Rotating machinery intelligent fault detection method based on nonparametric self-adaptive dynamic threshold

Abstract

The invention relates to an intelligent fault detection method for a rotating machine based on a nonparametric self-adaptive dynamic threshold. The method comprises the following steps: constructing a sample feature data set in a normal working state; generating a model through a preset mechanism characteristic, and generating a virtual fault sample; training a preset LSTM-AE network model by using the sample feature data set to obtain an identification model of normal engine data; identifying a preset test set by using the identification model of the normal engine data to obtain a reconstruction error of the test set; wherein the test set comprises sample feature data in a normal working state and a virtual fault sample; comparing the reconstruction error of the test set with the reconstruction error of the sample feature data set, and if an error threshold value exceeds a preset dynamic threshold value; and if yes, determining that the vibration signal is abnormal. According to the method, the detection process is more efficient and reliable, the false alarm rate and the missing report rate are remarkably reduced, and an effective solution is provided for fault detection and quality control of the rotating machine.

Description

Technical Field

[0001] The invention relates to the technical field of engine fault detection, and in particular to a rotating machinery intelligent fault detection method with a non-parameter adaptive dynamic threshold. Background Art

[0002] With the rapid development of new energy vehicle technology, despite the growing demand for clean energy vehicles in the market, fuel vehicles still maintain a huge number of vehicles in the world. However, in terms of maintenance and management of fuel vehicles, especially in engine fault detection technology, the existing methods are still relatively backward, which to a certain extent limits the optimization of vehicle performance and the guarantee of driving safety. The engine is the heart of the car, and its health is directly related to the operating efficiency and service life of the vehicle. Therefore, it is particularly important to develop advanced engine fault detection technology.

[0003] Traditional engine fault detection methods often rely on regular manual inspections or judgments based on experience. These methods are not only inefficient, but also difficult to detect early signs of faults in a timely manner. The use of cold test technology, that is, testing the engine before it is started, can effectively detect some faults that are not easy to detect at normal operating temperature, such as small cracks between components, imbalance or other potential problems. However, the data generated during the cold test contains complex noise and nonlinear characteristics, which puts forward high sensitivity and accuracy requirements for fault detection technology. At present, fault detection research is moving towards the advancement of signal processing technology, the integration of machine learning and deep learning, automated feature extraction and selection, improving model generalization ability, real-time detection and online monitoring, data set standardization, and multidisciplinary method integration. Although existing technologies perform well in specific situations, improving model generalization, building standardized public data sets, achieving real-time online monitoring, and enhancing detection robustness are still the focus and challenges of current research.

[0004] In summary, although current fault detection research has made progress in signal processing, machine learning, deep learning fusion, automated feature extraction, model generalization, real-time monitoring, etc., how to further improve the generalization ability of the model, build a standardized data set, realize real-time online monitoring, and enhance detection robustness are still the focus and challenges of research in this field. Summary of the invention

[0005] The purpose of the present invention is to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A non-parametric adaptive dynamic threshold intelligent fault detection method for rotating machinery, characterized by comprising:

[0008] Based on the vibration signals during the engine cold test under different working conditions, a sample feature data set of normal working conditions is constructed;

[0009] Generate a model through preset mechanism characteristics and generate virtual fault samples;

[0010] Using the sample feature data set to train a preset LSTM-AE network model to obtain a recognition model for normal engine data;

[0011] Using the recognition model of normal engine data, a preset test set is recognized to obtain a reconstruction error of the test set; wherein the test set includes: sample feature data of a normal working state and a virtual fault sample;

[0012] The reconstruction error of the test set is compared with the reconstruction error of the sample feature data set. If the error threshold exceeds a preset non-parametric dynamic threshold, the vibration signal is determined to be in an abnormal state.

[0013] Optionally, constructing the sample feature data set in a normal working state includes:

[0014] Collect vibration signals during engine cold tests under different working conditions, filter and reduce noise to obtain preset effective information;

[0015] Based on the preset effective signal, construct a sample data set in a normal working state;

[0016] Perform feature extraction on the sample data set to obtain the sample feature data set in a normal working state.

[0017] Optionally, the sample feature data set in the normal working state includes: time domain characteristics, frequency domain characteristics, average instantaneous power, cyclostationarity and cyclostationarity index of the normal vibration signal of the engine cold test.

[0018] Optionally, the preset mechanism feature generation model is:

[0019] x f =x+M·(x theo +∈)

[0020] Among them, x represents the source domain data with a length of L, x f Represents the frequency domain data of the generated fault state, x theo represents the theoretical defect frequency, ∈ represents the theoretical sideband frequency, M represents the commonality parameter vector, and M(·) represents the coefficient vector of the fault frequency amplitude variation.

[0021] Optionally, the preset LSTM-AE network includes:

[0022] Encoder: The encoder is the first half of LSTM-AE. The encoder uses LSTM units to process the input sequence time-step by time step. The goal of the encoder is to compress the input sequence into a latent vector of a fixed size that captures the preset important features of the input data.

[0023] Decoder: The decoder is the second half of the LSTM-AE. The decoder receives the latent vector output by the encoder and tries to reconstruct the original input sequence. The decoder is also composed of LSTM units, but it gradually reconstructs the sequence in a generative manner.

[0024] Optionally, obtaining the reconstruction error of the test set includes:

[0025] The test set is input into the recognition model trained with normal engine data. The encoder part of the recognition model is first used to encode the test set to obtain a low-dimensional feature representation. The decoder is used to remap the low-dimensional feature representation back to the dimension of the original data. On this basis, the difference between the original data of the test set and the decoded reconstructed data is evaluated. The reconstruction error is quantified by calculating the mean squared error (MSE) between the two, thereby measuring the model's ability to reconstruct the test data. Specifically, the mean squared error can be calculated using the following formula:

[0026]

[0027] in, represents the i-th sample in the test set, is the corresponding sample reconstructed after decoding, and n is the total number of samples in the test set. This indicator reflects the error size of the model in the reconstruction process and provides a basis for further analysis and judgment.

[0028] Optionally, using the sample feature data set to train a preset LSTM-AE network model includes:

[0029] Divide the sample feature data set in normal working state into a training set and a validation set;

[0030] The training set is used to train the preset LSTM-AE network model, an initial recognition model of normal engine data is obtained, and the network parameters are optimized by a back propagation algorithm;

[0031] The validation set is used to evaluate the model performance of the initial recognition model, encode and decode the sample feature data of the new normal working state, optimize the model structure parameters, and obtain the recognition model of normal engine data.

[0032] Optionally, the dynamic threshold is:

[0033] threshold=μ+k×σ

[0034] Wherein, threshold represents a non-parametric adaptive dynamic threshold, μ represents the mean of the reconstruction error of the training set, σ represents the standard deviation of the reconstruction error of the validation set, and k is a constant used to determine the sensitivity of the threshold. In the present invention, k is 3.

[0035] The beneficial effects of the present invention are:

[0036] The method proposed in the present invention not only improves the accuracy of extracting features from vibration signals by combining MCGM and LSTM-AE technology, but also effectively captures the complex nonlinear characteristics of rotating machinery during operation by using the ability of deep learning, and realizes high-sensitivity detection of fault signals.

[0037] In practical applications, even when the number of fault samples is limited, the method of the present invention can generate personalized virtual fault samples through MCGM, which enhances the model's ability to identify unknown faults and has high diagnostic accuracy and robustness.

[0038] The present invention solves the problem that the fixed threshold in the traditional method is difficult to adapt to different working conditions and samples by constructing a non-parametric adaptive dynamic threshold optimization method, making fault diagnosis more flexible and accurate and improving the generalization ability of the model;

[0039] Compared with the prior art, the present invention not only improves the efficiency and automation level of fault detection, but also realizes continuous monitoring and instant feedback of the operating status of rotating machinery through real-time adjustment of non-parametric adaptive dynamic thresholds, providing an effective solution for fault detection and quality control of rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0041] Figure 1 This is a main flow chart of the rotating machinery intelligent fault detection method according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the MCGM principle of an embodiment of the present invention;

[0043] Figure 3 The present invention is a schematic diagram of a specific process of a rotating machinery intelligent fault detection method with a non-parametric adaptive dynamic threshold according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] This embodiment proposes a non-parametric adaptive dynamic threshold intelligent fault detection method for rotating machinery, which is specifically designed to analyze the vibration signal during the engine cold test. Figure 1 As shown in the figure. The method first uses an angular velocity sensor to collect normal signals of the engine cold test to construct a data set. Then, the time domain features are extracted from the vibration signal, converted into frequency domain features through fast Fourier transform (FFT), and normalized. Then, the fault mechanism is combined with the individual characteristics of the cold test engine through the mechanism feature generation model (MCGM) to generate personalized virtual fault samples. The data is divided into a training set and a validation set, and then the LSTM-AE network model is constructed and trained. Finally, a non-parametric dynamic threshold is set based on the mean of the training data and the standard deviation of the validation data, and dynamically adjusted to identify reconstruction errors beyond the normal range, so as to achieve effective detection of abnormal engine operation. This method not only significantly improves the accuracy of fault diagnosis and the adaptability of the model, but also provides an effective technical means for the health management and predictive maintenance of rotating machinery.

[0047] The present embodiment proposes a non-parametric adaptive dynamic threshold intelligent fault detection method for rotating machinery, which specifically includes the following steps:

[0048] Based on the vibration signals during the engine cold test under different working conditions, a sample feature data set of normal working conditions is constructed;

[0049] Generate a model through preset mechanism characteristics and generate virtual fault samples;

[0050] The preset LSTM-AE network model is trained using the sample feature data set to obtain the recognition model of normal engine data;

[0051] Using the recognition model of normal engine data, a preset test set is recognized to obtain the reconstruction error of the test set; wherein the test set includes: sample feature data of normal working state and virtual fault samples;

[0052] The reconstruction error of the test set is compared with the feature data set and the reconstruction error. If the error threshold exceeds the preset dynamic threshold, the vibration signal is judged to be in an abnormal state.

[0053] Furthermore, constructing a sample feature data set in a normal working state includes:

[0054] Collect vibration signals during engine cold tests under different working conditions, filter and reduce noise to obtain preset effective information;

[0055] Based on the preset effective signals, a sample data set of normal working status is constructed;

[0056] Feature extraction is performed on the sample data set to obtain a sample feature data set in a normal working state. The sample feature data set in a normal working state includes: time domain features, frequency domain features, average instantaneous power, cyclostationarity and cyclostationarity index of the normal vibration signal of the engine cold test.

[0057] Specifically, in this embodiment, a normal vibration signal data set during the engine cold test is constructed, and feature extraction and normalization processing are performed on the data set. The specific implementation process includes:

[0058] Vibration sensors are used to collect vibration signals during the engine cold test under different working conditions. Stable signals are selected as effective audio, and stable and reliable signals are selected as effective information to create a sample data set of the normal working state of the cold test engine.

[0059] According to the time-frequency domain characteristics of the normal vibration signal of the engine cold test and the three characteristics of mean instantaneous power (MIP), cyclostationarity (DCSα) and cyclostationarity index (ICSnx), the fast Fourier transform (FFT) converts the time domain signal into the frequency domain signal, which can be expressed as:

[0060]

[0061] Where X[k] is the kth complex component of the signal in the frequency domain. x[n] is the nth sample value of the signal in the time domain. N is the total number of samples of the signal. j is an imaginary unit that satisfies j 2 = -1. e is the base of natural logarithms.

[0062] Mean instantaneous power (MIP) is a statistical feature used to describe the energy distribution of a signal. Its value reflects the average energy level of the signal in an instantaneous period of time. For a signal x(t), its MIP can be expressed as:

[0063]

[0064] N represents the length of the signal or the size of the time window.

[0065] Cyclostationarity (DCS α ) is used to evaluate the periodicity and frequency characteristics of a signal, and is particularly suitable for the analysis and identification of periodic signals. For a signal x(t), DCS α It can be expressed as:

[0066]

[0067] Where X(k) is the Fourier transform coefficient of the signal, α is the periodic frequency parameter, and N is the length of the signal.

[0068] Cyclostationary index (ICS nx ) is used to measure the periodic strength and frequency components of a signal, and can help analyze the periodic characteristics of a signal. For a signal x(t), ICS nx It can be expressed as:

[0069]

[0070] Where X(k) is the Fourier transform coefficient of the signal, X * (k) is its conjugate and n is the exponential parameter.

[0071] Then the eigenvectors and eigennumbers of the normal data are normalized using the MinMaxScaler normalization method, which performs minimum-maximum normalization, as follows:

[0072]

[0073] Where X is the original data point, X min is the minimum value in the feature column, X max is the maximum value in the feature column. norm is the normalized data point. Use normalization to scale the data between 0 and 1. MinMaxScaler calculates the minimum and maximum values ​​of each feature column in the fit step and applies the above formula in the transform step to normalize the data.

[0074] In this embodiment, a virtual fault sample is generated by generating a model through a preset mechanism feature; specifically, the process includes:

[0075] The signal characteristics caused by mechanical failure are common characteristics of rotating machinery, which are reflected in the fault characteristic frequency of the faulty component. Conduct in-depth analysis of the failure mechanism of the machine to understand the conditions, causes and processes of the failure. Collect data on the machine in normal and faulty states, including sensor data, operating parameters, environmental conditions, etc., and extract features from the collected data. These features represent the individual characteristics and failure modes of the machine. Based on the failure mechanism and the extracted features, build a model that can generate fault samples. According to the individual characteristics of the machine, adjust the model parameters to ensure that the generated fault samples can reflect the failure behavior of a specific machine under specific conditions. The MCGM principle is as follows: Figure 2 As shown in Figure 2. Virtual fault samples are generated using the adjusted model.

[0076] Specifically, the initial imbalance and initial misalignment of the engine under normal operating conditions during the cold test are taken as individual features, and the specific fault frequency characteristics shown in the spectrum diagram when the engine fails during the cold test are taken as common features. The common features under fault conditions are added to the individual features of the normal operation of the engine to generate virtual fault samples. In the frequency domain, the generation of fault state data can be expressed by the following formula:

[0077] x f =x+M·(x theo +∈)

[0078] Where x represents the source domain data with a length of L. f The frequency domain data representing the generated fault state is also of length L. theo represents the theoretical defect frequency, which is calculated based on the structural dimensions of the engine. ∈ represents the theoretical sideband frequency, which is associated with the fault characteristics. M represents the common parameter vector, which is a joint probability distribution consisting of the coefficients of the change in the fault frequency amplitude relative to the normal state. M(·) represents the coefficient vector of the change in the fault frequency amplitude, which reflects the impact of the fault on the frequency domain data.

[0079] The LSTM-AE network designed in this embodiment includes:

[0080] Encoder: The encoder is the first half of LSTM-AE, which uses LSTM units to process the input sequence time-step by time-step. Each LSTM unit includes an input gate, a forget gate, and an output gate, which control the flow and update of information. The goal of the encoder is to compress the input sequence into a latent vector of a fixed size that captures the important features of the input data;

[0081] Decoder: The decoder is the second half of the LSTM-AE, which receives the latent vector output by the encoder and tries to reconstruct the original input sequence. The decoder is also composed of LSTM units, but it gradually reconstructs the sequence in a generative manner. This process can be seen as the inverse process of the encoder's compressed information;

[0082] LSTM Cell: The core of LSTM-AE is the LSTM cell, which contains the following key components:

[0083] Forget Gate: Determines which information to discard from the current cell state.

[0084] Input Gate: Determines what new information is added to the cell state.

[0085] Cell State: Carries information about the observed input sequence and can maintain information between time steps.

[0086] Output Gate: Determines the output of the next hidden state.

[0087] Furthermore, the preset LSTM-AE network model is trained using the sample feature data set to obtain the recognition model of normal engine data including:

[0088] Divide the sample feature data set in normal working state into a training set and a validation set;

[0089] Use the training set to train the preset LSTM-AE network model, obtain the initial recognition model of normal engine data, and optimize the network parameters through the back propagation algorithm;

[0090] The validation set is used to evaluate the model performance of the initial recognition model, encode and decode the sample feature data of the new normal working state, optimize the model structure parameters, and obtain the recognition model of normal engine data.

[0091] Specifically, in this embodiment, the feature data is divided into a training set and a validation set through data partitioning, the training set is used as input, the LSTM-AE model is trained, a model of normal engine data is obtained, and the network parameters are optimized through the back propagation algorithm. Then, the model performance is evaluated on the validation set, the new engine operation data is encoded and decoded, the model structure parameters are optimized, and finally, the validation is performed on the validation set.

[0092] Furthermore, obtaining the reconstruction error of the test set includes:

[0093] The test set is input into the recognition model trained with normal engine data. First, the encoder part of the recognition model is used to encode the test set to obtain a low-dimensional feature representation. The decoder is used to remap the low-dimensional feature representation back to the dimension of the original data. On this basis, the difference between the original data of the test set and the decoded reconstructed data is evaluated. The reconstruction error is quantified by calculating the mean squared error (MSE) between the two, thereby measuring the model's ability to reconstruct the test data. Specifically, the mean squared error can be calculated using the following formula:

[0094]

[0095] in, represents the i-th sample in the test set, is the corresponding sample reconstructed after decoding, and n is the total number of samples in the test set. This indicator reflects the error size of the model in the reconstruction process and provides a basis for further analysis and judgment.

[0096] Specifically, in this embodiment, the processed rotating machinery vibration test set is imported into the trained LSTM-AE model, and the encoder part of the LSTM-AE model is used to encode the test data to obtain a low-dimensional feature representation. The decoder part of the model is then used to decode the encoded feature representation back to the original data dimension. The difference between the original test data and the decoded data is calculated, and the reconstruction error is determined using the mean square error (MSE).

[0097] In order to effectively identify samples with abnormal reconstruction errors, this embodiment adopts a dynamic threshold adjustment method based on non-parametric statistics. Specifically, the reconstruction error of the training set is first calculated from the cold test data set of the normal engine. Then, an initial non-parametric dynamic threshold is set based on the mean of the training set error and three times the standard deviation of the validation set error. This method takes into account the central position and dispersion of the data, so that the threshold can be more flexibly adapted to the characteristics of the data.

[0098] Furthermore, this implementation adaptively adjusts the initially set non-parametric dynamic threshold according to the central tendency of the training set and the dispersion of the validation set. This adjustment mechanism allows the threshold to more accurately capture samples whose reconstruction errors exceed the normal fluctuation range. In this way, we can more accurately identify potential abnormal or faulty samples.

[0099] In the implementation process, the reconstruction error is first calculated using the normal engine data set, and a non-parametric dynamic threshold is determined based on it. Then, this dynamic threshold is applied to the test set containing normal and faulty samples, and adaptive adjustments are made to identify samples whose errors exceed the threshold. Finally, the calculated dynamic threshold is saved, providing a reliable standard for subsequent anomaly detection.

[0100] Specifically, the calculation formula of the non-parametric dynamic threshold can be expressed as: the threshold is set based on the mean of the reconstruction error of the training set plus the standard deviation of the reconstruction error of the three-time validation set. The mean is the sum of all the values ​​in the data set divided by the total number of values, indicating the center position of the data set. train The mean calculation formula for N samples in is:

[0101]

[0102] Among them, E train，i is the reconstruction error of the ith training sample.

[0103] The standard deviation indicates the deviation of each value in the data set from the mean, and is a measure of the degree of dispersion of the data set. val The standard deviation of M samples is calculated as:

[0104]

[0105] Among them, E val，j is the reconstruction error of the jth validation sample, and μ val is the mean reconstruction error of the validation set.

[0106] The threshold is adjusted dynamically according to the central tendency of the training set and the dispersion of the validation set. Combined with the 3σ principle in the control chart, it assumes that the data is normally distributed and most of the data will fall within three standard deviations of the mean. In this case, if the reconstruction error exceeds the mean of the training set plus three times the standard deviation of the validation set, then the sample may be considered abnormal. The calculated threshold is saved in a file for subsequent fault detection. Threshold calculation formula:

[0107] threshold=μ+k×σ

[0108] Wherein, μ is the mean of the reconstruction error of the training set, σ is the standard deviation of the reconstruction error of the validation set, and k is a constant used to determine the sensitivity of the threshold, and in the present invention, k=3.

[0109] Finally, this embodiment uses a trained LSTM autoencoder model and a normalizer to perform feature extraction and anomaly detection on the newly input rotating machinery data, and identifies the abnormal state of the rotating machinery under operation based on the set dynamic threshold.

[0110] In the method provided in this embodiment, the fault mechanism is combined with the individual characteristics of the machine through a mechanism feature generation model (MCGM) to generate personalized virtual fault samples, which solves the problem of insufficient fault samples, improves the adaptability and diagnostic accuracy of the model to specific machines, and enhances the robustness of the model. The present invention is a rotating machinery intelligent fault detection method with a non-parametric adaptive dynamic threshold. The method is verified by mathematical modeling and simulation experiments, and can perform fault detection on rotating machinery in operation. Compared with the prior art, the present method solves the problem of insufficient fault samples, and uses the adjusted model to generate virtual fault samples, which can be used to test the fault diagnosis system and improve its accuracy and robustness. The rotating machinery intelligent fault detection method using a non-parametric adaptive dynamic threshold not only makes the detection process more efficient and reliable, but also significantly reduces the false alarm rate and missed alarm rate, saving time and cost. The flow chart of the rotating machinery intelligent fault detection method with a non-parametric adaptive dynamic threshold of the present invention is as follows Figure 3 shown.

[0111] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A non-parametric adaptive dynamic threshold intelligent fault detection method for rotating machinery, characterized in that: include: Based on the vibration signals during the engine cold test under different working conditions, a sample feature data set of normal working conditions is constructed; Generate a model through preset mechanism characteristics and generate virtual fault samples; Using the sample feature data set to train a preset LSTM-AE network model to obtain a recognition model for normal engine data; Using the recognition model of normal engine data, a preset test set is recognized to obtain a reconstruction error of the test set; wherein the test set includes: sample feature data of a normal working state and a virtual fault sample; The reconstruction error of the test set is compared with the reconstruction error of the sample feature data set. If the error threshold exceeds a preset non-parametric dynamic threshold, the vibration signal is determined to be in an abnormal state.

2. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: Constructing the sample feature data set in a normal working state includes: Collect vibration signals during engine cold tests under different working conditions, filter and reduce noise to obtain preset effective information; Based on the preset effective signal, construct a sample data set in a normal working state; Perform feature extraction on the sample data set to obtain the sample feature data set in a normal working state.

3. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: The sample feature data set in the normal working state includes: time domain characteristics, frequency domain characteristics, average instantaneous power, cyclostationarity and cyclostationarity index of the normal vibration signal of the engine cold test.

4. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: The preset mechanism feature generation model is: x f =x+M·(x theo +∈) Among them, x represents the source domain data of preset length, x f Represents the frequency domain data of the generated fault state, x theo represents the theoretical defect frequency, ∈ represents the theoretical sideband frequency, M represents the commonality parameter vector, and M(·) represents the coefficient vector of the fault frequency amplitude variation.

5. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: The preset LSTM-AE network model includes: Encoder: The encoder is the first half of LSTM-AE. The encoder uses LSTM units to process the input sequence time-step by time step. The goal of the encoder is to compress the input sequence into a latent vector of a fixed size that captures the preset important features of the input data. Decoder: The decoder is the second half of the LSTM-AE. The decoder receives the latent vector output by the encoder and tries to reconstruct the original input sequence. The decoder is also composed of LSTM units, but it gradually reconstructs the sequence in a generative manner.

6. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: Obtaining the reconstruction error of the test set includes: The test set is input into the recognition model trained with normal engine data. The encoder part of the recognition model is first used to encode the test set to obtain a low-dimensional feature representation. The decoder is used to remap the low-dimensional feature representation back to the dimension of the original data. The difference between the original data of the test set and the decoded reconstructed data is evaluated, and the reconstruction error is quantified by calculating the mean square error between the two. The mean square error is calculated using the following formula: in, represents the i-th sample in the test set, is the corresponding sample reconstructed after decoding, n is the total number of samples in the test set, and MSE stands for mean square error.

7. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 1 is characterized in that: Using the sample feature data set to train the preset LSTM-AE network model includes: Divide the sample feature data set in normal working state into a training set and a validation set; The training set is used to train the preset LSTM-AE network model, an initial recognition model of normal engine data is obtained, and the network parameters are optimized by a back propagation algorithm; The validation set is used to evaluate the model performance of the initial recognition model, encode and decode the sample feature data of the new normal working state, optimize the model structure parameters, and obtain the recognition model of normal engine data.

8. The rotating machinery intelligent fault detection method with non-parametric adaptive dynamic threshold according to claim 7 is characterized in that: The dynamic threshold is: threshold=μ+k×σ Among them, threshold represents the non-parametric adaptive dynamic threshold, μ represents the mean of the reconstruction error of the training set, σ represents the standard deviation of the reconstruction error of the validation set, and k is a constant used to determine the sensitivity of the threshold.

Images