Rotary machinery zero sample intelligent fault detection method and device driven by vibration signal analysis knowledge

Abstract

The invention relates to a vibration signal analysis knowledge-driven rotary machine zero sample intelligent fault detection method and device. The method comprises the following steps: carrying out quantitative scoring on multiple time-frequency analysis methods by adopting a multi-index evaluation mechanism; based on the similarity of the same-modal time-frequency maps of continuous normal state data and the dissimilarity of different-modal time-frequency maps, a zero-fault sample loss function is obtained based on similarity comparative analysis; a high-dimensional feature extractor based on Vision Transform is constructed, and a normal state sample and a constructed contrast loss function are utilized to realize model training; and calculating the similarity of various time-frequency spectrum high-dimensional features between the to-be-detected state data sample and the normal state data sample of the rotating machine, and realizing online detection of the to-be-detected state of the rotating machine. According to the method, the dependence of a traditional data-driven fault detection method on fault samples is broken through, the rotating machinery intelligent fault detection model can be constructed by using the normal state samples, and high-precision fault detection is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault detection technology for rotating machinery, and in particular to a zero-sample intelligent fault detection method and device for rotating machinery driven by vibration signal analysis knowledge. Background Technology

[0002] Rotating machinery (such as bearings and gears) is a crucial component of industrial equipment, widely used in key sectors such as energy and power, and rail transportation. Its stable operation is vital for production safety and efficiency. However, due to complex operating conditions, prolonged continuous operation, and component aging, rotating machinery is prone to typical faults such as bearing wear, rotor imbalance, and misalignment, leading to downtime losses and even safety accidents. Therefore, achieving accurate early fault detection is an urgent engineering need. 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, these methods heavily rely on large-scale, high-quality fault samples for training. In real-world industrial scenarios, rotating machinery faults occur infrequently and are not fully covered in terms of types. Obtaining complete fault samples is extremely costly, causing deep learning models to struggle with training difficulties due to sample imbalance, especially when zero fault samples are available, making it difficult to establish an intelligent fault detection model for rotating machinery. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a zero-sample intelligent fault detection method and device for rotating machinery driven by vibration signal analysis knowledge. Utilizing the prior knowledge of the temporal continuity of continuous normal state data of rotating machinery, it is assumed that normal state samples captured from adjacent time windows are similar, and therefore the time-frequency spectra obtained after the same time-frequency transformation will also be similar. Secondly, considering that different time-frequency analysis methods have different analytical angles for signals, it is assumed that the time-frequency spectra obtained by the same signal after different time-frequency transformation operations are not similar. Furthermore, based on the above assumptions regarding the comparison relationship of multimodal time-frequency spectra, a contrast loss function is constructed to drive the optimization of high-dimensional feature extractor parameters. Finally, by forming sample pairs between the test sample and normal benchmark samples, the samples are input into the pre-trained high-dimensional feature extractor, which outputs a high-dimensional feature vector and designs a fault detection matrix to detect the state of the rotating machinery.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a vibration signal analysis knowledge-driven zero-sample intelligent fault detection method for rotating machinery. This method employs a multi-index evaluation mechanism to select various time-frequency analysis methods and constructs a time-frequency analysis knowledge base. It performs multi-angle time-frequency analysis on the normal state monitoring signals of rotating machinery, and simultaneously constructs a contrastive loss function based on the obtained multi-modal time-frequency spectra. A high-dimensional feature extractor is constructed based on a deep neural network, and zero-fault sample training is performed using the contrastive loss function. The real-time state signal of the rotating machinery to be detected is analyzed using the constructed time-frequency analysis knowledge base. Then, high-dimensional features are extracted using the trained high-dimensional feature extractor, compared with various time-frequency spectra of the normal state signal, and a similarity threshold is determined. Based on the determination result, the abnormal state of the rotating machinery is identified.

[0005] In a preferred embodiment, the steps specifically include:

[0006] Step 1: Collect continuous normal state monitoring signals of rotating machinery, and extract samples from them through a sliding window as training samples;

[0007] Step 2: Generate various time-frequency spectra from normal state samples using multiple time-frequency analysis methods. Utilize an evaluation mechanism consisting of three indicators—Prewitt Operator, Signal-to-noise ratio, and Frequency Band Energy—to quantify and score the various time-frequency spectra. By sorting and filtering, select various time-frequency analysis methods to serve as a time-frequency analysis knowledge base for rotating machinery monitoring signals.

[0008] Step 3: Select normal state samples from adjacent sliding windows to form training sample pairs, and output various time-frequency spectra through the time-frequency analysis knowledge base;

[0009] Step 4: Construct a high-dimensional feature extractor based on the deep neural network ViT to encode the features of the multimodal time-frequency spectrum. The input of the encoder is the time-frequency spectrum, and the output is a high-dimensional feature vector.

[0010] Step 5: The continuous normal state monitoring data of rotating machinery has strong time-series characteristics. The time-frequency spectra obtained by the same time-frequency transformation of adjacent samples extracted by the sliding window should be similar, and the time-frequency spectra obtained by different time-frequency transformations should be different. Construct a comparison matrix and build a comparison loss function based on the meaning represented by the comparison matrix.

[0011] Step 6: Combine the constructed loss function based on similarity comparison relationship, and use this loss function to train the high-dimensional feature extractor;

[0012] Step 7: Using normal state data samples as benchmark samples, form sample pairs with online data to be detected. Obtain time-frequency spectra through time-frequency analysis knowledge base. Input the sample pairs into the trained high-dimensional feature extractor for feature encoding. Then, compare and analyze the output high-dimensional feature vector with the high-dimensional features of the normal state samples of rotating machinery. By comparing similarity thresholds, the normal state and abnormal state of rotating machinery can be distinguished.

[0013] In a preferred embodiment, in step 3, a normal data sample set is set. , of which S i Let i represent the i-th normal data sample, i = 1, 2, 3, ..., m-1, m, where m represents the number of normal data samples. 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 pairs to obtain the time-frequency spectra of each sample in the training sample pairs. The process is as follows:

[0014]

[0015]

[0016]

[0017]

[0018] in, This is the time-frequency spectrum corresponding to the first sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the second sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the nth sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the first sample in the i-th normal data sample. This is the time-frequency spectrum corresponding to the second sample in the i-th normal data sample. For the time-frequency spectrum corresponding to the nth sample in the (i-1)th normal data sample, TFAK(S) i-1 ) represents the time-frequency analysis knowledge base constructed from the (i-1)th normal data sample, TFAK(S) i ) represents the time-frequency analysis knowledge base constructed from the i-th normal data sample, and TFAK represents the constructed time-frequency analysis knowledge base. For time-frequency analysis of the library for the first sample, For the time-frequency analysis operation in the library for the second sample, For the time-frequency analysis operation in the library for the first sample, f is the obtained time-frequency spectrum.

[0019] In a preferred embodiment, in step 4, a high-dimensional feature extractor F(*) based on a deep neural network is used to encode and represent multiple time-frequency spectra. The process of obtaining high-dimensional features from different time-frequency spectra through F(*) is as follows:

[0020]

[0021]

[0022] in, as well as They represent respectively targeting and High-dimensional feature extraction is performed on the time-frequency spectra corresponding to two connected samples. This is the index for normal data samples, where n represents the maximum value of the normal data sample index. Indicating targeting The high-dimensional features obtained by a high-dimensional feature extractor after performing time-frequency operations on the x-th normal data sample in the sample. These represent the high-dimensional features obtained from sample i after time-frequency operations using a high-dimensional feature extractor.

[0023] In a preferred embodiment, in step 5, the obtained... and Construct the similarity between high-dimensional features of each pair of samples, where x = 1, 2, 3, ..., n-1, n; and calculate the similarity relationship between high-dimensional features after the same time-frequency analysis and different time-frequency analyses, which is expressed as:

[0024]

[0025] in, This is the set of similarity calculation results between pairs of samples. For sample index, For cosine similarity calculation, the above formula is used to obtain... Several similar calculated values, among which The number of time-frequency operations selected;

[0026] and This refers to high-dimensional features obtained by processing adjacent time window samples under normal operating conditions of rotating machinery; the obtained similarity array includes... When the similarity approaches 1, then the similarity has... The maximum similarity value; therefore, for a high-dimensional feature extractor, the output... and satisfy

[0027]

[0028] And when hour

[0029]

[0030] To obtain the optimization direction of the high-dimensional feature extractor, i.e., to optimize the network loss function, we need to...

[0031] .

[0032] In a preferred embodiment, in step 6, the normal state data sample set is used. Training sample pairs composed of adjacent sliding windows By using time-frequency maps obtained from various time-frequency variations and combining them with a constructed contrastive loss function, the high-dimensional feature extractor based on a deep neural network is optimized using the backpropagation algorithm. This continuously reduces the contrastive loss function, enabling effective optimization of the high-dimensional feature extractor under zero-fault samples.

[0033] In a preferred embodiment, in step 7, the normal state data sample is used as a reference sample. Set the sample to be tested as Obtained through time-frequency analysis knowledge base and Where b represents the normal state sample, x is the sample to be tested, and n is the time-frequency analysis operand; then, the trained high-dimensional feature extractor is used to extract high-dimensional features from the multi-dimensional time-frequency spectrum to obtain... and , where n is the time-frequency analysis operand.

[0034] In a preferred embodiment, in step 7, the similarity comparison relationship between the sample to be detected and the normal state of the rotating machinery is calculated by combining the high-dimensional feature vector obtained by the high-dimensional feature extractor. Each element in Similarity calculations are performed on each element in the dataset to obtain... Then, it is normalized and mapped to the interval [-0.5, 0.5]; where b represents the normal state sample, x is the sample to be tested, n is the time-frequency analysis operand, and i is the index of any normal state sample;

[0035] Next, set a threshold, name it threshold, and perform the following judgment.

[0036]

[0037] abs is the absolute value operation, and sum is the summation operation; when the above result is True, the detection result of the rotating machinery is a fault state, otherwise it is a normal state.

[0038] This invention also provides a vibration signal analysis-driven knowledge-based intelligent fault detection device for rotating machinery with zero samples, which operates the aforementioned vibration signal analysis-driven knowledge-based intelligent fault detection method for rotating machinery with zero samples.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. A time-frequency analysis knowledge base is constructed based on a multi-index evaluation mechanism. Various time-frequency analysis methods are used to perform multi-angle time-frequency analysis on normal state data to obtain multi-modal time-frequency maps, which provides support for the training of subsequent high-dimensional feature extractors.

[0041] 2. A high-dimensional feature extractor is constructed based on the deep neural network ViT. The multimodal time-frequency spectrum extracted from the time-frequency analysis knowledge base is captured by a multi-head self-attention mechanism to capture the deep relationships between features, thereby realizing a high-dimensional abstract representation of the rotating machinery state data.

[0042] 3. Design a novel contrastive loss function that can optimize the parameters of the high-dimensional feature extractor by using only sample pairs composed of normal state samples from adjacent time windows.

[0043] 4. Combine the trained high-dimensional feature extractor, input the sample pair of the sample to be detected and the normal benchmark 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 sample to finally obtain the detection result of the rotating machinery test state sample. Attached Figure Description

[0044] Figure 1 This invention provides a flow chart for constructing a contrast loss function and a fault detection algorithm.

[0045] Figure 2 This describes the training process of the high-dimensional feature extractor constructed in this invention;

[0046] Figure 3 The test results are presented for the fault detection method under zero-fault sample conditions proposed in this invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0050] This invention provides an end-to-end intelligent fault detection method for rotating machinery with zero-fault samples. It constructs a multi-index evaluation mechanism to quantitatively score various time-frequency analysis methods, selecting at least two time-frequency analysis methods with higher scores to build a time-frequency analysis knowledge base. Utilizing the temporal similarity of normal rotating machinery data samples, it assumes that adjacent samples undergoing the same time-frequency transformation have similar time-frequency spectra, while time-frequency spectra obtained from different time-frequency transformations are different. Training sample pairs are constructed, and a contrastive loss function is constructed based on this assumption. A high-dimensional feature extractor based on a ViT deep neural network is constructed and trained using the contrastive loss function and training sample pairs. The sample to be tested is paired with a normal baseline sample and input into the trained high-dimensional feature extractor, combined with the constructed fault detection matrix, to detect the state of the rotating machinery. The method includes the following steps:

[0051] Step 1: Collect continuous normal state monitoring data of rotating machinery, and extract samples from it through a sliding window as training samples;

[0052] Step 2: Generate various time-frequency spectra from normal state samples using multiple time-frequency analysis methods. Quantify and score these spectra using an evaluation mechanism composed of three indicators: PO, SNR, and FBE. Build a time-frequency analysis knowledge base by sorting and selecting at least two time-frequency analysis methods.

[0053] Step 3: Select normal state data samples extracted from adjacent sliding windows to form training sample pairs, and output various time-frequency spectra through the time-frequency analysis knowledge base;

[0054] Step 4: Construct a high-dimensional feature extractor based on the ViT deep neural network to encode features from various time-frequency spectra and output high-dimensional feature vectors;

[0055] Step 5: The continuous normal state monitoring data of rotating machinery has strong time series characteristics. The time-frequency spectrum obtained by the same time-frequency transformation of adjacent samples extracted by the sliding window should be similar. The time-frequency spectrum obtained by different time-frequency transformations is not similar. The similarity analysis of the high-dimensional features of different time-frequency analysis results between two continuous normal state samples is performed. Combined with prior knowledge, the training loss function of the high-dimensional feature extractor under zero-fault samples is obtained.

[0056] Step 6: Train the constructed high-dimensional feature extractor by combining the constructed contrastive loss function with the constructed training sample pairs;

[0057] Step 7: Using normal state data samples as benchmark samples, form sample pairs with online data to be detected, obtain time-frequency spectra through time-frequency analysis knowledge base, input them into the trained high-dimensional feature extractor for feature encoding, then perform pairwise similarity calculation and analysis on the output high-dimensional feature vectors, and combine with the set threshold to detect the state of rotating machinery.

[0058] The following is a detailed implementation process of the present invention.

[0059] Please see Figure 1 , Figure 2 and Figure 3 The present invention provides an intelligent fault detection method for zero-fault samples of rotating machinery from bottom to end, which mainly includes the following steps:

[0060] Step 1: Select the rotor shaft system of a steam turbine generator set as a typical rotating machine;

[0061] Specifically, fault experiments were conducted using the test platform, obtaining data for four single faults: bearing wear, rotor imbalance, misalignment, and bearing housing loosening. During each fault experiment, vibration signals were collected by vibration sensors 1-8, installed at the following locations: Y-direction of the front bearing of the coupling, Z-direction of the front bearing of the coupling, Y-direction of the front bearing of the rotor, Z-direction of the front bearing of the rotor, X-direction of the front bearing of the rotor, Y-direction of the rear bearing of the rotor, Z-direction of the rear bearing of the rotor, and X-direction of the rear bearing of the rotor. The corresponding sensor channels were numbered 1-8, and the data acquisition frequency of the test platform was 20kHz. A sliding window was used to sample vibration data under normal conditions, obtaining 1000 samples as training samples. 1000 samples were also sampled for each type of fault condition, resulting in a total of 5000 samples for testing, along with the normal condition samples.

[0062] Step 2: Select 1000 normal state samples, use a multi-index evaluation mechanism to quantify and score various time-frequency analysis methods, and select at least two time-frequency analysis methods to construct a time-frequency analysis knowledge base.

[0063] Specifically, time-frequency changes of the first 20 normal state samples were analyzed using various time-frequency analysis methods to generate time-frequency spectra. An evaluation mechanism consisting of three indicators, PO, SNR, and FBN, was used to quantify and score the various time-frequency spectra. Among them, continuous wavelet transform (CWT), short-time Fourier transform (STFT), and linear frequency modulated wavelet transform (Chirplet transform) scored high. A time-frequency analysis knowledge base was constructed based on the above three time-frequency analysis methods.

[0064] Step 3: Select 1000 normal state data samples, select adjacent samples to form training sample pairs, and output 1000 sets of time-frequency maps by constructing a time-frequency analysis knowledge base. Each set contains 3 different time-frequency maps, and each map is 3×128×128 in size.

[0065] Specifically, let the normal state data sample set be... Adjacent samples form training sample pairs. Obtained through time-frequency analysis knowledge base and .

[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 multilayer perceptron is 128. The input of this high-dimensional feature extractor is a color time-frequency spectrum, and the input data format is [3, 128, 128], where 3 represents the three RGB channels of the image, and 128 and 128 represent the height and width of the image, respectively. The output format is [1, 128], where 1 represents the batch size, that is, one sample is processed at a time, and 128 represents the dimension of the output feature vector.

[0067] Specifically, a constructed high-dimensional feature extractor is used to characterize the state of the multimodal time-frequency spectra. Let the model be F, and the high-dimensional feature vectors obtained by extracting different time-frequency spectra through F are: and .

[0068] Step 5: Continuous normal state monitoring data has strong temporal characteristics. The time-frequency maps obtained by the same time-frequency transformation from adjacent samples extracted by the sliding window should be similar, while the time-frequency maps obtained by different time-frequency transformations are not similar. Construct a contrast matrix and build a contrast loss function based on the above assumptions.

[0069] Specifically, using the obtained and ,conduct Correlation calculations are performed. Based on the similarity comparison results, a training loss function for the high-dimensional feature extractor is constructed using zero-fault samples. The network parameters are then optimized through the loss function to ensure that the output of the high-dimensional feature extractor satisfies... and Furthermore, when the loss function stops decreasing, the training of the high-dimensional feature extractor is complete.

[0070] Step 6: Combine the designed zero-fault sample loss function with the training sample pairs to train the network parameters of the constructed high-dimensional feature extractor F.

[0071] Specifically, for training sample pairs consisting of two consecutive normal state samples, the pairwise high-dimensional feature similarity is calculated to construct a loss function, and the network parameters of the high-dimensional feature extractor are optimized. During training, the training period is set to 80, the learning rate is set to 0.00001, the training sample batch size is set to 2, and the initial threshold is set to 0.2.

[0072] Step 7: Select 1000 samples from normal state and 1000 samples from various fault states to form a test set, totaling 5000 samples. The time-frequency spectra of the samples to be tested and the normal state baseline samples are obtained through a time-frequency analysis knowledge base. These spectra are then input into the trained high-dimensional feature extractor F for feature encoding. The output high-dimensional feature vector is then compared with the high-dimensional feature vector of the normal state baseline samples for similarity calculation. Combined with a similarity threshold, the state of the rotating machinery is determined.

[0073] Specifically, it means that Normalization is performed on each group of elements, mapping each group to [-0.5, 0.5]. Next, the absolute value of the sum of each group of elements is calculated and compared to 0.2. When there exists... If the condition is met, the rotating machinery is in a fault state; otherwise, it is in a normal state.

[0074] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A knowledge-driven method for zero-sample intelligent fault detection in rotating machinery based on vibration signal analysis, characterized in that, A multi-index evaluation mechanism was employed to select various time-frequency analysis methods and construct a time-frequency analysis knowledge base. Multi-angle time-frequency analysis was performed on the normal state monitoring signals of rotating machinery. Simultaneously, a contrastive loss function was constructed based on the obtained multimodal time-frequency spectra. A high-dimensional feature extractor was built based on a deep neural network, and zero-fault sample training was conducted using the contrastive loss function. The real-time state signal of the rotating machinery under test was analyzed using the constructed time-frequency analysis knowledge base. Then, high-dimensional features were extracted using the trained high-dimensional feature extractor and compared with various time-frequency spectra of the normal state signal. A similarity threshold was determined, and the abnormal state of the rotating machinery was identified based on the judgment results.

2. The method for zero-sample intelligent fault detection of rotating machinery driven by vibration signal analysis knowledge according to claim 1, characterized in that, The method includes the following steps: Step 1: Collect continuous normal state monitoring signals of rotating machinery, and extract samples from them through a sliding window as training samples; Step 2: Generate various time-frequency spectra from normal state samples using multiple time-frequency analysis methods. Utilize an evaluation mechanism consisting of three indicators: Prewitt Operator, Signal-to-noise ratio, and Frequency Band Energy, to quantify and score the various time-frequency spectra. By sorting and filtering, select various time-frequency analysis methods to serve as a time-frequency analysis knowledge base for rotating machinery monitoring signals. Step 3: Select normal state samples from adjacent sliding windows to form training sample pairs, and output various time-frequency spectra through the time-frequency analysis knowledge base; Step 4: Construct a high-dimensional feature extractor based on the deep neural network ViT to encode the features of the multimodal time-frequency spectrum. 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 rotating machinery has strong time-series characteristics. The time-frequency spectra obtained by the same time-frequency transformation of adjacent samples extracted by the sliding window should be similar, and the time-frequency spectra obtained by different time-frequency transformations should be different. Construct a comparison matrix and build a comparison loss function based on the meaning represented by the comparison matrix. Step 6: Combine the constructed loss function based on similarity comparison relationship, and use this loss function to train the high-dimensional feature extractor; Step 7: Using normal state data samples as benchmark samples, form sample pairs with online data to be detected. Obtain time-frequency spectra through time-frequency analysis knowledge base. Input the sample pairs into the trained high-dimensional feature extractor for feature encoding. Then, compare and analyze the output high-dimensional feature vector with the high-dimensional features of the normal state samples of rotating machinery. By comparing similarity thresholds, the normal state and abnormal state of rotating machinery can be distinguished.

3. The method for zero-sample intelligent fault detection of rotating machinery driven by vibration signal analysis knowledge according to claim 2, characterized in that, In step 3, let's assume a normal data sample set. , among which, S i Let i represent the i-th normal data sample, i = 1, 2, 3, ..., m-1, m, where m represents the number of normal data samples. 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 pairs to obtain the time-frequency spectra of each sample in the training sample pairs. The process is as follows: , , , ,in, This is the time-frequency spectrum corresponding to the first sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the second sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the nth sample in the (i-1)th normal data sample. This is the time-frequency spectrum corresponding to the first sample in the i-th normal data sample. This is the time-frequency spectrum corresponding to the second sample in the i-th normal data sample. For the time-frequency spectrum corresponding to the nth sample in the (i-1)th normal data sample, TFAK(S) i-1 ) represents the time-frequency analysis knowledge base constructed from the (i-1)th normal data sample, TFAK(S) i ) represents the time-frequency analysis knowledge base constructed from the i-th normal data sample, and TFAK represents the constructed time-frequency analysis knowledge base. For time-frequency analysis of the library for the first sample, For the time-frequency analysis operation in the library for the second sample, This is a time-frequency analysis operation in the library for the first sample.

4. The method for zero-sample intelligent fault detection of rotating machinery driven by vibration signal analysis knowledge according to claim 2, characterized in that, In step 4, a high-dimensional feature extractor F(*) based on a deep neural network is used to encode and represent various time-frequency spectra. The process of obtaining high-dimensional features from different time-frequency spectra through F(*) is as follows: , ,in, as well as They represent respectively targeting and High-dimensional feature extraction is performed on the time-frequency spectra corresponding to two connected samples. This is the index for normal data samples, where n represents the maximum value of the normal data sample index. Indicating targeting The high-dimensional features obtained by a high-dimensional feature extractor after performing time-frequency operations on the x-th normal data sample in the sample. These represent the high-dimensional features obtained from sample i after time-frequency operations using a high-dimensional feature extractor.

5. The method for zero-sample intelligent fault detection of rotating machinery driven by vibration signal analysis knowledge according to claim 2, characterized in that, In step 5, using the obtained and Construct the similarity between high-dimensional features of each pair of samples, where x = 1, 2, 3, ..., n-1, n; and calculate the similarity relationship between high-dimensional features after the same time-frequency analysis and different time-frequency analyses, which is expressed as: ,in, This is the set of similarity calculation results between each pair of samples. For sample index, For cosine similarity calculation, the above formula is used to obtain... Several similar calculated values, among which The number of time-frequency operations selected; and This refers to the high-dimensional features obtained by processing adjacent time window samples under normal operating conditions of rotating machinery; the obtained similarity array includes... When the similarity approaches 1, then the similarity has... The maximum similarity value; therefore, for a high-dimensional feature extractor, the output... and satisfy , and when hour To obtain the optimization direction of the high-dimensional feature extractor, i.e., the network optimization loss function loss. 。 6. The vibration signal analysis-driven knowledge-based intelligent fault detection method for rotating machinery with zero samples according to claim 2, characterized in that, In step 6, the normal state data sample set is used. Training sample pairs composed of adjacent sliding windows By using time-frequency maps obtained from various time-frequency variations, combined with the constructed contrastive loss function, and employing the backpropagation algorithm to optimize the high-dimensional feature extractor based on a deep neural network, the contrastive loss function is continuously reduced, thus achieving effective optimization of the high-dimensional feature extractor under zero-fault samples.

7. The method for zero-sample intelligent fault detection of rotating machinery driven by vibration signal analysis knowledge according to claim 2, characterized in that, In step 7, the normal state data sample is used as the baseline sample. Set the sample to be tested as Obtained through time-frequency analysis knowledge base and Where b represents the normal state sample, x is the sample to be tested, and n is the time-frequency analysis operand; then, the trained high-dimensional feature extractor is used to extract high-dimensional features from the multi-dimensional time-frequency spectrum to obtain... and , where n is the time-frequency analysis operand.

8. The vibration signal analysis-driven knowledge-based intelligent fault detection method for rotating machinery with zero samples according to claim 7, characterized in that, In step 7, the high-dimensional feature vectors obtained by the high-dimensional feature extractor are combined to calculate the similarity comparison between the sample to be detected and the normal state of the rotating machinery. Each element in Similarity calculations are performed on each element in the dataset to obtain... Then, it is normalized and mapped to the interval [-0.5, 0.5]. Where b represents a normal state sample, x is the sample to be tested, n is the time-frequency analysis operand, and i is the index of any normal state sample. Next, set a threshold, name it threshold, and perform the following judgment. abs is the absolute value operation, and sum is the summation operation; when the above result is True, the detection result of the rotating machinery is a fault state, otherwise it is a normal state.

9. A vibration signal analysis-driven, zero-sample intelligent fault detection device for rotating machinery, characterized in that, The method described in any one of claims 1-8 is a vibration signal analysis-driven intelligent fault detection method for rotating machinery with zero samples.

Images