequency domain feature vector extracted from the original time-domain segment, which centrally reflects the main distribution of signal energy in the frequency domain during that time period.

[0027] Step S14: Perform wavelet packet transform on the entire signal segment, extract the energy sequence of the preset frequency band and divide it into windows to generate a time-frequency energy vector.

[0028] Specifically, wavelet packet transform is performed on a complete, preprocessed long sequence of vibration or current signals. Wavelet packet transform is a time-frequency analysis method that can decompose a signal into multiple different frequency bands and provide time-varying analysis within each band. Based on the frequency range of possible fault characteristics of the diagnosed equipment, a set of frequency bands requiring focus are pre-defined. Through wavelet packet transform, the decomposed coefficient sequence of the signal in each preset frequency band can be obtained, and the length of this sequence is the same as the number of time points of the original signal. Then, the energy of the coefficient sequence of each preset frequency band in each small time interval is calculated, usually by calculating the sum of squares of the coefficients within that time interval. To align with the time-domain analysis window in step S12, the boundaries of these small time intervals are consistent with the window boundaries used for time-domain segmentation. In this way, for each preset frequency band, a sequence of energy changing over time (by window index) is obtained. Extracting the energy values ​​of all preset frequency bands within the same time window and arranging them in frequency band order constitutes the time-frequency energy vector of that window, which represents the distribution of signal energy in different frequency bands within that time interval.

[0029] Step S15: Organize the time-domain waveform segment vectors, frequency-domain feature vectors, and time-frequency energy vectors from different channels into the multi-scale waveform feature primitive set.

[0030] Specifically, all feature vectors generated in steps S11 to S14 are summarized and structured. Specifically, the set includes all time-domain waveform segment vectors, all frequency-domain feature vectors, and all time-frequency energy vectors from the vibration channel, as well as the aforementioned three types of vectors from the current channel. During organization, necessary metadata tags are attached to each vector. These tags at least indicate the physical channel (vibration or current) to which the vector belongs, the analysis scale used (time domain, frequency domain, or time-frequency domain), the corresponding time window index, and, for frequency-domain and time-frequency domain vectors, their frequency-related information. Finally, all these clearly identified feature vectors are stored in a unified data structure or database, collectively forming a multi-scale waveform feature primitive set. This set is a rich, multi-perspective feature pool with clear physical and mathematical semantics, providing comprehensive material for subsequent analysis.

[0031] The method provided in this embodiment, on the one hand, eliminates noise interference and dimensional differences by jointly denoising and normalizing the original signals from multiple channels, thus providing a clean and comparable data source for subsequent multi-scale analysis; on the other hand, by using non-overlapping windows of different lengths for time-domain segmentation, it captures local waveform details of the signal at different time resolutions, thus taking into account both transient events and long-term trends; furthermore, by performing spectral analysis on each time-domain segment and extracting the main amplitude, and by performing wavelet packet transform on the entire signal to extract specific frequency band energy, it reveals the inherent patterns of the signal's periodicity and modulation from two complementary dimensions of frequency and time-frequency, and finally forms a multi-scale feature primitive set through structured organization.

[0032] In some embodiments, performing a fast Fourier transform on each time-domain waveform segment vector to extract its first N main spectral amplitudes to form a frequency domain feature vector includes: step S131, applying a windowed fast Fourier transform algorithm to each time-domain waveform segment vector to obtain the corresponding complex spectrum.

[0033] Specifically, for each time-domain waveform segment vector to be processed in step S13, a windowing operation is first performed. A suitable window function is selected, such as the Hanning window, and the coefficient sequence of the window function is multiplied one by one with each corresponding element of the time-domain waveform segment vector. The purpose of this operation is to reduce the spectral leakage effect caused by signal truncation, so that the spectrum obtained by the subsequent transformation can better reflect the frequency characteristics of the signal itself. After windowing, the Fast Fourier Transform algorithm is applied to the new sequence. This algorithm transforms the time-domain sequence into a complex sequence of the same length through efficient recursive calculation. This complex sequence is the complex spectrum, where each complex number corresponds to a specific frequency component, the modulus of the complex number represents the amplitude of the frequency component, and the argument of the complex number represents the phase of the frequency component.

[0034] Step S132: Calculate the amplitude of each frequency component in the complex spectrum to generate a spectrum amplitude sequence.

[0035] Specifically, after obtaining the complex spectrum, amplitude information needs to be extracted. For each complex element in the complex spectrum sequence, its modulus is calculated. The formula for calculating the modulus is to take the square root of the sum of the squares of the real and imaginary parts of the complex number. Performing this calculation on the complex numbers corresponding to all frequency components in the spectrum yields a sequence of non-negative real numbers, the length of which is the same as the complex spectrum. This sequence of real numbers is the spectral amplitude sequence, which visually represents the energy distribution of each frequency component in the original time-domain segment, where the frequency points with high amplitudes correspond to the main frequency components with concentrated energy in the signal.

[0036] Step S133: Sort the spectrum amplitude sequence in descending order according to amplitude size.

[0037] Specifically, the spectral amplitude sequence generated in step S132 is sorted. The purpose of sorting is to identify the frequency components with the most concentrated energy. The sorting is performed in descending order of amplitude. Using a sorting algorithm, the element with the largest amplitude in the original spectral amplitude sequence is placed at the first position of the new sequence, the second largest at the second position, and so on, until the element with the smallest amplitude is placed at the end. This sorting process does not change any amplitude data itself, but only changes their order in the sequence, thus making it easier for subsequent steps to locate and select the components with the highest energy.

[0038] Step S134: Select the N largest amplitudes and their corresponding frequency numbers from the sorted spectrum amplitude sequence.

[0039] Specifically, in the descending-ordered sequence of spectral amplitudes, the first N elements are extracted sequentially, starting from the beginning of the sequence (i.e., the position with the largest amplitude). These N elements represent the amplitudes of the N frequency components with the highest energy in the original signal during that time period. Simultaneously, the position indices of these N amplitudes in the original, unsorted complex spectrum must be recorded. These position indices are the frequency serial numbers, which directly correspond to the specific frequency values ​​(frequency serial number multiplied by frequency resolution). Therefore, the output of this step is two sets of one-to-one data: one set contains the N largest amplitudes, and the other set contains the frequency serial numbers corresponding to each of these N amplitudes.

[0040] Step S135: Arrange the selected N amplitude values ​​in order of their corresponding frequency numbers to form an N-dimensional real vector; the N-dimensional real vector is the frequency domain feature vector, which is used to characterize the energy concentration distribution pattern of the time domain waveform segment vector in the frequency domain.

[0041] Specifically, the N amplitude values ​​selected in step S134 are rearranged according to their corresponding frequency indices in ascending order. That is, the magnitude order of the amplitude values ​​themselves is ignored; instead, they are organized according to the natural order of the frequencies they represent from low to high. This arrangement yields a one-dimensional real number array of length N. This array is the final frequency domain feature vector. Its significance lies in the fact that it preserves the skeleton information of the original signal spectrum in a compact form (containing only the N most important components), and the order of its elements is consistent with the frequency axis order. This makes the frequency domain feature vectors of different time segments comparable and can effectively characterize the distribution pattern of the signal's most concentrated energy in the frequency domain.

[0042] The method provided in this embodiment, on the one hand, suppresses spectral leakage by performing Fourier transform on the time-domain segment after windowing, thereby obtaining a complex spectrum that more accurately reflects the frequency components of the signal; on the other hand, it obtains an amplitude sequence by calculating the modulus of the complex spectrum, and selects the top N maximum values ​​from the sorted amplitude sequence, thereby focusing on the most significant and energy-concentrated frequency components in the signal, achieving effective compression and purification of spectral information; furthermore, it constructs a feature vector by rearranging the selected amplitudes according to the frequency sequence, thereby maintaining the physical order of the frequency structure in the generated feature vector, which facilitates subsequent quantitative comparison and analysis based on waveform morphology similarity.

[0043] In some embodiments, the step of calculating the generalized Tanimoto coefficients between different physical channels and different analysis scales based on the multi-scale waveform feature primitive set, and constructing cross-scale and cross-channel self-similarity tensors, includes: Step S21: Select any two feature vectors from the vibration and current channels, and from the time domain, frequency domain, or time-frequency domain, from the multi-scale waveform feature element set.

[0044] Specifically, the multi-scale waveform feature primitive set is used as the operation object. The selection process follows a set of predefined rules based on the physical mechanisms of fault diagnosis. For example, the rules may specify the need to compare time-domain segments of vibration signals and current signals at the same moment to examine the direct similarity between the two waveforms; or to compare the frequency domain characteristics of the vibration signal at the current moment with the time-frequency energy of the current signal several time intervals later to examine the delay correlation between vibration features and current modulation. During selection, two feature vectors that conform to the rules are identified and extracted from the set based on the vector's metadata tags (channel, scale, time window). These vector pairs constitute the basic unit for subsequent similarity calculations, and their relationship type (such as "vibration time domain - current frequency domain") is predefined, ensuring the target nature and engineering significance of the calculation.

[0045] Step S22: Calculate the generalized Tanimoto coefficient between each pair of feature vectors. The formula for calculating the generalized Tanimoto coefficient is the inner product of the two vectors divided by the sum of the squares of the magnitudes of the two vectors minus the inner product.

[0046] Specifically, for each pair of feature vectors selected in step S21, the generalized Tanimoto coefficient is calculated. The calculation first requires finding the inner product of the two vectors, i.e., multiplying corresponding elements and then summing the results. Next, the squares of the magnitudes of each vector are calculated, i.e., the squares of all elements of each vector are summed. Then, according to the formula, the inner product value is used as the numerator. The denominator is the square of the magnitude of the first vector plus the square of the magnitude of the second vector, minus the inner product value just calculated. Finally, the numerator is divided by the denominator to obtain a scalar value. This value is the generalized Tanimoto coefficient, which describes the degree of overlap and the similarity of the overall shape of the two vectors in the feature space they span. It is relatively insensitive to the absolute magnitude of the vectors, but sensitive to the consistency of the vector element distribution.

[0047] Step S23: The generalized Tanimoto coefficients calculated for all selected feature vector pairs are arranged into a three-dimensional data structure according to the time window ordinal number associated with the feature vector pair and the channel and scale combination type to which the feature vector pair belongs. The three-dimensional data structure is the self-similarity tensor.

[0048] Specifically, after calculating the generalized Tanimoto coefficients for all preset vector pairs, these results need to be systematically stored. To this end, a three-dimensional array is constructed. The indices of the first and second dimensions of this array represent the time window numbers to which the two feature vectors in the vector pair belong, respectively. The index of the third dimension is a classification identifier, uniquely corresponding to a specific "channel and scale combination type," i.e., the relationship type defined when selecting vector pairs according to rules in step S21. Each calculated generalized Tanimoto coefficient is accurately filled into the corresponding coordinate position of this three-dimensional array according to its two corresponding time window numbers and its relationship type. Once all coefficients are filled, this complete three-dimensional array constitutes a self-similarity tensor. This tensor demonstrates the evolution of similarity in the time dimension (first and second dimensions) and reveals the differences in the association patterns between different physical quantities and different feature perspectives in the relationship dimension (third dimension).

[0049] The method provided in this embodiment, on the one hand, selects feature vector pairs from a multi-scale primitive set according to physical mechanism rules, ensuring that similarity calculation focuses on specific correlations with diagnostic value, thereby avoiding meaningless calculations and enhancing the specificity of features; on the other hand, by using the generalized Tanimoto coefficient as a similarity measure, it effectively captures the discriminative common patterns caused by faults in the multi-dimensional feature space, rather than simple amplitude differences, by utilizing its sensitivity to morphological consistency and relative insensitivity to absolute amplitude; furthermore, by organizing the calculated similarity coefficients into a three-dimensional tensor structure according to time windows and relationship types, a three-dimensional data model capable of simultaneously characterizing the temporal evolution of signals and the complexity of multi-dimensional correlations is constructed, laying the foundation for subsequent deep decomposition analysis.

[0050] In some embodiments, the step of performing tensor decomposition on the self-similarity tensor, extracting the time intensity vector of the atomic self-similarity mode corresponding to the fault mode, and fusing them into a decoupled self-similarity feature vector includes: Step S31: Perform a normalized parallel factorization on the self-similarity tensor to obtain a core tensor and multiple factor matrices.

[0051] Specifically, the self-similarity tensor constructed in step S23 is used as input, and a parallel factorization algorithm is applied. This algorithm is a multilinear algebraic analysis method whose goal is to decompose a high-order tensor into the best approximation of the product of several factor matrices and a core tensor. The decomposition process is iteratively optimized, using numerical methods such as alternating least squares to solve for the factor matrices and core tensor that minimize the difference between the decomposed model and the original tensor data. For a three-dimensional tensor, the decomposition typically yields three factor matrices, corresponding to the three dimensions of the tensor (i.e., two time dimensions and one relation type dimension), and a core tensor of the same order, which characterizes the strength of the interactions between the column vectors of the factor matrices. Performing this decomposition aims to decouple the observed complex self-similarity tensor, which mixes multiple modes, into a series of simpler, more clearly interpretable components.

[0052] Step S32: Select several pre-defined column vectors related to the target fault mechanism from the factor matrix associated with the channel-scale combination type dimension. Each column vector represents the temporal intensity distribution of an atomic self-similar mode.

[0053] Specifically, in the factor matrix obtained from the decomposition, the factor matrix corresponding to the third dimension (i.e., the "channel-scale combination type" dimension) of the self-similarity tensor is located. Each column vector of this matrix has a clear physical meaning: it represents an "atomic" self-similarity pattern. This pattern is a pure, basic correlation type, such as "high synchronization between vibration shock energy and specific sideband energy of current". The values ​​of this column vector in each row represent the activation intensity or manifestation degree of this atomic pattern over the entire analysis time range (corresponding to the time window sequence), and therefore it is called the temporal intensity distribution vector of the pattern. Based on prior knowledge of different fault mechanisms of the diagnosed equipment (e.g., a certain fault will specifically excite certain atomic patterns), several specific column vectors related to these faults are pre-selected from this factor matrix.

[0054] Step S33: Calculate the statistical characteristics of each column of the vector, including peak factor, permutation entropy and main period intensity.

[0055] Specifically, statistical features are extracted from the time intensity distribution vectors of each atomic mode selected in step S32. First, the peak factor is calculated; this factor is the ratio of the maximum value to the root mean square value of the vector, used to quantify the prominence of the impact component in the mode's intensity sequence. Second, the permutation entropy is calculated; this entropy value is obtained through ordinal pattern analysis of the vector sequence and is used to measure the complexity and randomness of the mode's intensity changes. A low entropy value indicates strong regularity and mode stability. Then, the dominant periodic intensity is calculated; by performing autocorrelation analysis or spectral analysis on the vector, its most dominant periodic component is identified, and the amplitude or energy of this periodic component is used as a measure of periodic intensity. These statistics characterize the dynamic behavior of the same atomic mode in the time dimension from different mathematical perspectives.

[0056] Step S34: Concatenate all the calculated statistical features in order to form a one-dimensional numerical vector, which is the decoupled self-similar feature vector.

[0057] Specifically, the statistical features (peak factor, permutation entropy, main period intensity, etc.) calculated for all selected atomic pattern column vectors in step S33 are aggregated. The aggregation follows a fixed order: for example, all statistics for the first selected pattern are arranged first, then all statistics for the second selected pattern, and so on. Finally, all these statistics are sequentially concatenated to form a longer one-dimensional real number array. This newly generated one-dimensional vector is the decoupled self-similar feature vector. It is no longer a direct time series or similarity matrix, but rather a highly condensed set of comprehensive quantitative indicators directly related to specific fault mechanisms, obtained from the original signal after multiple transformations and refinements.

[0058] The method provided in this embodiment, on the one hand, decouples the various intertwined correlation patterns in the observation data into independent atomic components by performing parallel factor decomposition on the self-similarity tensor, thereby achieving mathematical separation of complex fault coupling phenomena; on the other hand, by selecting atomic pattern vectors that are a priori related to the target fault mechanism from the decomposed factor matrix, the focus of analysis is locked on the information components most directly related to the diagnostic target, filtering out irrelevant or interfering patterns; furthermore, by calculating the peak factor, permutation entropy, and periodic intensity statistics of each atomic pattern time intensity vector, and fusing statistics from multiple patterns and multiple aspects to form the final feature vector, a highly discriminative feature descriptor is constructed that both characterizes the existence of fault modes and describes their dynamic behavior characteristics.

[0059] In some embodiments, the step of performing weighted distance calculation and hierarchical comparison between the decoupled self-similar feature vector and a pre-stored standard fault feature template, and outputting a fault type identification conclusion based on the comparison result, includes: Step S41: Calculate the weighted Mahalanobis distance between the decoupled self-similar feature vector and each standard fault feature template in the feature template library, wherein the feature template library stores standard feature vectors and weight vectors corresponding to known fault types.

[0060] Specifically, first, a pre-generated and stored feature template library is accessed. This library stores one record for each known fault type (such as "bearing outer ring damage" or "rotor misalignment"). Each record contains a standard feature vector and a weight vector. The standard feature vector is the standard value of the aforementioned decoupled self-similar feature vector extracted under the typical fault condition. The weight vector indicates the importance of each feature component in the standard feature vector for identifying this type of fault; components with higher importance have greater weight in the distance calculation. For each fault template in the library, the following operations are performed sequentially: its standard feature vector and weight vector are read; then, following the weighted Mahalanobis distance calculation process, the distance value between the decoupled self-similar feature vector to be diagnosed and the template is calculated. This distance value comprehensively considers the overall deviation between the vector to be diagnosed and the standard vector, the correlation between each feature component, and the identification weight of each component.

[0061] Step S42: Select all standard fault feature templates whose weighted Mahalanobis distance is less than the first preset threshold to form a candidate fault set.

[0062] Specifically, after calculating the weighted Mahalanobis distance with all templates in the template library, a distance list is obtained. Each distance value in this list is compared with a pre-defined first-level threshold. This threshold defines the boundary of "sufficient similarity." All fault templates with distance values ​​less than this threshold are selected; the fault types represented by these templates are considered potential candidates for the current state to be diagnosed. The identifiers of these candidate templates (such as fault type names or codes) are collected together to form a set called the candidate fault set. This step is a coarse screening of diagnostic possibilities, aiming to eliminate fault types that clearly do not match the current state.

[0063] Step S43: If there is only one element in the candidate fault set, output the fault type corresponding to this element.

[0064] Specifically, the candidate fault set formed in step S42 is checked. If the check finds that the set contains only one fault template, that is, only one fault type has a distance below the threshold, then the diagnostic process ends here. The system directly outputs the fault type corresponding to the unique candidate fault template as the final diagnostic conclusion. This situation corresponds to the state to be diagnosed having a high degree of feature matching with a certain known fault type and being significantly different from other types.

[0065] Step S44: If the candidate fault set contains multiple elements, calculate the sum of the absolute deviations between the decoupled self-similar feature vector and each element in the candidate fault set on its top K feature components with the highest weights, and output the fault type corresponding to the minimum sum of deviations.

[0066] Specifically, if the candidate fault set formed in step S42 contains two or more fault templates, it means that multiple fault types are quite close to the current state in terms of weighted Mahalanobis distance, requiring more refined differentiation. At this point, the second-level arbitration mechanism is initiated. For each fault template in the candidate set, the feature dimension indices corresponding to the K largest components in its weight vector are extracted. Then, the absolute values ​​of the differences between the values ​​of the decoupling self-similar feature vector to be diagnosed on these key dimensions of the template and the values ​​of the corresponding standard feature vectors are calculated, and these K absolute deviations are summed to obtain a "sum of key feature deviations". The sum of deviations between the vector to be diagnosed and each template in the candidate set is calculated separately. Finally, these sums of deviations are compared, and the one with the smallest value is selected. The system outputs the fault type corresponding to this smallest sum of deviations as the final diagnostic conclusion. This step utilizes the subtle differences between different faults on their most sensitive features for the final decision.

[0067] Step S45: If the candidate fault set is empty, output the unknown fault status identifier.

[0068] Specifically, if the comparison result of step S42 is that the weighted Mahalanobis distance of all templates is greater than or equal to the first preset threshold, i.e., the candidate fault set is an empty set, this indicates that the current state to be diagnosed differs significantly from the standard characteristics of all known fault types. In this case, the system cannot classify it into any known category. Therefore, the system will output a specific identifier, such as "unknown fault state" or "novel state," to prompt the operator that the current state may be a new fault type not included in the database, or a complex state formed by a combination of multiple known faults, requiring further expert analysis and confirmation.

[0069] The method provided in this embodiment, on the one hand, quickly eliminates obviously mismatched fault types by calculating the weighted Mahalanobis distance and performing coarse screening based on a first threshold, thereby narrowing the identification range and improving processing efficiency; on the other hand, when multiple candidates appear after coarse screening, it calculates the sum of deviations of the vector to be diagnosed on the most critical features of each candidate template for fine arbitration, thereby utilizing the most discriminative feature differences of different faults for final differentiation, improving the accuracy of distinguishing similar faults; furthermore, by setting a processing path with an empty candidate set and outputting an unknown state identifier, it provides a clear output mechanism for the system to identify novel or compound faults, enhancing the completeness and practicality of the method.

[0070] In some embodiments, calculating the weighted Mahalanobis distance between the decoupled self-similar feature vector and each standard fault feature template in the feature template library includes: step S411, reading a standard fault feature template stored in the feature template library, wherein the standard fault feature template includes a standard feature vector and a corresponding weight vector.

[0071] Specifically, data from the feature template library is loaded from persistent storage. When it is necessary to calculate the distance to a specific fault type (denoted as type X), the record corresponding to type X is retrieved and read from the library. This record contains two key data objects: the first is the standard feature vector, which is a one-dimensional real number array representing the typical or average state of the decoupled self-similar feature vectors extracted through this method under fault type X. The second is the weight vector, which is a one-dimensional real number array with the same length as the standard feature vector. Each element in the vector indicates the importance weight of the feature component at the corresponding position in the standard feature vector for determining whether it is a fault type X. The weight values ​​are usually positive and may have been normalized.

[0072] Step S412: Calculate the difference between the decoupled self-similar feature vector and the standard feature vector in each feature dimension.

[0073] Specifically, the decoupled self-similar feature vector to be diagnosed is subtracted element-wise from the standard feature vector read in step S411. Specifically, for the i-th dimension of the feature vector, the value of the standard feature vector in the i-th dimension is subtracted from the value of the feature vector to be diagnosed in the i-th dimension to obtain a difference. This operation is repeated for all dimensions (from the 1st to the Nth dimension), resulting in a new vector consisting of N differences, called the difference vector. This difference vector visually represents the specific direction and magnitude of the deviation of the state to be diagnosed from the standard fault state in each feature component.

[0074] Step S413: Obtain the inverse of a pre-computed global covariance matrix.

[0075] Specifically, during system initialization or template library establishment, a global covariance matrix of decoupled self-similar feature vectors has been calculated based on a training dataset containing samples of various states (normal, various faults). This matrix describes the linear correlation between the feature components. To calculate the Mahalanobis distance, the inverse of this global covariance matrix is ​​needed. In this step, this pre-calculated inverse matrix is ​​directly retrieved from storage. This inverse matrix is ​​an N×N square matrix (N is the feature vector dimension), which plays a role in standardization and decorrelation in distance calculation, enabling meaningful comparisons of feature components with different dimensions and correlations on the same scale.

[0076] Step S414: Use the weight vector to weight each component of the difference.

[0077] Specifically, the difference vector calculated in step S412 is combined with the weight vector read in step S411. The combination method is element-wise multiplication (Hadamard product). That is, for the i-th dimension, the value of the difference vector in the i-th dimension is multiplied by the value of the weight vector in the i-th dimension to obtain a new weighted difference. This process is performed on all dimensions to generate a weighted difference vector. The effect of this step is that the original difference is amplified in the calculation for feature components with high discriminative importance, while the original difference is reduced for components with low importance. This reflects the idea of ​​"differentiating" different feature components, focusing on key discriminative features.

[0078] Step S415: Calculate a scalar distance value according to the Mahalanobis distance formula based on the weighted difference vector and the inverse of the global covariance matrix; the scalar distance value is the weighted Mahalanobis distance between the decoupled self-similar feature vector and the current standard fault feature template.

[0079] Specifically, first, the weighted difference vector obtained in step S414 is transposed to obtain a row vector. Then, this row vector, the inverse of the global covariance matrix obtained in step S413, and the original weighted difference vector (column vector) are continuously multiplied according to the matrix multiplication rules: first, the product of the row vector and the inverse matrix is ​​calculated to obtain an intermediate row vector; then, this intermediate row vector is multiplied by the original weighted difference column vector, resulting in a single scalar value. This scalar value is the final weighted Mahalanobis distance. It comprehensively reflects the "statistical distance" in the feature space between the vector to be diagnosed and the current standard fault feature template after considering feature correlation and the discrimination weights of each component. This distance value will be used for comparison and screening in step S42.

[0080] Please see Figure 2 , Figure 2 A schematic block diagram of the waveform self-similarity fault type identification system 100 based on Tanimoto coefficients provided in this application embodiment is shown below. Figure 2 As shown, the waveform self-similarity fault type identification system 100 based on Tanimoto coefficients provided in this application includes: The acquisition module 110 is used to acquire multi-channel time-series waveform data, preprocess the waveform data and perform multi-scale partitioning to generate a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors. The calculation module 120 is used to calculate the generalized Tanimoto coefficients between different physical channels and different analysis scales based on the multi-scale waveform feature primitive set, constructing a cross-scale and cross-channel self-similarity tensor. The decomposition module 130 is used to perform tensor decomposition on the self-similarity tensor, extract the time intensity vectors of atomic self-similar modes corresponding to the fault modes, and fuse them into a decoupled self-similar feature vector.

[0081] The comparison module 140 is used to perform weighted distance calculation and hierarchical comparison between the decoupled self-similar feature vector and the pre-stored standard fault feature template, and output a fault type identification conclusion based on the comparison result.

[0082] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and each module described above can be referred to the process in the aforementioned embodiment of the waveform self-similarity fault type identification method based on Tanimoto coefficients, and will not be repeated here.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying fault types based on waveform self-similarity coefficients, characterized in that, include: Acquire multi-channel time-series waveform data, preprocess and divide the waveform data into multiple scales to generate a multi-scale waveform feature primitive set containing time-domain, frequency-domain and time-frequency-domain feature vectors; Based on the multi-scale waveform feature primitive set, the generalized Tanimoto coefficients between different physical channels and different analysis scales are calculated to construct cross-scale and cross-channel self-similarity tensors. The self-similarity tensors are decomposed to extract the time intensity vectors of atomic self-similarity patterns corresponding to the fault modes, and then fused into decoupled self-similarity feature vectors. The decoupled self-similarity feature vectors are compared with pre-stored standard fault feature templates using weighted distance calculation and hierarchical comparison. Based on the comparison results, the fault type identification conclusion is output.

2. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 1, characterized in that, The process of acquiring multi-channel time-series waveform data, preprocessing and multi-scale partitioning the waveform data, and generating a multi-scale waveform feature primitive set containing time-domain, frequency-domain, and time-frequency-domain feature vectors includes: Simultaneously acquire raw data of vibration and current signals, and perform joint noise reduction and amplitude normalization on the raw data; Multiple non-overlapping windows of different lengths are used to segment the processed signals of each channel, and the data point sequence in each window is used as a time-domain waveform segment vector. Perform a Fast Fourier Transform on each of the time-domain waveform segment vectors to extract its first N main spectral amplitudes to form a frequency domain feature vector; Perform wavelet packet transform on the entire signal segment, extract the energy sequence of the preset frequency band and divide it into windows to generate a time-frequency energy vector; The time-domain waveform segment vectors, frequency-domain feature vectors, and time-frequency energy vectors from different channels are collectively organized into the multi-scale waveform feature primitive set.

3. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 2, characterized in that, The step of performing a fast Fourier transform on each time-domain waveform segment vector to extract its first N main spectral amplitudes to form a frequency domain feature vector includes: applying a windowed fast Fourier transform algorithm to each time-domain waveform segment vector to obtain the corresponding complex spectrum; Calculate the amplitude of each frequency component in the complex spectrum to generate a spectrum amplitude sequence; Sort the spectrum amplitude sequence in descending order of amplitude magnitude; From the sorted sequence of spectral amplitudes, select the N largest amplitudes and their corresponding frequency indices; The selected N amplitude values ​​are arranged in order of their corresponding frequency numbers to form an N-dimensional real vector; the N-dimensional real vector is the frequency domain feature vector, which is used to characterize the energy concentration distribution pattern of the time domain waveform segment vector in the frequency domain.

4. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 1, characterized in that, The process of calculating generalized Tanimoto coefficients between different physical channels and different analysis scales based on the multi-scale waveform feature primitive set, and constructing cross-scale and cross-channel self-similarity tensors, includes: Select any two feature vectors from the vibration and current channels, and from the time domain, frequency domain, or time-frequency domain, from the set of multi-scale waveform feature elements; Calculate the generalized Tanimoto coefficient between each pair of the eigenvectors, wherein the generalized Tanimoto coefficient is calculated by dividing the inner product of the two vectors by the sum of the squares of the magnitudes of the two vectors minus the inner product. The generalized Tanimoto coefficients calculated for all selected feature vector pairs are arranged into a three-dimensional data structure based on the time window ordinal number associated with the feature vector pair and the channel and scale combination type to which the feature vector pair belongs. The three-dimensional data structure is the self-similarity tensor.

5. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 1, characterized in that, The step of performing tensor decomposition on the self-similarity tensor, extracting the time intensity vector of the atomic self-similarity mode corresponding to the fault mode, and fusing them into a decoupled self-similarity feature vector includes: Perform normalized parallel factorization on the self-similarity tensor to obtain a core tensor and multiple factor matrices; From the factor matrix associated with the channel-scale combination type dimension, select a number of pre-defined column vectors related to the target fault mechanism. Each column of the vector represents the temporal intensity distribution of an atomic self-similar mode. Calculate the statistical characteristics of each column of the vector, including peak factor, permutation entropy, and main periodic intensity; All the calculated statistical features are concatenated in sequence to form a one-dimensional numerical vector, which is the decoupled self-similar feature vector.

6. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 1, characterized in that, The step of performing weighted distance calculation and hierarchical comparison between the decoupled self-similar feature vector and the pre-stored standard fault feature template, and outputting a fault type identification conclusion based on the comparison result, includes: Calculate the weighted Mahalanobis distance between the decoupled self-similar feature vector and each standard fault feature template in the feature template library, wherein the feature template library stores standard feature vectors and weight vectors corresponding to known fault types; All standard fault feature templates whose weighted Mahalanobis distance is less than a first preset threshold are selected to form a candidate fault set; If the candidate fault set contains only one element, then output the fault type corresponding to this element; If the candidate fault set contains multiple elements, then calculate the sum of the absolute deviations between the decoupled self-similar feature vector and each element in the candidate fault set on its top K feature components with the highest weights, and output the fault type corresponding to the minimum sum of deviations. If the candidate fault set is empty, then an unknown fault status identifier is output.

7. The waveform self-similarity fault type identification method based on Tanimoto coefficient according to claim 6, characterized in that, The step of calculating the weighted Mahalanobis distance between the decoupled self-similar feature vector and each standard fault feature template in the feature template library includes: reading a standard fault feature template stored in the feature template library, wherein the standard fault feature template includes a standard feature vector and a corresponding weight vector; Calculate the difference between the decoupled self-similar feature vector and the standard feature vector in each feature dimension; Obtain the inverse of a pre-computed global covariance matrix; The weight vector is used to weight each component of the difference; A scalar distance value is calculated using the Mahalanobis distance formula based on the weighted difference vector and the inverse of the global covariance matrix; the scalar distance value is the weighted Mahalanobis distance between the decoupled self-similar feature vector and the current standard fault feature template.

8. A waveform self-similarity fault type identification system based on Tanimoto coefficients, characterized in that, include: The acquisition module is used to acquire multi-channel time-series waveform data, preprocess the waveform data and divide it into multiple scales, and generate a multi-scale waveform feature primitive set containing time-domain, frequency-domain and time-frequency domain feature vectors. The calculation module is used to calculate the generalized Tanimoto coefficients between different physical channels and different analysis scales based on the multi-scale waveform feature primitive set, and to construct cross-scale and cross-channel self-similarity tensors. The decomposition module is used to perform tensor decomposition on the self-similarity tensor, extract the time intensity vector of the atomic self-similarity mode corresponding to the fault mode, and fuse them into a decoupled self-similarity feature vector; the comparison module is used to perform weighted distance calculation and hierarchical comparison between the decoupled self-similarity feature vector and the pre-stored standard fault feature template, and output the fault type identification conclusion based on the comparison result.

Images