Early Fault Detection Method for Equipment Based on Fast Spectral Kujicism and Time Series Large Model

Abstract

This invention discloses an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. First, the equipment to be detected is selected. During equipment operation, sensors record the real-time operating signals F of key components. Then, signal F is divided into N segments, and further subdivided into N segments by optimizing the objective function ξ. D The normalized sub-signals are normalized to a set value μ, thus obtaining the normalized signal FN of F. The signal FN is decomposed into different frequency bands, and the frequency band that maximizes the LSESG value is selected for filtering to obtain the filtered signal FNF of signal FN. Finally, FNF is randomly divided into K sub-signals and data enhancement is performed to output the fault detection result. This invention helps to identify early equipment faults and provides theoretical support and decision-making basis for intelligent operation and maintenance and health management of equipment.

Description

Technical Field

[0001] This invention belongs to the field of early equipment fault detection technology, specifically involving a method for early equipment fault detection based on fast spectral kurtosis and a large time series model. Background Technology

[0002] Equipment failures not only cause resource waste and safety hazards, but the extent of loss also increases exponentially with the severity of the failure. Therefore, timely detection of early equipment failures is essential to effectively reduce losses. Large-scale time series models, through text alignment and time series embedding, demonstrate advantages such as automation and high accuracy in equipment failure detection. However, early failure signals are easily masked by noise, affecting the detection accuracy of large-scale time series models. The FastKurtogram (FK) algorithm is used to filter failure noise components; however, the filtering performance of the FK algorithm is also susceptible to strong noise interference. Therefore, this invention proposes an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. It utilizes Multiple Greedy Gaussian Segmentation (MGGS) to normalize equipment signals, improving overall fault detection efficiency. A combined index (LSESGASG) of Log-Squared Envelope Spectrum Gini Coefficient (LSESGC) and Spectral Gini Coefficient (SGC) is used instead of the kurtosis criterion in the FK algorithm, improving the filtering accuracy of the FK algorithm. Finally, this invention proposes using a highly accurate large time series model for fault detection, thereby improving the accuracy of early fault detection for equipment. Summary of the Invention

[0003] The purpose of this invention is to provide a method for early equipment fault detection based on fast spectral kurtosis and time series large models, which helps to identify early equipment faults and provides theoretical support and decision-making basis for intelligent equipment operation and maintenance and health management.

[0004] The technical solution adopted in this invention is an early fault detection method for equipment based on fast spectral kurtosis and a large time series model, characterized by the following steps:

[0005] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0006] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0007] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0008] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0009] The present invention is also characterized in that:

[0010] Step 1 is implemented in the following steps:

[0011] Select the equipment for which fault detection is required, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signal F of the key components. The matrix representation of F is as follows:

[0012]

[0013] Among them, X j ∈R n Let represent the real-time signal collected by the j-th sensor, m represent the total number of sensors, n represent the total signal value of F, and x represent the total signal value of F. ji This represents the i-th signal collected by the j-th sensor.

[0014] Step 2 is implemented in the following steps:

[0015] Step 2.1: Divide F into N segments, with t as the signal quantity in each segment. Let the set of signals after the division be denoted as . For the i-th sub-segment, give Matrix representation:

[0016]

[0017] Step 2.2, Definition Divided by K points a1, a2, ..., a K Divide into K+1 segments, let ω f represent The f-th segment in ω f N(μ) follows a Gaussian distribution f ,∑ f The Gaussian parameters are optimized according to the objective function ξ constructed by formula (3), and then F is further divided to obtain N. D Individual signals:

[0018]

[0019] Where δ represents the regularization parameter, C = -(mt / 2)(log(2π)+1), S f ω f The empirical covariance;

[0020] Steps 2.3 and 2.2 yield N. D To normalize the signal, a fixed mean μ is set for each sub-signal. The mean of each sub-signal is normalized to μ, thus obtaining the normalized signal FN = [Y1, Y2, ..., Y]. j ,...,Y m ] T , where Y j Represents X j The standardized signal.

[0021] Step 3 is implemented in the following steps:

[0022] Step 3.1: Obtain the normalized signal FN = [Y1, Y2, ..., Y] from Step 2. j ,...,Y m ] T Construct two quasi-analytical low-pass and high-pass analysis filters h0(n) and h1(n) for Y. j Perform multi-level decomposition of j = 1, 2, ..., m, and let... This represents the frequency band signal filtered by the i-th filter in the i-th layer;

[0023] Step 3.2, for the signal Perform a Hilbert transform to obtain Analyzed signal

[0024]

[0025] Where ι is the imaginary unit, yes Hilbert transform;

[0026] Step 3.3: Based on the analytical signal obtained in Step 3.2 calculate envelope

[0027]

[0028] Step 3.4: Based on the envelope obtained in Step 3.3 calculate Logarithmic square envelope

[0029]

[0030] Where τ is a very small constant value, used to avoid zero values ​​in the logarithmic function;

[0031] Step 3.5: Based on the logarithmic square envelope obtained in Step 3.4 calculate The log-squared envelope spectrum Gini coefficient LSESGC:

[0032]

[0033] Where, vector for The ascending sequence, R is semaphores;

[0034] Step 3.6: Set the number of points N for the Fast Fourier Transform. FFT N FFT Greater than or equal to the signal length R, and N FFT Powers of 2:

[0035]

[0036] Where FFT stands for Fast Fourier Transform;

[0037] Step 3.7, Signal Perform a Fast Fourier Transform to obtain the frequency domain signal c. f :

[0038]

[0039] Step 3.8: Calculate the frequency domain signal c f Normalized amplitude spectrum

[0040]

[0041] Step 3.9: Based on the normalized amplitude spectrum obtained in Step 3.8 Calculate signal Spectral Gini coefficients SGC:

[0042]

[0043] Where, vector for An ascending sequence;

[0044] Step 3.10: Obtain the frequency band signal from steps 3.5 and 3.9. The log-squared envelope spectral Gini coefficient LSESGC and the spectral Gini coefficient SGC are used to calculate the combined index LSESGASG of LSESGC and SGC:

[0045]

[0046] Among them, LSESGC reflects the sparsity characteristics of the signal envelope, and SGC uses frequency domain information to extract the time-frequency characteristics of the signal. Since the spectrum reflects the center frequency range of the signal, the weight of LSESGC in LSESGASG should be appropriately reduced to maintain the accuracy of the spectrum information.

[0047] Step 3.11: Select the frequency band signal that maximizes LSESG. Obtain the corresponding level and center frequency f c Filter Y j Signal Z is obtained j Thus, the filtered signal FNF = [Z1, Z2, ..., Zn] is obtained. j ,...,Z m ] T ;

[0048] Step 4 is implemented in the following steps:

[0049] Step 4.1: Randomly divide the FNF into K sub-signals and perform data augmentation to generate a signal set including the K sub-signals and their positive and negative signals.

[0050] Step 4.2, compress u i Embedding e for low-dimensional encoder i ;

[0051] Step 4.3: Embed the encoder in e i Convert to text embedding

[0052] Step 4.4: Construct soft hints (pe) as guiding information for the signal FNF to assist the large model in fault detection;

[0053] Step 4.5: Select BERT as the fault detection model;

[0054] Step 4.6: Define the loss function to optimize model performance.

[0055] Step 4.1 is as follows:

[0056] Randomly divide FNF into K non-overlapping sub-signals. s k Let represent the k-th sub-signal. For sub-signal s... k Perform data augmentation to generate positive signals (to s) k Add random noise or divide s k (Segments of random length, randomly reassembled) and negative signals (s k The non-overlapping signals are ultimately used to generate sub-signals s. k The signal set including its positive and negative signals u i This represents the i-th sub-signal in the signal set u.

[0057] Step 4.2 is as follows:

[0058] Self-encoder f ed By encoder f e With decoder f d It consists of two parts, for the signal set obtained in step 4.1 encoder f e compressed signal u i To an embedding space with M features and through decoder f d Generate reconstructed signal f d (e i Embedded space e i The signal embedding instance e should be made close to the positive signal e. + At the same time, stay away from negative signals e - Furthermore, to avoid an excessively small embedding space, the embedding space e should be ensured. i There are sufficient differences between the various features to fully cover the diverse information of the signal.

[0059] Step 4.3 is as follows:

[0060] Large models are inherently text-oriented and lack the ability to handle signal processing tasks. Therefore, P text embeddings t are selected. i i = 1, 2, ..., p is used as the prototype, and the text is embedded by alignment. p With signal embedding space e i This activates the potential of large models to process signal tasks, thereby obtaining the text embedding of the signal FNF as...

[0061] Step 4.5 is as follows:

[0062] BERT was selected as the fault detection model. The specific architecture of BERT includes an input layer, multi-layer Transformer encoders, pooling layers, and an output layer. Each Transformer encoder layer includes a self-attention calculation module, a residual connection and layer normalization module, and a feedforward neural network processing module. The specific steps for using BERT for device fault detection are as follows:

[0063] First, the associated signal text embedding With soft prompt PE Transformation via word embedding matrix of BERT input layer It is a high-dimensional semantic vector, and an absolute position encoding technique is used to generate a feature representation with temporal correlation;

[0064] Then, after the feature representation establishes global dependencies through multiple Transformer encoders, it goes through a pooling layer and the global representation vector is extracted through the hidden state of the special classification label [CLS].

[0065] Finally, the global representation vector undergoes a linear transformation via a fully connected classifier in the output layer, and a fault type probability distribution is generated using a Softmax function to output the fault detection result. For healthy state H or fault state {F1,F2,...,F...} p One of them.

[0066] Step 4.6 is as follows:

[0067] Define the following loss function to optimize model performance:

[0068] (1) Autoencoder loss Used to optimize autoencoder f ed The performance is as shown in formula (13):

[0069]

[0070] Where sim is the similarity function, and B represents the total number of inputs to the autoencoder;

[0071] (2) Comparison of losses Used to optimize the embedding space e k The distance between the signals is given by formula (14):

[0072]

[0073] Among them, σ(e,e + / - ) is used to measure e and e + / - Similarity between them;

[0074] (3) Embedding loss Used to optimize the embedding space e k Different characteristics, such as in formula (15):

[0075]

[0076] Where, m i This represents the i-th feature of instance m. and This represents the corresponding characteristics of positive and negative signals m;

[0077] (4) Alignment loss Used to align text embeddings p With signal embedding space e i As shown in formula (16):

[0078]

[0079] Where, sim(t) i ,e) indicates the alignment item, For comparison;

[0080] (5) Soft cue embedding loss For optimizing soft suggestions (PE), as in formula (17):

[0081]

[0082] The beneficial effects of this invention are as follows: An early fault detection method for equipment based on fast spectral kurtosis and a large time series model. To improve overall fault detection efficiency, this invention utilizes multiple greedy Gaussian segmentation to normalize equipment signals. To enhance the accuracy and reliability of early fault detection, this invention leverages the advantages of large time series models, such as automation and high accuracy, and proposes combining large time series models for fault detection. Furthermore, this invention replaces the kurtosis criterion in the FK algorithm with a comprehensive index of the logarithmic squared envelope spectral Gini coefficient and the spectral Gini coefficient, improving the filtering accuracy of the algorithm and thus enhancing the fault detection accuracy of the large model. Attached Figure Description

[0083] Figure 1 This is the overall flowchart of the device early fault detection method based on fast spectral kurtosis and time series large model of the present invention;

[0084] Figure 2 This is a diagram illustrating the signal normalization and filtering framework of the device early fault detection method based on fast spectral kurtosis and a large time series model, as presented in this invention.

[0085] Figure 3 This invention provides an example of an early fault detection method for equipment based on a fast spectral kurtosis and a large time series model: vibration signal of an inner ring fault in a mechanical bearing.

[0086] Figure 4 This is an example of the device early fault detection method based on fast spectral kurtosis and time series large model of the present invention, which is a normalized mechanical bearing inner ring fault vibration signal;

[0087] Figure 5 This invention provides an example of an early fault detection method for equipment based on fast spectral kurtosis and a large time series model, comparing the filtered signal with the original signal.

[0088] Figure 6This is the envelope spectrum of an example filtered signal of the device early fault detection method based on fast spectral kurtosis and a large time series model according to the present invention. Detailed Implementation

[0089] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0090] In the early stages of equipment fault detection, real-time signals from key components during equipment operation are typically collected by various sensors. These signals are presented as a continuous time series. To improve the accuracy and efficiency of fault detection, the signals need to be properly normalized and filtered to remove noise interference. Furthermore, due to the weak impact of early faults, effective fault detection methods should be employed. The main detection process is as follows: First, select the equipment to be fault-detected. During equipment operation, sensors record the real-time operating signals F of key components (such as bearings, gears, etc.). Second, divide signal F into N segments, and further subdivide F into N segments by optimizing the objective function ξ. D The normalized sub-signals are normalized to a set value μ, thus obtaining the normalized signal FN of F. Then, signal FN is decomposed into different frequency bands, and the frequency band that maximizes the LSEGASG value is selected for filtering, resulting in the filtered signal FNF of signal FN. Finally, FNF is randomly divided into K sub-signals and data augmentation is performed to output the fault detection result.

[0091] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0092] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0093] Step 1 is implemented in the following steps:

[0094] Select the equipment for which fault detection is required, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signal F of the key components. The matrix representation of F is as follows:

[0095]

[0096] Among them, X j ∈R n Let represent the real-time signal collected by the j-th sensor, m represent the total number of sensors, n represent the total signal value of F, and x represent the total signal value of F. ji This represents the i-th signal collected by sensor j.

[0097] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0098] Step 2 is implemented in the following steps:

[0099] Step 2.1: Divide F into N segments, with t as the signal quantity in each segment. Let the set of signals after the division be denoted as . For the i-th sub-segment, give Matrix representation:

[0100]

[0101] Step 2.2, Definition Divided by K points a1, a2, ..., a K Divide into K+1 segments, let ω f represent The f-th segment in ω f N(μ) follows a Gaussian distribution f ,∑ f The Gaussian parameters are optimized according to the objective function ξ constructed by formula (3), and then F is further divided to obtain N. D Individual signals:

[0102]

[0103] Where δ represents the regularization parameter, C = -(mt / 2)(log(2π)+1), S f ω f The empirical covariance;

[0104] Steps 2.3 and 2.2 yield N. D To normalize the signal, a fixed mean μ is set for each sub-signal. The mean of each sub-signal is normalized to μ, thus obtaining the normalized signal FN = [Y1, Y2, ..., Y]. j ,...,Y m ] T , where Y j X represents j The standardized signal.

[0105] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0106] Step 3 is implemented in the following steps:

[0107] Step 3.1: Obtain the normalized signal FN = [Y1, Y2, ..., Y] from Step 2. j ,...,Y m ] T Construct two quasi-analytical low-pass and high-pass analysis filters h0(n) and h1(n) for Y. j Perform multi-level decomposition of j = 1, 2, ..., m, and let... This represents the frequency band signal filtered by the i-th filter in the i-th layer;

[0108] Step 3.2, for the signal Perform a Hilbert transform to obtain Analyzed signal

[0109]

[0110] Where ι is the imaginary unit, yes Hilbert transform;

[0111] Step 3.3: Based on the analytical signal obtained in Step 3.2 calculate envelope

[0112]

[0113] Step 3.4: Based on the envelope obtained in Step 3.3 calculate Logarithmic square envelope

[0114]

[0115] Where τ is a very small constant value, used to avoid zero values ​​in the logarithmic function;

[0116] Step 3.5: Based on the logarithmic square envelope obtained in Step 3.4 calculate The log-squared envelope spectrum Gini coefficient LSESGC:

[0117]

[0118] Where, vector for The ascending sequence, R is semaphores;

[0119] Step 3.6: Set the number of points N for the Fast Fourier Transform. FFT N FFTGreater than or equal to the signal length R, and N FFT Powers of 2:

[0120]

[0121] Where FFT stands for Fast Fourier Transform;

[0122] Step 3.7, Signal Perform a Fast Fourier Transform to obtain the frequency domain signal c. f :

[0123]

[0124] Step 3.8: Calculate the frequency domain signal c f Normalized amplitude spectrum

[0125]

[0126] Step 3.9: Based on the normalized amplitude spectrum obtained in Step 3.8 Calculate signal Spectral Gini coefficients SGC:

[0127]

[0128] Where, vector for An ascending sequence;

[0129] Step 3.10: Obtain the frequency band signal from steps 3.5 and 3.9. The log-squared envelope spectral Gini coefficient LSESGC and the spectral Gini coefficient SGC are used to calculate the combined index LSESGASG of LSESGC and SGC:

[0130]

[0131] Among them, LSESGC reflects the sparsity characteristics of the signal envelope, and SGC uses frequency domain information to extract the time-frequency characteristics of the signal. Since the spectrum reflects the center frequency range of the signal, the weight of LSESGC in LSESGASG should be appropriately reduced to maintain the accuracy of the spectrum information.

[0132] Step 3.11: Select the frequency band signal that maximizes LSESG. Obtain the corresponding level and center frequency f c Filter Y j Signal Z is obtained j Thus, the filtered signal FNF = [Z1, Z2, ..., Zn] is obtained. j ,...,Zm ] T ;

[0133] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0134] Step 4 is implemented in the following steps:

[0135] Step 4.1: Randomly divide the FNF into K sub-signals and perform data augmentation to generate a signal set including the K sub-signals and their positive and negative signals.

[0136] Step 4.1 is as follows:

[0137] Randomly divide FNF into K non-overlapping sub-signals. s k Let represent the k-th sub-signal. For sub-signal s... k Perform data augmentation to generate positive signals (to s) k Add random noise or divide s k (Segments of random length, randomly reassembled) and negative signals (s k The non-overlapping signals are ultimately used to generate sub-signals s. k The signal set including its positive and negative signals u i This represents the i-th sub-signal in the signal set u.

[0138] Step 4.2, compress u i Embedding e for low-dimensional encoder i ;

[0139] Step 4.2 is as follows:

[0140] Self-encoder f ed By encoder f e With decoder f d It consists of two parts, for the signal set obtained in step 4.1 encoder f e compressed signal u i To an embedding space with M features and through decoder f d Generate reconstructed signal f d (e i Embedded space e i The signal embedding instance e should be made close to the positive signal e. + At the same time, stay away from negative signals e - Furthermore, to avoid an excessively small embedding space, the embedding space e should be ensured. iThere are sufficient differences between the various features to fully cover the diverse information of the signal.

[0141] Step 4.3: Embed the encoder in e i Convert to text embedding

[0142] Step 4.3 is as follows:

[0143] Large models are inherently text-oriented and lack the ability to handle signal processing tasks. Therefore, P text embeddings t are selected. i i = 1, 2, ..., p is used as the prototype, and the text is embedded by alignment. p With signal embedding space e i This activates the potential of large models to process signal tasks, thereby obtaining the text embedding of the signal FNF as...

[0144] Step 4.4: Construct soft hints (pe) as guiding information for the signal FNF to assist the large model in fault detection;

[0145] Step 4.5: Select BERT as the fault detection model;

[0146] Step 4.5 is as follows:

[0147] BERT was selected as the fault detection model. The specific architecture of BERT includes an input layer, multi-layer Transformer encoders, pooling layers, and an output layer. Each Transformer encoder layer includes a self-attention calculation module, a residual connection and layer normalization module, and a feedforward neural network processing module. The specific steps for using BERT for device fault detection are as follows:

[0148] First, the associated signal text embedding With soft prompt PE Transformation via word embedding matrix of BERT input layer It is a high-dimensional semantic vector, and an absolute position encoding technique is used to generate a feature representation with temporal correlation;

[0149] Then, after the feature representation establishes global dependencies through multiple Transformer encoders, it goes through a pooling layer and the global representation vector is extracted through the hidden state of the special classification label [CLS].

[0150] Finally, the global representation vector undergoes a linear transformation via a fully connected classifier in the output layer, and a fault type probability distribution is generated using a Softmax function to output the fault detection result. For healthy state H or fault state {F1,F2,...,F...} p One of them.

[0151] Step 4.6: Define the loss function to optimize model performance.

[0152] Step 4.6 is as follows:

[0153] Define the following loss function to optimize model performance:

[0154] (1) Autoencoder loss Used to optimize autoencoder f ed The performance is as shown in formula (13):

[0155]

[0156] Where sim is the similarity function, and B represents the total number of inputs to the autoencoder;

[0157] (2) Comparison of losses Used to optimize the embedding space e k The distance between the signals is given by formula (14):

[0158]

[0159] Among them, σ(e,e + / - ) is used to measure e and e + / - Similarity between them;

[0160] (3) Embedding loss Used to optimize the embedding space e k Different characteristics, such as in formula (15):

[0161]

[0162] Where, m i This represents the i-th feature of instance m. and This represents the corresponding characteristics of positive and negative signals m;

[0163] (4) Alignment loss Used to align text embeddings p With signal embedding space e i As shown in formula (16):

[0164]

[0165] Where, sim(t) i ,e) indicates the alignment item, For comparison;

[0166] (5) Soft cue embedding loss For optimizing soft suggestions (PE), as in formula (17):

[0167]

[0168] Example 1

[0169] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0170] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0171] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0172] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0173] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0174] Example 2

[0175] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0176] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0177] Step 1 is implemented in the following steps:

[0178] Select the equipment for which fault detection is required, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signal F of the key components. The matrix representation of F is as follows:

[0179]

[0180] Among them, X j ∈R n Let represent the real-time signal collected by the j-th sensor, m represent the total number of sensors, n represent the total signal value of F, and x represent the total signal value of F. ji This represents the i-th signal collected by sensor j.

[0181] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0182] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0183] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0184] Example 3

[0185] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0186] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0187] Step 1 is implemented in the following steps:

[0188] Select the equipment for which fault detection is required, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signal F of the key components. The matrix representation of F is as follows:

[0189]

[0190] Among them, X j ∈R n Let represent the real-time signal collected by the j-th sensor, m represent the total number of sensors, n represent the total signal value of F, and x represent the total signal value of F. ji This represents the i-th signal collected by sensor j.

[0191] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0192] Step 2 is implemented in the following steps:

[0193] Step 2.1: Divide F into N segments, with t as the signal quantity in each segment. Let the set of signals after the division be denoted as . For the i-th sub-segment, give Matrix representation:

[0194]

[0195] Step 2.2, Definition Divided by K points a1, a2, ..., a K Divide into K+1 segments, let ω f represent The f-th segment in ω f N(μ) follows a Gaussian distribution f ,∑ f The Gaussian parameters are optimized according to the objective function ξ constructed by formula (3), and then F is further divided to obtain N. D Individual signals:

[0196]

[0197] Where δ represents the regularization parameter, C = -(mt / 2)(log(2π)+1), S f ω f The empirical covariance;

[0198] Steps 2.3 and 2.2 yield N. D To normalize the signal, a fixed mean μ is set for each sub-signal. The mean of each sub-signal is normalized to μ, thus obtaining the normalized signal FN = [Y1, Y2, ..., Y]. j ,...,Y m ] T , where Y j Represents X j The standardized signal.

[0199] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0200] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0201] Example 4

[0202] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0203] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0204] Step 1 is implemented in the following steps:

[0205] Select the equipment for which fault detection is required, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signal F of the key components. The matrix representation of F is as follows:

[0206]

[0207] Among them, X j ∈R n Let represent the real-time signal collected by the j-th sensor, m represent the total number of sensors, n represent the total signal value of F, and x represent the total signal value of F. ji This represents the i-th signal collected by sensor j.

[0208] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0209] Step 2 is implemented in the following steps:

[0210] Step 2.1: Divide F into N segments, with t as the signal quantity in each segment. Let the set of signals after the division be denoted as . For the i-th sub-segment, give Matrix representation:

[0211]

[0212] Step 2.2, Definition Divided by K points a1, a2, ..., a K Divide into K+1 segments, let ω f represent The f-th segment in ω f N(μ) follows a Gaussian distribution f ,∑ f The Gaussian parameters are optimized according to the objective function ξ constructed by formula (3), and then F is further divided to obtain N. D Individual signals:

[0213]

[0214] Where δ represents the regularization parameter, C = -(mt / 2)(log(2π)+1), S f ω f The empirical covariance;

[0215] Steps 2.3 and 2.2 yield N. DTo normalize the signal, a fixed mean μ is set for each sub-signal. The mean of each sub-signal is normalized to μ, thus obtaining the normalized signal FN = [Y1, Y2, ..., Y]. j ,...,Y m ] T , where Y j Represents X j The standardized signal.

[0216] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0217] Step 3 is implemented in the following steps:

[0218] Step 3.1: Obtain the normalized signal FN = [Y1, Y2, ..., Y] from Step 2. j ,...,Y m ] T Construct two quasi-analytical low-pass and high-pass analysis filters h0(n) and h1(n) for Y. j Perform multi-level decomposition of j = 1, 2, ..., m, and let... This represents the frequency band signal filtered by the i-th filter in the i-th layer;

[0219] Step 3.2, for the signal Perform a Hilbert transform to obtain Analyzed signal

[0220]

[0221] Where ι is the imaginary unit, yes Hilbert transform;

[0222] Step 3.3: Based on the analytical signal obtained in Step 3.2 calculate envelope

[0223]

[0224] Step 3.4: Based on the envelope obtained in Step 3.3 calculate Logarithmic square envelope

[0225]

[0226] Where τ is a very small constant value, used to avoid zero values ​​in the logarithmic function;

[0227] Step 3.5: Based on the logarithmic square envelope obtained in Step 3.4 calculate The log-squared envelope spectrum Gini coefficient LSESGC:

[0228]

[0229] Where, vector for The ascending sequence, R is semaphores;

[0230] Step 3.6: Set the number of points N for the Fast Fourier Transform. FFT N FFT Greater than or equal to the signal length R, and N FFT Powers of 2:

[0231]

[0232] Where FFT stands for Fast Fourier Transform;

[0233] Step 3.7, Signal Perform a Fast Fourier Transform to obtain the frequency domain signal c. f :

[0234]

[0235] Step 3.8: Calculate the frequency domain signal c f Normalized amplitude spectrum

[0236]

[0237] Step 3.9: Based on the normalized amplitude spectrum obtained in Step 3.8 Calculate signal Spectral Gini coefficients SGC:

[0238]

[0239] Where, vector for An ascending sequence;

[0240] Step 3.10: Obtain the frequency band signal from steps 3.5 and 3.9. The log-squared envelope spectral Gini coefficient LSESGC and the spectral Gini coefficient SGC are used to calculate the combined index LSESGASG of LSESGC and SGC:

[0241]

[0242] Among them, LSESGC reflects the sparsity characteristics of the signal envelope, and SGC uses frequency domain information to extract the time-frequency characteristics of the signal. Since the spectrum reflects the center frequency range of the signal, the weight of LSESGC in LSESGASG should be appropriately reduced to maintain the accuracy of the spectrum information.

[0243] Step 3.11: Select the frequency band signal that maximizes LSESG. Obtain the corresponding level and center frequency f c Filter Y j Signal Z is obtained j Thus, the filtered signal FNF = [Z1, Z2, ..., Zn] is obtained. j ,...,Z m ] T ;

[0244] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0245] Example 5

[0246] This invention relates to an early fault detection method for equipment based on fast spectral kurtosis and a large time series model. The flowchart is as follows: Figure 1 As shown, please follow these steps:

[0247] Step 1: Select the equipment that needs to be fault detected. When the equipment is in operation, the sensors record the real-time operating signals F of the key components of the equipment.

[0248] Step 2: Divide the signal F into N segments, and further divide F into N segments by optimizing the objective function ξ. D Each sub-signal is normalized to a set value μ, thus obtaining the normalized signal FN of F;

[0249] Step 3: Decompose the signal FN into different frequency bands, select the frequency band that maximizes the LSEGASG value for filtering, and obtain the filtered signal FNF of signal FN;

[0250] Step 4: Randomly divide FNF into K sub-signals and perform data enhancement, then output the fault detection results.

[0251] Step 4 is implemented in the following steps:

[0252] Step 4.1: Randomly divide the FNF into K sub-signals and perform data augmentation to generate a signal set including the K sub-signals and their positive and negative signals.

[0253] Step 4.1 is as follows:

[0254] Randomly divide FNF into K non-overlapping sub-signals. s k Let represent the k-th sub-signal. For sub-signal s... k Perform data augmentation to generate positive signals (to s) k Add random noise or divide s k (Segments of random length, randomly reassembled) and negative signals (s k The non-overlapping signals are ultimately used to generate sub-signals s. k The signal set including its positive and negative signals u i This represents the i-th sub-signal in the signal set u.

[0255] Step 4.2, compress u i Embedding e for low-dimensional encoder i ;

[0256] Step 4.2 is as follows:

[0257] Self-encoder f ed By encoder f e With decoder f d It consists of two parts, for the signal set obtained in step 4.1 encoder f e compressed signal u i To an embedding space with M features and through decoder f d Generate reconstructed signal f d (e i Embedded space e i The signal embedding instance e should be made close to the positive signal e. + At the same time, stay away from negative signals e - Furthermore, to avoid an excessively small embedding space, the embedding space e should be ensured. i There are sufficient differences between the various features to fully cover the diverse information of the signal.

[0258] Step 4.3: Embed the encoder in e i Convert to text embedding

[0259] Step 4.3 is as follows:

[0260] Large models are inherently text-oriented and lack the ability to handle signal processing tasks. Therefore, P text embeddings t are selected. i i = 1, 2, ..., p is used as the prototype, and the text is embedded by alignment. p With signal embedding space e iThis activates the potential of large models to process signal tasks, thereby obtaining the text embedding of the signal FNF as...

[0261] Step 4.4: Construct soft hints (pe) as guiding information for the signal FNF to assist the large model in fault detection;

[0262] Step 4.5: Select BERT as the fault detection model;

[0263] Step 4.5 is as follows:

[0264] BERT was selected as the fault detection model. The specific architecture of BERT includes an input layer, multi-layer Transformer encoders, pooling layers, and an output layer. Each Transformer encoder layer includes a self-attention calculation module, a residual connection and layer normalization module, and a feedforward neural network processing module. The specific steps for using BERT for device fault detection are as follows:

[0265] First, the associated signal text embedding With soft prompt PE Transformation via word embedding matrix of BERT input layer It is a high-dimensional semantic vector, and an absolute position encoding technique is used to generate a feature representation with temporal correlation;

[0266] Then, after the feature representation establishes global dependencies through multiple Transformer encoders, it goes through a pooling layer and the global representation vector is extracted through the hidden state of the special classification label [CLS].

[0267] Finally, the global representation vector undergoes a linear transformation via a fully connected classifier in the output layer, and a fault type probability distribution is generated using a Softmax function to output the fault detection result. For healthy state H or fault state {F1,F2,...,F...} p One of them.

[0268] Step 4.6: Define the loss function to optimize model performance.

[0269] Step 4.6 is as follows:

[0270] Define the following loss function to optimize model performance:

[0271] (1) Autoencoder loss Used to optimize autoencoder f ed The performance is as shown in formula (13):

[0272]

[0273] Where sim is the similarity function, and B represents the total number of inputs to the autoencoder;

[0274] (2) Comparison of losses Used to optimize the embedding space e k The distance between the signals is given by formula (14):

[0275]

[0276] Among them, σ(e,e + / - ) is used to measure e and e + / - Similarity between them;

[0277] (3) Embedding loss Used to optimize the embedding space e k Different characteristics, such as in formula (15):

[0278]

[0279] Where, m i This represents the i-th feature of instance m. and This represents the corresponding characteristics of positive and negative signals m;

[0280] (4) Alignment loss Used to align text embeddings p With signal embedding space e i As shown in formula (16):

[0281]

[0282] Where, sim(t) i ,e) indicates the alignment item, For comparison;

[0283] (5) Soft cue embedding loss Used to optimize soft suggestions (PE), such as formula (17):

[0284]

[0285] like Figure 4 As shown, compared to Figure 3 The original signal in the original data, combined with the MGGSN segmentation algorithm, achieves reasonable signal normalization while preserving the basic fault characteristics. Meanwhile, as... Figure 5 and Figure 6As shown, this invention filters the normalized signal, enabling the extraction of fault signals with distinct characteristics. This filtering method improves the efficiency of subsequent fault detection. Furthermore, leveraging the advantages of large-scale time-series models, such as automation and high accuracy, this invention proposes combining large-scale time-series models for fault detection. When using relevant algorithms for fault detection in industrial equipment condition monitoring applications, combining the proposed large-scale time-series model fault detection method can effectively reduce equipment downtime for troubleshooting, significantly lowering the false positive rate while improving fault detection efficiency.

[0286] Example 6

[0287] To verify the feasibility of this invention, it is further described in conjunction with the embodiments and accompanying drawings. The data selected in this study is the Xi'an Jiaotong University Bearing Dataset (XJTU-SY). XJTU-SY contains the full life cycle vibration signals of 15 rolling bearings under three operating conditions. Two directional sensors are used to record the full life cycle vibration signals of the bearings in the horizontal and vertical directions, respectively. In XJTU-SY, the failure causes of bearings include inner ring wear, cage fracture, outer ring wear, and outer ring cracking. This invention selects bearing inner ring failure as an example to explain the effectiveness of the invention. The bearing inner ring failure acquisition signal F is shown below. Figure 3 As shown.

[0288] First, the bearing inner ring fault signal F is normalized by setting a fixed segment length of 4096, dividing F into multiple segments. Then, F is further subdivided by minimizing the objective function ξ. A fixed mean μ = 0.5 is set, and each segment is processed to make the mean of each segment equal to the set value μ, resulting in the normalized signal FN of F, as shown below. Figure 4 As shown, with Figure 3 In comparison, the normalized signal FN is more standardized, and the envelope spectrum characteristics of the signal acquired by the horizontal sensor are clearer. Then, the frequency band signal that maximizes LSESG is selected. Obtain the corresponding level and center frequency f c The values ​​are 2133.333 and 1.585 respectively. The filtered signal FNF is obtained by filtering FN. The comparison between the filtered signal FNF and the original signal F is shown below. Figure 5 As shown, compared to the original signal, the filtered signal effectively removes interference components and reflects more obvious inner-circle fault characteristics. The envelope spectrum of the filtered signal is shown below. Figure 6 As shown. Figure 6 In the middle, the characteristic frequency f of the inner ring fault i and twice the characteristic frequency, the frequency f rThe frequency and its harmonics are clearly distinguishable, and a significant difference between the characteristic frequency and the frequency is also visible. The filtering results demonstrate that this invention effectively captures the characteristics of fault signals and achieves reasonable filtering, laying the foundation for subsequent fault detection work and helping to improve the accuracy and efficiency of detection.

Claims

1. A method for early equipment fault detection based on fast spectral kurtosis and a large time series model, characterized in that, The specific steps are as follows: Step 1: Select the equipment that needs to be fault detected. During equipment operation, sensors record the real-time operating signals of key components. ; Step 2, transfer the signal Divided into Segment, and optimize the objective function Further division for Each sub-signal, the mean of the normalized sub-signal is a set value. Thus obtain Normalized signal ; Step 3: Decompose the signal Different frequency bands are selected for filtering to maximize the LSESG value, thus obtaining the signal. Filtered signal ; Step 3 is implemented in the following steps: Step 3.1: Obtain the signal from Step 2. Normalized signal Construct two quasi-analytical low-pass and high-pass analysis filters. and ,right Perform multi-level decomposition, assuming Indicates the first The layer consists of the first Each filter filters the frequency band signal; Step 3.2, for the signal Perform a Hilbert transform to obtain Analyzed signal : (4) in, The imaginary unit, yes Hilbert transform; Step 3.3: Based on the analytical signal obtained in Step 3.2 ,calculate envelope : (5) Step 3.4: Based on the envelope obtained in Step 3.3 ,calculate Logarithmic square envelope : (6) in, It is a very small constant value used to avoid zero values ​​in the logarithmic function; Step 3.5: Based on the logarithmic square envelope obtained in Step 3.4 ,calculate Log-squared envelope spectrum Gini coefficient : (7) Where, vector for ascending sequence, for semaphores; Step 3.6: Set the number of points for Fast Fourier Transform , Greater than or equal to signal length ,and Powers of 2: (8) in, Indicates Fast Fourier Transform; Step 3.7, Signal Perform a Fast Fourier Transform to obtain the frequency domain signal. : (9) Step 3.8: Calculate the frequency domain signal Normalized amplitude spectrum : (10) Step 3.9: Based on the normalized amplitude spectrum obtained in Step 3.8 Calculate signal Spectral Gini coefficient : (11) Where, vector for An ascending sequence; Step 3.10: Obtain the frequency band signal from steps 3.5 and 3.

9. Log-squared envelope spectrum Gini coefficient With spectral Gini coefficient ,calculate and Comprehensive indicators : (12) in, This reflects the sparsity characteristics of the signal envelope. Extracting the time-frequency features of a signal using frequency domain information; Step 3.11, Select the option that makes Maximum frequency band signal To obtain the corresponding layer number and center frequency , filtering Receive signal Thus obtain Filtered signal ; Step 4: Randomly divide for Each sub-signal is processed and data enhancement is performed to output the fault detection result. ; Step 4 is implemented in the following steps: Step 4.1, Random Segmentation for Each sub-signal is processed and data augmentation is performed to generate a result including... The signal set including individual sub-signals and their positive and negative signals ; Step 4.2, Compression Embedding for low-dimensional encoders ; Step 4.3: Embed the encoder Convert to text embedding ; Step 4.4: Build soft suggestions As a signal The guidance information assists the large model in fault detection; Step 4.5: Select BERT as the fault detection model; Step 4.6: Define the loss function to optimize model performance.

2. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 1, characterized in that, Step 1 is implemented in the following steps: Select the equipment that needs fault detection, fix one or more sensors to the key components of the selected equipment, and record the real-time operating signals of the key components. , The matrix representation is as follows: (1) in, , indicating that by the first Real-time signals collected by a sensor Indicates the total number of sensors. express The total amount of signal, Indicates sensor The first collection One signal.

3. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 2, characterized in that, Step 2 is implemented in the following steps: Step 2.1, Division for Segments, each segment's semaphore is Let the set of signals after partitioning be ? , For the division of the first Each sub-segment is given. Matrix representation: (2) Step 2.2, Definition quilt Dividing points Divided into Duan, Ling represent The first in Each segment , Follows a Gaussian distribution The objective function constructed according to formula (3) Optimize Gaussian parameters, and then... To further divide, in order to obtain Individual signals: (3) in, Represents the regularization parameter. , for The empirical covariance; Steps 2.3 and 2.2 yielded the following results: For each sub-signal, a fixed mean is set to be used for signal normalization. The mean of each sub-signal is normalized to be 1. Thus, the signal is obtained. Normalized signal ,in, express The standardized signal.

4. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 3, characterized in that, Step 4.1 is as follows: Random partitioning for Non-overlapping sub-signals , Indicates the first Individual signal, pair signal Perform data augmentation to generate positive signals and negative signals Ultimately, this generates sub-signals. The signal set including its positive and negative signals , Represents signal set The first in Sub-signals.

5. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 4, characterized in that, Step 4.2 is as follows: Self-encoder By encoder With decoder It consists of two parts, for the signal set obtained in step 4.1 encoder compressed signal To have Embedding space of features and through the decoder Generate reconstructed signal Embedded space The signal should be embedded in the instance. Approaching positive signal At the same time, stay away from negative signals Furthermore, to avoid an excessively small embedding space, the embedding space should be ensured. There are sufficient differences between the various features to fully cover the diverse information of the signal.

6. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 5, characterized in that, Step 4.3 is as follows: Large models are text-oriented and lack the ability to process signals; therefore, we choose... Text embedding As a prototype, through aligned text embedding With signal embedding space This activates the potential of large models for signal processing tasks, thereby acquiring signals. The text embedding is .

7. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 6, characterized in that, Step 4.5 is as follows: BERT was selected as the fault detection model. The specific architecture of BERT includes an input layer, multi-layer Transformer encoders, pooling layers, and an output layer. Each Transformer encoder layer includes a self-attention calculation module, a residual connection and layer normalization module, and a feedforward neural network processing module. The specific steps for using BERT for device fault detection are as follows: First, the associated signal text embedding With soft prompts for Transformation through the word embedding matrix of the BERT input layer It is a high-dimensional semantic vector, and an absolute position encoding technique is used to generate a feature representation with temporal correlation; Then, after the feature representation establishes global dependencies through multiple Transformer encoders, it goes through a pooling layer and the global representation vector is extracted through the hidden state of the special classification label [CLS]. Finally, the global representation vector undergoes a linear transformation via a fully connected classifier in the output layer, and a fault type probability distribution is generated using a Softmax function to output the fault detection result. , For a healthy state or fault status One of them.

8. The method for early equipment fault detection based on fast spectral kurtosis and a large time series model according to claim 7, characterized in that, Step 4.6 is as follows: Define the following loss function to optimize model performance: (1) Self-encoder loss Used to optimize autoencoders The performance is as shown in formula (13): (13) in, For similarity function, This indicates the total number of inputs to the autoencoder; (2) Comparison of losses Used to optimize embedding space The distance between the signals is given by formula (14): (14) in, Used to measure and Similarity between them; (3) Embedding loss Used to optimize embedding space Different characteristics, such as in formula (15): (15) in, Representation of instances The One characteristic, and express Corresponding characteristics of positive and negative signals; (4) Alignment loss Used to align text embeddings With signal embedding space For example, in formula (16): (16) in, Indicates alignment item, For comparison; (5) Soft cue embedding loss Used to optimize soft suggestions For example, in formula (17): (17)。

Images