A vibration signal analysis knowledge-driven rotating machinery zero-sample intelligent fault detection method and device
By constructing a multi-index evaluation mechanism and a high-dimensional feature extractor based on deep neural networks, combined with a contrastive loss function, the problem of fault detection under zero-fault samples of rotating machinery was solved, achieving efficient fault detection results.
Patent Information
- Application Number
- CN202511357955.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2026-01-09
- Estimated Expiration
- 2045-09-23
AI Technical Summary
Existing technologies struggle to establish effective intelligent fault detection models in rotating machinery fault detection due to the lack of large-scale, high-quality fault samples, especially in the case of zero fault samples, making accurate detection difficult.
By constructing a multi-index evaluation mechanism to screen time-frequency analysis methods, building a time-frequency analysis knowledge base, using a deep neural network to construct a high-dimensional feature extractor, and combining it with a contrastive loss function for training, high-dimensional feature extraction is achieved, and similarity judgment is performed using normal state samples.
It enables accurate detection of the condition of rotating machinery under zero-fault sample conditions, improving the accuracy and reliability of fault detection.
Smart Images
Figure CN120873554B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent fault detection of rotating machinery, and in particular to a vibration signal analysis knowledge driven intelligent fault detection method and device for rotating machinery with zero samples. BACKGROUND
[0002] Rotating machinery (such as bearings, gears, etc.) is a crucial component in industrial equipment, widely used in key fields such as energy power and rail transportation. Its stable operation is crucial for production safety and efficiency. However, due to factors such as complex working conditions, long-term continuous operation, and aging of parts, rotating machinery is prone to typical faults such as bearing wear, rotor imbalance, and centering deterioration, leading to downtime losses and even safety accidents. Therefore, there is an urgent need to achieve early and accurate fault detection. In recent years, data-driven intelligent fault detection technology has shown significant potential due to the advantages of deep learning models in feature mining. However, this method highly depends on large-scale and high-quality fault samples for training. However, in actual industrial scenarios, rotating machinery faults occur infrequently and are not fully covered, making it extremely costly to obtain complete fault sample sets. This leads to a training dilemma for deep learning models due to sample imbalance, especially when there are zero fault samples, making it difficult to establish an intelligent fault detection model for rotating machinery. SUMMARY
[0003] Therefore, the purpose of the present application is to provide a vibration signal analysis knowledge driven intelligent fault detection method and device for rotating machinery with zero samples, which utilizes the time series continuity prior knowledge of rotating machinery continuous normal state data. It is assumed that the normal state samples taken by adjacent time windows are similar, and then the time-frequency spectrum obtained by the same time-frequency transformation also has similarity. Secondly, combining the different analysis angles of signals by different time-frequency analysis methods, it is assumed that the time-frequency spectrum obtained by the same signal through different time-frequency transformation operations does not have similarity. Thirdly, based on the comparison relationship of the multi-modal time-frequency spectrum of the above assumptions, a comparison loss function is constructed to drive the parameter optimization of the high-dimensional feature extractor. Finally, the sample pair composed of the test sample and the normal reference sample is input into the high-dimensional feature extractor which has been trained, and the high-dimensional feature vector is output. A fault detection matrix is designed to detect the state of the rotating machinery.
[0004] To achieve the above object, the application adopts the following technical scheme: a vibration signal analysis knowledge-driven rotating machinery zero-sample intelligent fault detection method, adopts a multi-index evaluation mechanism to screen multiple time-frequency analysis methods and construct a time-frequency analysis knowledge base, performs multi-angle time-frequency analysis on the rotating machinery normal state monitoring signal, and simultaneously constructs a comparison loss function based on the obtained multi-modal time-frequency spectrum; a high-dimensional feature extractor is constructed based on a deep neural network, and the high-dimensional feature extractor is trained based on the comparison loss function; the real-time state signal of the rotating machinery to be detected is subjected to time-frequency analysis through the constructed time-frequency analysis knowledge base, then the trained high-dimensional feature extractor is used for high-dimensional feature extraction, and the multiple time-frequency spectra of the normal state signal are compared to perform similarity threshold judgment, and the abnormal state of the rotating machinery is determined according to the judgment result.
[0005] In a preferred embodiment, it specifically comprises the following steps:
[0006] Step 1: Collecting rotating machinery continuous normal state monitoring signals, and intercepting samples from the signals through a sliding window as training samples;
[0007] Step 2: Generating multiple time-frequency spectra from the normal state samples through multiple time-frequency analysis methods, and quantitatively scoring the multiple time-frequency spectra by using an evaluation mechanism composed of three indexes of Prewitt Operator operator, signal-to-noise ratio (SNR) and frequency band energy (FBE), and selecting multiple time-frequency analysis methods through sorting as a time-frequency analysis knowledge base of the rotating machinery monitoring signal;
[0008] Step 3: Selecting adjacent sliding window intercepted normal state samples to form training sample pairs, and outputting multiple time-frequency spectra through the time-frequency analysis knowledge base;
[0009] Step 4: Constructing a high-dimensional feature extractor based on a deep neural network ViT, and performing feature coding on the multi-modal time-frequency spectrum, wherein the input of the encoder is the time-frequency spectrum, and the output is a high-dimensional feature vector;
[0010] Step 5: The rotating machinery continuous normal state monitoring data has strong time sequence characteristics, the time-frequency spectra obtained by the same time-frequency transformation from adjacent samples intercepted by a sliding window should have similarity, and the time-frequency spectra obtained by different time-frequency transformations are different from each other; a comparison matrix is constructed, and a comparison loss function is constructed based on the meaning represented by the comparison matrix;
[0011] Step 6: Training the high-dimensional feature extractor by using the loss function based on the similarity comparison relationship;
[0012] Step 7: Using normal state data samples as reference samples, forming sample pairs with online data to be detected, obtaining time-frequency spectrum through time-frequency analysis knowledge base, inputting sample pairs into trained high-dimensional feature extractor to encode features, then comparing output high-dimensional feature vector with high-dimensional features of normal state samples of rotating machinery, realizing discrimination of normal state and abnormal state of rotating machinery through similarity threshold comparison.
[0013] In a preferred embodiment, in step 3, set normal data sample set , wherein S i represents the i-th normal data sample, i = 1, 2, 3,..., m-1, m, m represents the number of normal data samples, and two adjacent samples are selected to form a training sample pair , time-frequency analysis is performed on the training sample pair through time-frequency analysis knowledge base to obtain time-frequency spectrum of each sample in the training sample pair, and the process is represented as follows
[0014]
[0015]
[0016]
[0017]
[0018] , wherein is the time-frequency spectrum corresponding to the first sample in the i-1-th normal data sample, is the time-frequency spectrum corresponding to the second sample in the i-1-th normal data sample, is the time-frequency spectrum corresponding to the n-th sample in the i-1-th normal data sample, is the time-frequency spectrum corresponding to the first sample in the i-th normal data sample, is the time-frequency spectrum corresponding to the second sample in the i-th normal data sample, is the time-frequency spectrum corresponding to the n-th sample in the i-1-th normal data sample, TFAK(S i-1 ) represents the time-frequency analysis knowledge base constructed by the i-1-th normal data sample, TFAK(S i ) represents the time-frequency analysis knowledge base constructed by the i-th normal data sample, and TFAK represents the constructed time-frequency analysis knowledge base, is the time-frequency analysis operation in the library for the first sample, is the time-frequency analysis operation in the library for the second sample, is the time-frequency analysis operation in the library for the first sample, and f is the obtained time-frequency spectrum.
[0019] In a preferred embodiment, in step 4, a high-dimensional feature extractor F(*) constructed based on a deep neural network is used to encode and represent multiple time-frequency maps, and the process of extracting high-dimensional features from different time-frequency maps by F(*) is represented as follows
[0020]
[0021]
[0022] wherein, and respectively represent the high-dimensional features of the time-frequency map corresponding to the i-th sample and the (i+1)-th sample, respectively. and are the high-dimensional features of the time-frequency map corresponding to the i-th sample and the (i+1)-th sample, respectively. is the index of the normal data sample, and n represents the maximum value of the index of the normal data sample. represents the high-dimensional feature of the x-th normal data sample in the i-th sample obtained by the high-dimensional feature extractor after time-frequency operation, represents the high-dimensional feature of the x-th normal data sample in the i-th sample obtained by the high-dimensional feature extractor after time-frequency operation. and represent the high-dimensional features of the i-th sample obtained by the high-dimensional feature extractor after time-frequency operation.
[0023] In a preferred embodiment, in step 5, the obtained and are used to construct the similarity between the high-dimensional features of each pair of samples, where x = 1, 2, 3,..., n-1, n; and the similarity relationship between the high-dimensional features after the same time-frequency analysis and different time-frequency analyses is calculated and represented as
[0024]
[0025] wherein, is the set of similarity calculation results between each pair of samples, is the sample index, is the cosine similarity calculation, and the above formula is used to obtain similarity calculation values, wherein is the number of selected time-frequency operations.
[0026] and are the high-dimensional features of the adjacent time window samples obtained by processing under the normal state of the rotating machinery; the obtained similarity array has the following characteristics: when , the similarity tends to 1, and at this time there are maximum similarity values; therefore, for the high-dimensional feature extractor, the output and satisfy
[0027]
[0028] While ,
[0029]
[0030] Obtain the optimization direction of high-dimensional feature extractor, that is, the network optimization loss function loss, that is
[0031] .
[0032] In a preferred embodiment, in step 6, the training sample pair composed of adjacent sliding windows in the normal state data sample set is used Through a variety of time-frequency changes, the time-frequency spectrum is obtained, combined with the constructed contrast loss function, and the high-dimensional feature extractor based on deep neural network is optimized by using the back propagation algorithm, so that the contrast loss function is continuously reduced, and the effective optimization of the high-dimensional feature extractor under the zero fault sample is realized.
[0033] In a preferred embodiment, in step 7, the normal state data sample is used as a reference sample set , and the sample to be detected is set as , and and are obtained through the time-frequency analysis knowledge base, where b represents the normal state sample, x is the sample to be detected, and n is the time-frequency analysis operation number; Then, the trained high-dimensional feature extractor is used to extract high-dimensional features from the multi-dimensional time-frequency spectrum to obtain and , and n is the time-frequency analysis operation number.
[0034] In a preferred embodiment, in step 7, the similarity contrast relationship between the sample to be detected and the normal state of the rotating machinery is calculated by combining the high-dimensional feature vectors obtained by the high-dimensional feature extractor, that is, each element in is respectively calculated with each element in to obtain , and then normalized to map to the interval [-0.5, 0.5]; Where b represents the normal state sample, x is the sample to be detected, n is the time-frequency analysis operation number, and i is any normal state sample index.
[0035] Then, a threshold is set, the threshold is named threshold, and the following formula is judged
[0036]
[0037] abs is an absolute value operation, sum is a summation operation; when the above result is True, then the detection result of the rotating machinery state at this time is a fault state, otherwise it is a normal state.
[0038] The application also provides a vibration signal analysis knowledge-driven rotating machinery zero-sample intelligent fault detection device.
[0039] Compared with the prior art, the application has the following beneficial effects:
[0040] 1. Construct a time-frequency analysis knowledge base based on a multi-index evaluation mechanism, use multiple time-frequency analysis methods to perform multi-angle time-frequency analysis on normal state data, obtain multi-modal time-frequency spectrum, and provide support for training of a subsequent high-dimensional feature extractor.
[0041] 2. Construct a high-dimensional feature extractor based on a deep neural network ViT, use a multi-head self-attention mechanism to capture deep relationships between features for multi-modal time-frequency spectrum extracted from the time-frequency analysis knowledge base, and realize high-dimensional abstract representation of rotating machinery state data.
[0042] 3. Design a brand new contrast loss function, which can realize parameter optimization of the above high-dimensional feature extractor by using only sample pairs composed of adjacent time window normal state samples.
[0043] 4. Combine the trained high-dimensional feature extractor, input the sample pair composed of the to-be-detected sample and the normal reference sample, and perform similarity calculation and comparison with the high-dimensional features of the time-frequency analysis results of the rotating machinery normal state data samples, and finally obtain the detection result of the rotating machinery to-be-tested state sample. BRIEF DESCRIPTION OF DRAWINGS
[0044] Figure 1 The application provides a contrast loss function construction and a fault detection algorithm process.
[0045] Figure 2 The application provides a training process of the constructed high-dimensional feature extractor.
[0046] Figure 3 The application provides a fault detection method test result under the condition of zero fault samples. DETAILED DESCRIPTION
[0047] The application will be further described below in combination with the drawings and examples.
[0048] It should be pointed out that the following detailed description is exemplary and is intended to provide further description of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.
[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application; as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.
[0050] The application provides a rotating machine zero-failure sample end-to-end intelligent fault detection method, a multi-index evaluation mechanism is constructed to quantitatively score a plurality of time-frequency analysis methods, and at least two time-frequency analysis methods with higher scores are selected to construct a time-frequency analysis knowledge base; the time sequence similarity of normal data samples of the rotating machine is utilized, it is assumed that the time-frequency spectrum obtained by the same time-frequency change of adjacent samples is similar, and the time-frequency spectrum obtained by different time-frequency transformation has difference, a training sample pair is constructed, and a contrast loss function is constructed based on the assumption relationship; a high-dimensional feature extractor based on a ViT deep neural network is constructed, and the contrast loss function and the training sample pair are combined to train the high-dimensional feature extractor; the sample to be detected and the normal reference sample form a sample pair, which is input into the trained high-dimensional feature extractor, and a fault detection matrix is constructed to detect the state of the rotating machine. The method comprises the following steps:
[0051] Step 1: Collecting continuous normal state monitoring data of the rotating machine, and cutting samples from the data through a sliding window as training samples;
[0052] Step 2: Generating a plurality of time-frequency spectra from the normal state samples through a plurality of time-frequency analysis methods, quantitatively scoring the time-frequency spectra by using an evaluation mechanism composed of PO, SNR and FBE, and selecting at least two time-frequency analysis methods through sorting to construct a time-frequency analysis knowledge base;
[0053] Step 3: Selecting adjacent sliding window cut normal state data samples to form a training sample pair, outputting a plurality of time-frequency spectra through the time-frequency analysis knowledge base;
[0054] Step 4: Constructing a high-dimensional feature extractor based on a ViT deep neural network, performing feature coding on the plurality of time-frequency spectra, and outputting high-dimensional feature vectors;
[0055] Step 5: The rotating machinery continuous normal state monitoring data has strong time sequence characteristics, adjacent samples intercepted by a sliding window should have similarity through time-frequency spectrum obtained by the same time-frequency transform, and the time-frequency spectrum obtained by different time-frequency transforms does not have similarity, similarity analysis is performed on high-dimensional features of different time-frequency analysis results between two continuous normal state samples, and prior knowledge is combined to obtain a high-dimensional feature extractor training loss function under zero fault samples;
[0056] Step 6: The constructed high-dimensional feature extractor is trained in combination with the constructed contrast loss function and the composed training sample pair.
[0057] Step 7: The normal state data sample is used as a reference sample, combined with an online data to be detected to form a sample pair, time-frequency spectrum is obtained through a time-frequency analysis knowledge base, and is input into the trained high-dimensional feature extractor for feature coding, then similarity calculation and analysis are performed on the output high-dimensional feature vectors, and the rotating machinery state is detected in combination with a set threshold.
[0058] The following is a specific implementation process of the application.
[0059] Please refer to Figure 1 、 Figure 2 and Figure 3 , the application provides an end-to-end intelligent fault detection method under zero fault samples of rotating machinery, mainly including the following steps:
[0060] Step 1, select a typical rotating machinery of a steam turbine generator set rotor shaft system;
[0061] Specifically, a fault experiment is performed by using a test platform, a total of 4 single fault experiment data are obtained, which are bearing wear, rotor imbalance, centering deterioration and bearing seat loosening. In each fault experiment, the vibration signal is collected by 1-8 vibration sensors, which are respectively installed in the front bearing Y direction of the coupling, the front bearing Z direction of the coupling, the front bearing Y direction of the rotor, the front bearing Z direction of the rotor, the front bearing X direction of the rotor, the rear bearing Y direction of the rotor, the rear bearing Z direction of the rotor, and the rear bearing X direction of the rotor. The sensor corresponds to channel numbers 1-8, and the data acquisition frequency of the test bench is 20Kh. The normal state vibration data is sampled by using a sliding window, a total of 1000 samples are obtained, which are used as training samples, and 1000 samples of each type of fault state data are sampled, and a total of 5000 samples of normal state samples are used for testing.
[0062] Step 2, select 1000 normal state samples, use a multi-index evaluation mechanism to quantitatively score a plurality of time-frequency analysis methods, and select at least two time-frequency analysis methods to construct a time-frequency analysis knowledge base.
[0063] Specifically, the first 20 normal state samples are subjected to time-frequency changes by using various time-frequency analysis methods to generate time-frequency spectrum, and the evaluation mechanism composed of PO, SNR and FBN three indexes is used to quantitatively score the various time-frequency spectrum, wherein the continuous wavelet change (CWT), short-time Fourier transform (STFT) and chirplet transform have high scores, and the time-frequency analysis knowledge base is constructed based on the above three time-frequency analysis methods.
[0064] Step 3, 1000 normal state data samples are selected, adjacent samples are selected to form training sample pairs, and 1000 groups of time-frequency spectrum are output by constructing the time-frequency analysis knowledge base, each group containing 3 different time-frequency spectrum, each with a size of 3x128x128.
[0065] Specifically, let the normal state data sample set be , adjacent samples form training sample pairs , and and are obtained through the time-frequency analysis knowledge base.
[0066] Step 4, construct a high-dimensional feature extractor F(*) based on the ViT deep neural network, the number of Transformer modules is 3, the number of multi-attention heads is 8, and the dimension of the multi-layer perceptron is 128, then the input of the high-dimensional feature extractor is the color time-frequency spectrum, the input data format is [3, 128, 128], where 3 represents the RGB three channels of the image, 128 and 128 represent the height and width of the image respectively, and the output format is [1, 128], where 1 represents the batch size, i.e. processing one sample at a time, and 128 represents the dimension of the output feature vector.
[0067] Specifically, the constructed high-dimensional feature extractor is used to represent the state of the multi-modal time-frequency spectrum, let the model be F, and the high-dimensional feature vectors obtained by F from different time-frequency spectrum are and .
[0068] Step 5, the continuous normal state monitoring data has strong time sequence characteristics, and the time-frequency spectrum obtained by the same time-frequency transformation from adjacent samples intercepted by a sliding window should have similarity, and the time-frequency spectrum obtained by different time-frequency transformations do not have similarity, a comparison matrix is constructed, and a comparison loss function is constructed based on the above assumptions.
[0069] Specifically, using the obtained and , the The correlation calculation is performed, the training loss function of the high-dimensional feature extractor is constructed under the zero fault sample based on the similarity comparison result, the network parameters are optimized through the loss function, and the result output by the high-dimensional feature extractor meets With . And when the loss function is not descending, the high-dimensional feature extractor training is completed.
[0070] Step 6, the designed zero fault sample loss function is combined with the training sample pair, and the network parameters of the constructed high-dimensional feature extractor F are trained.
[0071] Specifically, two consecutive normal state samples constitute a training sample pair, the high-dimensional feature similarity of each pair is calculated, the loss function is constructed, and the network parameters of the high-dimensional feature extractor are optimized. In the training process, the training period is set to 80, the learning rate is set to 0.00001, the training sample batch is set to 2, and the initial threshold threshold is set to 0.2.
[0072] Step 7, 1000 normal state samples and 1000 fault state samples of each type are selected to form a test set, a total of 5000 samples. The to-be-detected sample and the normal state reference sample pair are obtained by time-frequency analysis knowledge base to obtain the time-frequency spectrum, which is input into the trained high-dimensional feature extractor F for feature coding, then the output high-dimensional feature vector is compared with the high-dimensional feature vector of the normal state reference sample, and the similarity threshold is combined to judge the rotating machinery state.
[0073] Specifically, the is normalized respectively, and each group of elements is mapped to [-0.5, 0.5]. Then, the absolute value of the sum of each group of elements is calculated and compared with 0.2. When , the rotating machinery is in a fault state, otherwise it is in a normal state.
[0074] The above is the preferred embodiment of the present application, any change made according to the technical scheme of the present application, which does not exceed the range of the technical scheme of the present application, belongs to the protection scope of the present application.
Claims
1. A vibration signal analysis knowledge-driven rotating machinery zero-sample intelligent fault detection method, characterized in that, A multi-index evaluation mechanism is used to screen a plurality of time-frequency analysis methods and construct a time-frequency analysis knowledge base, and normal state monitoring signals of the rotating machinery are analyzed from multiple angles in time-frequency, and a comparison loss function is constructed based on the obtained multi-modal time-frequency spectrum; a high-dimensional feature extractor is constructed based on a deep neural network, and the high-dimensional feature extractor is trained based on the comparison loss function; real-time state signals of the rotating machinery to be detected are analyzed in time-frequency through the constructed time-frequency analysis knowledge base, and then high-dimensional features are extracted by using the trained high-dimensional feature extractor, and the high-dimensional features are compared with the plurality of time-frequency spectrums of the normal state signals, similarity threshold value judgment is performed, and the abnormal state of the rotating machinery is determined according to the judgment result; The method comprises the following steps: Step 1: collecting continuous normal state monitoring signals of the rotating machinery, and cutting samples from the signals through a sliding window as training samples; Step 2: generating a plurality of time-frequency spectrums from the normal state samples through a plurality of time-frequency analysis methods, quantitatively scoring the plurality of time-frequency spectrums by using an evaluation mechanism composed of three indexes of PrewittOperator, Signal-to-noise ratio and Frequency Band Energy, and screening the plurality of time-frequency analysis methods through sorting as a time-frequency analysis knowledge base of the rotating machinery monitoring signals; Step 3: selecting normal state samples cut by adjacent sliding windows to form training sample pairs, and outputting a plurality of time-frequency spectrums through the time-frequency analysis knowledge base; Step 4: constructing a high-dimensional feature extractor based on a deep neural network ViT, and performing feature coding on the multi-modal time-frequency spectrums, wherein the input of the encoder is the time-frequency spectrum, and the output is a high-dimensional feature vector; Step 5: the continuous normal state monitoring data of the rotating machinery has strong time sequence characteristics, the time-frequency spectrums obtained by the same time-frequency transformation from adjacent samples cut by a sliding window should have similarity, and the time-frequency spectrums obtained by different time-frequency transformations are different from each other; a comparison matrix is constructed, and a comparison loss function is constructed based on the meaning represented by the comparison matrix; Step 6: training the high-dimensional feature extractor by using the loss function based on the similarity comparison relationship; Step 7: using normal state data samples as reference samples, forming sample pairs with online data to be detected, obtaining time-frequency spectrums through the time-frequency analysis knowledge base, inputting the sample pairs into the trained high-dimensional feature extractor for feature coding, then comparing the output high-dimensional feature vectors with the high-dimensional features of the normal state samples of the rotating machinery, and realizing the discrimination between the normal state and the abnormal state of the rotating machinery through similarity threshold value comparison.
2. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method according to claim 1, characterized in that, In step 3, a normal data sample set is set wherein S i represents the ith normal data sample, i = 1, 2, 3,..., m-1, m, m represents the number of normal data samples, and two adjacent samples are selected to form a training sample pair The time-frequency analysis knowledge base is used to perform time-frequency analysis on the training sample pair to obtain the time-frequency spectrum of each sample in the training sample pair, and the process is represented as follows wherein, is the time-frequency spectrum corresponding to the 1st sample in the (i-1)th normal data sample, is the time-frequency spectrum corresponding to the 2nd sample in the (i-1)th normal data sample, is the time-frequency spectrum corresponding to the nth sample in the (i-1)th normal data sample, is the time-frequency spectrum corresponding to the 1st sample in the ith normal data sample, is the time-frequency spectrum corresponding to the 2nd sample in the ith normal data sample, is the time-frequency spectrum corresponding to the nth sample in the ith normal data sample, i-1 represents the time-frequency analysis knowledge base constructed by the (i-1)th normal data sample, i represents the time-frequency analysis knowledge base constructed by the ith normal data sample, and TFAK represents the time-frequency analysis knowledge base constructed, is the time-frequency analysis operation in the base for the 1st sample, is the time-frequency analysis operation in the base for the 2nd sample, is the time-frequency analysis operation in the base for the nth sample.
3. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method according to claim 1, characterized in that, In step 4, the high-dimensional feature extractor F(•) constructed based on the deep neural network is used to code and represent the plurality of time-frequency spectrums, and the process of extracting high-dimensional features from different time-frequency spectrums by F(•) is as follows wherein, and respectively represent the time-frequency spectrum of the i-th sample in the i-th group of samples, and perform high-dimensional feature extraction on the time-frequency spectrum corresponding to the two connected samples, is the index of the normal data sample, n represents the maximum value of the index of the normal data sample; respectively represent the time-frequency spectrum of the i-th sample in the i-th group of samples, is the high-dimensional feature obtained by the high-dimensional feature extractor after the time-frequency operation on the x-th normal data sample in the sample, respectively represent the high-dimensional feature obtained by the high-dimensional feature extractor after the time-frequency operation on the i-th sample.
4. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method of claim 1, wherein, In step 5, the obtained and The similarity between the high-dimensional features between each pair of samples is constructed, where x = 1, 2, 3,..., n-1, n; and the similarity relationship between the high-dimensional features after the same kind of time-frequency analysis and different kinds of time-frequency analysis is calculated, which is represented as wherein, is a set of similarity computation results between pairs of samples, is a sample index, is a cosine similarity computation, obtained by the above equation is a number of similarity computations, wherein is a number of selected time-frequency operations; and is the high-dimensional feature of the adjacent time window sample of the rotating machinery in normal state; the similarity array obtained has , the similarity factor tends to 1, and at this time, there is maximum similarity value; therefore, the output of the high-dimensional feature extractor and satisfies And when time The optimization direction of the high-dimensional feature extractor is obtained, that is, the network optimization loss function loss is 。 5. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method of claim 1, wherein, In step 6, the training sample pair composed of the normal state data sample set and the adjacent sliding window set in the fault state data sample set Through the time-frequency spectrum obtained by multiple time-frequency changes, combined with the constructed contrast loss function, the high-dimensional feature extractor based on the deep neural network is optimized by using the back propagation algorithm, so that the contrast loss function is continuously reduced, and the effective optimization of the high-dimensional feature extractor under the zero fault sample is realized.
6. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method of claim 1, wherein, In step 7, the normal state data sample is set as a reference sample , the sample to be detected is set as , and the time-frequency analysis knowledge base is used to obtain and , wherein b represents the normal state sample, x is the sample to be detected, and n is the time-frequency analysis operation number; then, the trained high-dimensional feature extractor is used to extract high-dimensional features from the multi-dimensional time-frequency spectrum to obtain and , n is the time-frequency analysis operation number.
7. The vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection method according to claim 6, characterized in that, In step 7, the similarity comparison relationship between the to-be-detected state sample and the normal state of the rotating machinery is calculated by combining the high-dimensional feature vector obtained by the high-dimensional feature extractor, that is, the similarity between each element in and each element in is calculated respectively to obtain , and then normalized to map to the interval [-0.5, 0.5]. Wherein b represents normal state samples, x is a sample to be tested, n is a time-frequency analysis operation number, and i is an arbitrary normal state sample index; Next, a threshold is set, the threshold is named as threshold, and the following formula is judged Abs is an absolute value operation, and sum is a summation operation; when the above result is True, the detection result of the rotating machinery state is a fault state at this time, otherwise, it is a normal state.
8. A vibration signal analysis knowledge-driven rotating machinery zero- sample intelligent fault detection apparatus, characterized in that, Run the vibration signal analysis knowledge-driven rotating machinery zero-sample intelligent fault detection method of any one of claims 1-7.