A method and system for fault diagnosis of rotating machinery based on high-precision deterministic learning

Abstract

This invention discloses a method and system for diagnosing rotating machinery faults based on high-precision deterministic learning, belonging to the field of rotating machinery fault diagnosis technology. The method includes the following steps: acquiring vibration signal data of the rotating machinery and preprocessing the vibration signal data; pre-training a neural network using sampling deterministic learning theory to obtain the weight values ​​of the inherent dynamic patterns of the vibration signal data; reconstructing the sampling deterministic learning theory using a Runge-Kutta approximation model; retraining the pre-trained neural network model using the sampling deterministic learning theory incorporating the Runge-Kutta method to obtain the dynamic patterns of the vibration signals to be diagnosed; comparing the dynamic patterns of the vibration signals to be diagnosed with the dynamic patterns in a preset pattern library using a dynamic estimator to obtain the fault diagnosis result. This invention provides an efficient, high-precision, and reliable method for diagnosing rotating machinery faults, improving the accuracy of diagnostic analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rotating machinery fault diagnosis, and in particular to a rotating machinery fault diagnosis method and system based on high-precision deterministic learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] With the rapid development of science and technology and modern manufacturing technology, the functions of mechanical equipment in modern industry are becoming more and more complex, and the rotating machinery, which is the main production tool in various industries, is developing towards large-scale, intelligentization and automation, such as high-precision numerical control machine tools, generators, aircraft engines, etc. Despite these advances, due to their harsh environment, heavy load and continuous operation, rotating machinery is becoming more and more prone to failure, thus reducing the efficiency and quality of the production system, resulting in economic losses and even catastrophic casualties in some cases. Therefore, early fault prediction of the running state of rotating machinery, reliable and stable rotating machinery fault diagnosis, and timely equipment maintenance are extremely important for the progress of productivity and the safety of life and property.

[0004] The analysis of rotating machinery vibration signals is one of the main methods to solve the problem of fault diagnosis. Through the processing technology of vibration signals, important diagnostic information can be obtained from the vibration signals. Different operating states of the equipment itself can be represented by different patterns, including normal and various fault types. The essence of rotating machinery fault diagnosis can be regarded as a pattern recognition problem. However, due to the diversity and richness of vibration signals, it is almost impossible to directly perform pattern recognition on the target signal. Therefore, the existing rotating machinery fault diagnosis generally includes two steps of feature extraction and fault recognition. Common artificial feature extraction methods include wavelet decomposition, empirical mode decomposition, etc. With the development of artificial intelligence technology, convolutional neural networks, as a method of automatically extracting features, have also been introduced into fault diagnosis and have achieved good fault detection results. From a system perspective, rotating machinery is a nonlinear system, and existing research shows that when the system fails, it is often because the healthy dynamics has changed to fault dynamics. However, the existing automatic feature extraction method of convolutional neural networks is not accurate enough for the dynamic pattern recognition of this nonlinear system, which affects the fault diagnosis results in the pattern recognition stage. Therefore, how to obtain the internal dynamic pattern of the rotating machinery vibration signal and achieve fast and accurate fault diagnosis has become a problem to be solved in the existing technology. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide a rotating machinery fault diagnosis method and system based on high-precision deterministic learning, which combines the fourth-order Runge-Kutta method in numerical analysis to enable the deterministic learning mechanism to obtain a higher-precision system identification result, and further utilizes the obtained accurate dynamic mode to complete rotating machinery fault diagnosis through a dynamic pattern recognition method. The present application provides an efficient, high-precision and reliable method for rotating machinery fault diagnosis, and improves the accuracy of diagnostic analysis.

[0006] In order to achieve the above-mentioned purpose, the present application is implemented by the following technical solutions:

[0007] The present application provides a rotating machinery fault diagnosis method based on high-precision deterministic learning in the first aspect, comprising the following steps:

[0008] Obtaining vibration signal data of the rotating machinery, and pre-processing the vibration signal data;

[0009] Extracting vibration signal dynamics information from the pre-processed vibration signal data by using a sampling deterministic learning theory fused with a Runge-Kutta method to obtain a vibration signal dynamics mode to be diagnosed; wherein the vibration signal dynamics information extraction process is as follows: pre-training a neural network by using the sampling deterministic learning theory, identifying the system dynamics of the pre-processed vibration signal data to obtain a constant radial basis function neural network expression of the internal dynamics mode of the vibration signal data; reconstructing the sampling deterministic learning theory by using a Runge-Kutta approximation model, re-training the pre-trained neural network model by using the sampling deterministic learning theory fused with the Runge-Kutta method, processing the weights of the re-trained neural network, and obtaining the vibration signal dynamics mode to be diagnosed;

[0010] Comparing the vibration signal dynamics mode to be diagnosed with the dynamics modes in the preset mode library through a dynamic estimator to obtain a fault diagnosis result.

[0011] Further, the specific steps of pre-processing the vibration signal data are as follows:

[0012] Performing high-gain observation on the vibration signal data to expand the signal dimension;

[0013] Performing normalization processing on each dimension of the vibration signal.

[0014] Further, the specific steps of pre-training the neural network by using the sampling deterministic learning theory are as follows:

[0015] Expressing the rotating machinery vibration signal by using an Euler approximation model, and constructing a dynamic neural network identifier;

[0016] Under the condition of continuous incentive, the dynamic neural network identifier is updated by using a parameter updating algorithm, so that the neural network parameters converge to a small neighborhood of the ideal true value.

[0017] The updated dynamic neural network identifier can locally and accurately approximate the unknown dynamics of the vibration signal trajectory in the form of a constant neural network.

[0018] Further, the neural network is a radial basis function neural network.

[0019] Further, the parameter updating algorithm is designed based on Lyapunov stability theory.

[0020] Further, the sampling-determined learning theory is reconstructed by using the Runge-Kutta approximation model, and the specific steps of using the sampling-determined learning theory combined with the Runge-Kutta method to retrain the pre-trained neural network model are as follows:

[0021] The Runge-Kutta approximation model is used to express the vibration signal, and the neural network identifier with the structure of the Runge-Kutta method is reconstructed;

[0022] The weight values of the internal dynamics mode of the pre-trained vibration signal data are used as the initial weights to retrain the neural network.

[0023] Further, the construction steps of the preset mode library are as follows: saving all the dynamics modes of the trained rotating machinery vibration signals in the form of constant neural network weight values to form a mode library.

[0024] Further, the construction process of the dynamic estimator is as follows:

[0025] Considering the constant neural network weight value vector saved in the mode library, the constant neural network of the vibration signal is reconstructed;

[0026] For each set of dynamics mode weight value vector, a dynamic estimator is built based on the constructed neural network;

[0027] Each set of dynamic estimator of the sampling data represents a vibration signal dynamics mode in the mode library.

[0028] Further, the specific steps of comparing the to-be-diagnosed vibration signal dynamics mode with the dynamics mode in the preset mode library through the dynamic estimator to obtain the fault diagnosis result are as follows:

[0029] The to-be-diagnosed vibration signal is input into the dynamic estimator in a parallel manner, the generated estimation mode is compared with the to-be-diagnosed mode, and the identification error corresponding to different dynamics modes in the vibration signal mode library is generated;

[0030] The norm of the identification error is calculated, and the identification error norms of the dynamics modes of the same category are integrated.

[0031] According to the integrated identification error norm, the most similar signal dynamics mode in the mode library is determined, and a fault diagnosis result is obtained.

[0032] The second aspect of the present application provides a rotating machinery fault diagnosis system based on high-precision deterministic learning, comprising:

[0033] A data acquisition module is configured to acquire vibration signal data of the rotating machinery and pre-process the vibration signal data.

[0034] A feature extraction module is configured to extract vibration signal dynamics information from the pre-processed vibration signal data by using a sampling deterministic learning theory fused with a Runge-Kutta method, to obtain a vibration signal dynamics mode to be diagnosed.

[0035] A fault diagnosis module is configured to compare the vibration signal dynamics mode to be diagnosed with dynamics modes in a preset mode library through a dynamic estimator, to obtain a fault diagnosis result.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] The present application discloses a rotating machinery fault diagnosis method and system based on high-precision deterministic learning, which can realize accurate identification of the internal dynamics of rotating machinery vibration signals with high precision, and can obtain a fault diagnosis result with high accuracy.

[0038] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0039] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, together with its

[0040] Figure 1 Flow chart of the rotating machinery fault diagnosis method based on high-precision deterministic learning in embodiment one of the present application;

[0041] Figure 2 Data collection example diagram in embodiment one of the present application;

[0042] Figure 3 Schematic diagram of training neural network again by fusing Runge-Kutta method and deterministic learning theory in embodiment one of the present application;

[0043] Figure 4 Schematic diagram of fault diagnosis result in embodiment one of the present application. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed description is exemplary in nature, intended to provide further description of the application. Unless otherwise defined, 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 belongs.

[0045] It is also important to note that the use of the term "example" herein is used in the sense of serving as an example, instance, or illustration. That is, it is not meant to be construed as limiting in any way. It should also be noted that the term "or" is used herein in the inclusive sense, such that any of the described options are acceptable under the present application unless otherwise indicated. In addition, it should be noted that the term "comprising" is used herein in the inclusive sense, such that it is intended to be construed to include the presence of features, steps, operations, devices, components, and / or combinations thereof;

[0046] Embodiment one:

[0047] Runge-Kutta method is an iterative method for solving nonlinear ordinary differential equations in numerical analysis, which is widely used in engineering and is a high-precision single-step algorithm. For differential equation (1), the general form of typical fourth-order Runge-Kutta is shown in formula (2).

[0048]

[0049]

[0050] The method of the vibration signal modeling and identification stage of the present application is derived from the deterministic learning mechanism. The deterministic learning theory provides an effective method for coping with the challenges in knowledge acquisition, expression and utilization in dynamic systems. Based on the deterministic learning theory, the present application further proposes a Runge-Kutta high-precision identification deterministic learning mechanism. The proposed theory can realize local accurate identification or modeling of a nonlinear dynamic system when the periodic or regression trajectory satisfies the sustained excitation condition. The modeling dynamics can be expressed and stored by a constant RBF neural network. Moreover, the stored knowledge can be effectively used for tasks in dynamic environments, such as dynamic pattern recognition and intelligent control.

[0051] From the system perspective, a rotating machine can be regarded as a nonlinear dynamic system, and the vibration signal generated by rotation can be represented by a system of partial differential equations. Therefore, the vibration time series signal generated by a rotating machine hides rich nonlinear dynamic information. Due to the harsh environment, heavy load and continuous operation, the normal vibration state changes to the fault state, and the nonlinear dynamic information of the rotating machine changes accordingly. The essence of the rotating machine fault diagnosis method based on the Runge-Kutta algorithm high-precision identification deterministic learning mechanism is how to effectively save and distinguish the dynamic patterns of normal and fault vibrations generated by the nonlinear dynamic system of the rotating machine, and quickly diagnose the fault of the rotating machine vibration signal to be detected. The specific steps are as follows:

[0052] Embodiment one of the present application provides a rotating machine fault diagnosis method based on high-precision deterministic learning, as shown in 1, comprising the following steps:

[0053] Step 1, collecting the running data of the rotating machine, preferably the vibration signal data, determining the sampling period T of the sensor, and pre-processing the vibration signal data.

[0054] Step 1.1, the collected vibration signal is at the driving end of the motor, and the sampling frequency is determined according to the used sensor. As shown in 2, it is an example of the collected vibration signal data. Figure 2

[0055] Step 1.2, high-gain observation is performed on the collected normal and different fault type rotating machine vibration signal data, the signal dimension is expanded, and the vibration signal data of the second state of the rotating machine nonlinear system is obtained. The purpose of the high-gain observer is to observe the internal state from the collected one-dimensional vibration signal, and the observed signal is taken as the second state in the rotating machine nonlinear system, or multiple sensors can be installed near the rotating parts of the rotating machine to collect multi-point data to form a regression trajectory, which satisfies the sustained excitation condition in the deterministic learning mechanism.

[0056] ​Step 1.3, normalize the vibration signal of each dimension, and obtain the vibration signal and the amplitude of the signal, which is the basis for subsequent determination of the learning mechanism to build neural network.

[0057] Step 2, use the sampling determination learning theory of the fused Runge-Kutta method to extract the vibration signal dynamics information of the preprocessed vibration signal data, and obtain the vibration signal dynamics mode to be diagnosed. The determination learning of the sampling data provides an effective method for modeling the internal dynamics in the sampling data sequence. In this embodiment, the neural network selects a radial basis function (RBF) neural network. The RBF network is a feedforward neural network with a hidden layer. As long as a sufficient number of neuron nodes are arranged, it can approximate any continuous function, i.e. dynamic mode, near the trajectory.

[0058] Step 2.1: pre-train the neural network using the sampling determination learning theory, and perform system dynamics identification on the preprocessed vibration signal data to obtain the constant radial basis function neural network expression of the internal dynamics mode of the vibration signal data. The neural network pre-training ensures the local accurate approximation of unknown dynamics, so that the determination learning mechanism with the Runge-Kutta method can meet the continuous incentive condition in the secondary learning process, and ensure that the parameters can converge to the ideal true value.

[0059] Step 2.1.1, use the Euler approximation model to express the rotating machinery vibration signal, and construct a dynamic neural network identifier.

[0060] In a specific embodiment, the nonlinear dynamics of the rotating machinery vibration signal is expressed as follows:

[0061]

[0062] where x = [x1, x2] T ∈ R n , n = 2, represents the state vector of the nonlinear dynamics system (3), f(x) = [f1(x; p s ), f2(x; p s )] T represents the internal dynamics of a rotating machinery vibration signal, p s represents the system parameter vector.

[0063] Use the Euler approximation model to express the discrete vibration signal for one of the states in the system:

[0064] x i (k+1) = x i (k) + Tf(x(k)) (4)

[0065] In order to accurately identify the internal dynamics of the vibration signal, the following dynamic neural network identifier is constructed,

[0066]

[0067] wherein, represents the state of the dynamic neural network identifier at the kth moment, 0 < a < 1 represents the identifier gain to be designed, represents the RBF neural network for learning the system dynamics, is the weight to be estimated, S(x(k)) is a regression vector composed of radial basis functions, and T represents the sampling time.

[0068] Step 2.1.2, under the condition of meeting the persistent excitation condition (PE), selecting appropriate identifier gain and learning gain, and using a parameter updating algorithm to update the dynamic neural network identifier, so that the neural network parameters converge to a small neighborhood of the ideal true value; wherein the parameter updating algorithm is designed according to Lyapunov stability theory.

[0069] The parameter convergence of the above neural network has two cases: one is that the neurons along the trajectory of the rotating machinery system satisfy the persistent excitation condition, and their weights converge to the vicinity of the ideal true value; the other is that the neurons far away from the system trajectory are not excited, and their weights cannot be updated and remain the initial value, which is almost zero.

[0070] The unknown dynamics of the updated dynamic neural network identifier along the vibration signal trajectory can be locally accurately approximated in the form of a constant neural network.

[0071] In a specific embodiment, the weight estimation of the neural network is updated in the following manner, i represents the ith dimension,

[0072]

[0073] wherein γ represents the learning gain, represents the tracking error.

[0074] By selecting appropriate identifier gain a and learning gain γ, the unknown dynamics f i (x; p s ) along the trajectory of the rotating machinery vibration signal can be locally accurately approximated in the form of a constant RBF neural network .

[0075]

[0076] wherein, is the average weight of the neural network weight after convergence for a period of time, and ∈' is the modeling error of the RBF neural network.

[0077] The RBF neural network is pre-trained by the sampling determination learning method of step 2.1, and an accurate model along the state trajectory of the vibration signal can be obtained. However, this method is built on the basis of the Euler model, and the Euler formula is a first-order method, and the highest theoretical precision can only reach O (h 2 ), wherein h is an iteration step length. The model established by this method is not accurate enough, and the error is also gradually accumulated with continuous estimation of the state. In order to further solve this pain point, the present application proposes a new method and solution based on the sampling data determination learning algorithm. First, the sampling determination learning algorithm based on the Euler model is used for pre-training to ensure the local accurate approximation of unknown dynamics, and then the Runge-Kutta method is introduced to establish a more accurate sampling data expression model to improve the identification accuracy of unknown system dynamics.

[0078] Step 2.2, after the local accurate approximation of the internal dynamics mode of the vibration signal in step 2.1 is completed, the Runge-Kutta method is introduced to establish a more accurate sampling data expression model to improve the identification accuracy of unknown system dynamics. The sampling determination learning theory is reconstructed by using the Runge-Kutta approximation model, the pre-trained neural network model is trained again by using the sampling determination learning theory fused with the Runge-Kutta method, and the weights of the neural network after retraining are processed to obtain the vibration signal dynamics mode to be diagnosed.

[0079] It should be particularly pointed out that the secondary learning is not equal to simply learning twice by using the determination learning mechanism, and the secondary learning effect of the introduced Runge-Kutta precision improvement method is obviously better than that of simply learning twice by using the determination learning mechanism.

[0080] Step 2.2.1, the vibration signal is expressed by using the Runge-Kutta approximation model, and a neural network identifier with the structure of the Runge-Kutta method is reconstructed;

[0081] Step 2.2.2, the weight values of the internal dynamics mode of the vibration signal data obtained by pre-training are used as initial weights to train the neural network again. Specifically, the mean values of each weight in a period after the convergence of the weight values of the dynamic RBF neural network are saved to generate a constant RBF neural network, the parameters of the obtained constant RBF neural network are used as the initial weights of secondary learning, and the internal dynamics mode of the rotating machinery vibration signal is represented. The above dynamics approximation is performed on all rotating machinery vibration signal samples to obtain the dynamics mode of various samples.

[0082] In a specific embodiment, the discrete vibration signal can be expressed in the following form by fusing the Runge-Kutta method:

[0083]

[0084] wherein,

[0085] f N1 = f i (x(k))

[0086]

[0087]

[0088] f N4 = f i (x(k)+Tf N3 )

[0089] The dynamic neural network identifier using the Runge-Kutta method is built, and the sampling of the above steps determines the average weight value learned by pre-training As the initial weight of the newly built neural network identifier.

[0090]

[0091] Wherein,

[0092]

[0093]

[0094]

[0095]

[0096] Indicates the state of the dynamic neural network identifier, β indicates the gain of the identifier to be designed, N f1 ,N f2 ,N f3 ,N f4 Indicates the RBF neural network used to learn the system dynamics, wherein Indicates the weight to be estimated of the RBF neural network, S: R n →R m Indicates the regression vector, and the dimension m is the number of neuron nodes of the RBF neural network. The Runge-Kutta precision improvement method neural network structure diagram is as shown in Figure 3

[0097] The weight update rate of the neural network becomes:

[0098]

[0099] Wherein,

[0100]

[0101] ​​Step 3, compare the vibration signal dynamics mode to be diagnosed with the dynamics mode in the preset mode library through a dynamic estimator to obtain a fault diagnosis result.

[0102] Step 3.1, constructing a mode library. Save the dynamics mode of all training rotating machinery vibration signals in the form of constant RBF neural network weights to form a mode library. Among them, all training rotating machinery vibration signals include the normal operation state of rotating machinery and the operation state in the case of different faults. Each operation state uses multiple sets of sampling data to enrich the dynamics mode of each operation state in the mode library and improve the detection efficiency of fault detection.

[0103] In a specific embodiment, the average value of the weight vector of the neural network after secondary learning and convergence represents the internal dynamics mode of the rotating machinery vibration signal, which is saved to the mode library, and the selection method is as follows:

[0104]

[0105] Step 3.2, constructing a dynamic estimator. The dynamic estimator can quickly recall the internal dynamics mode of the vibration signal learned through the deterministic learning mechanism, and the state of the dynamic estimator is the dynamic behavior of the rotating machinery vibration signal dynamics mode. In this way, the knowledge learned through the above high-precision identification of Runge-Kutta deterministic learning can be used again in the identification stage.

[0106] Step 3.2.1, considering the constant RBF neural network weight vector saved in the mode library, reconstruct the constant RBF neural network of the vibration signal;

[0107] Step 3.2.2, build a dynamic estimator for each set of dynamics mode weight vector combined with the constructed neural network; the dynamic estimator of each set of sampling data represents a vibration signal dynamics mode in the mode library.

[0108] In a specific embodiment, the dynamic estimator is represented as:

[0109]

[0110] Among them,

[0111]

[0112]

[0113]

[0114] Among them, represents the state of the dynamic estimator of the vibration signal, superscript s represents the dynamic mode, subscript i represents the i-th dimension of the dynamic mode, represents the input of the dynamic estimator of the vibration signal, i.e. the vibration signal of the rotating machinery to be diagnosed, b i is a positive number less than 1, representing the gain of the estimator, N s1 is the number of dimensions of the dynamic mode, s2 is the number of dimensions of the dynamic mode, s3 is the number of dimensions of the dynamic mode, s4 represents the learned constant RBF neural network, is the dynamic mode saved in the mode library of the vibration signal of the rotating machinery, i.e. the parameters of the learned constant RBF neural network determined by the learning.

[0115] Step 3.3, compare the dynamic mode of the vibration signal to be diagnosed with the dynamic mode in the preset mode library through the dynamic estimator to obtain the fault diagnosis result.

[0116] Step 3.3.1, construct the recognition error. In a parallel manner, input the vibration signal to be diagnosed into the dynamic estimator, compare the generated estimation mode with the mode to be diagnosed, and generate the recognition error corresponding to different dynamic modes in the mode library of the vibration signal. The recognition error reflects the size of the dynamic difference between the modes. Small dynamic difference between the modes will result in small recognition error.

[0117] In a specific embodiment, the output generated by the dynamic estimator is subtracted from the mode of the rotating machinery to be diagnosed to obtain the recognition error of different training modes in the mode library:

[0118]

[0119] wherein, represents the recognition error corresponding to the i-th dimension of the s-th mode in the mode library.

[0120] Step 3.3.2, classify the rotating machinery fault. Calculate the L1 norm of the recognition error and integrate the recognition error norm of the dynamic modes of the same category. The L1 norm of the vibration signal to be diagnosed and all the dynamic modes of the vibration signal in the mode library needs to be calculated before each fault diagnosis.

[0121] In a specific embodiment, the rotating machinery fault is classified: take the L1 norm of the constructed recognition error:

[0122]

[0123] wherein, the interval [k n ,k m ] is the value range of the recognition error for calculating the L1 norm.

[0124] Pattern integration is performed on the same category of dynamic patterns:

[0125]

[0126] wherein M is the number of the same category of dynamic patterns in the pattern library, N is the total dimension of the same dynamic pattern, E M is the average of the integrated identification error L1 norm.

[0127] Step 3.3.3, determine the most similar signal dynamic pattern in the pattern library according to the integrated identification error norm, and obtain the fault diagnosis result.

[0128] The final calculated integrated result E M is taken as the decision index of fault diagnosis. The norm of the constructed identification error is taken because the same category of vibration signal dynamic patterns is composed of multiple vibration signal samples, so the pattern integration is performed on the same category of dynamic patterns, and the average of the identification error L1 norm corresponding to which category is the smallest after integration, which means that the vibration signal dynamic pattern of the to-be-diagnosed rotating machine is the most similar to the corresponding category in the pattern library, and finally the result of rotating machine fault diagnosis is given. The data length for generating the average L1 norm of the identification error is a hyperparameter that can be adjusted, and a small data length can make the fault diagnosis process faster.

[0129] The idea of rotating machine fault diagnosis in this embodiment is that the L1 norm of the to-be-diagnosed rotating machine vibration signal and all vibration signal dynamic patterns in the pattern library is calculated before each fault diagnosis, the pattern integration is performed on the same category of dynamic patterns, the average E M of the identification error L1 norm corresponding to which category is the smallest after integration, which means that the vibration signal dynamic pattern of the to-be-diagnosed rotating machine is the most similar to the corresponding category in the pattern library, and finally the result of rotating machine fault diagnosis is given as shown in Figure 4 , and category 1 is the identification result of the to-be-diagnosed vibration signal.

[0130] The determination learning method based on the Runge-Kutta high recognition accuracy in the application can realize accurate recognition of the internal dynamics of the rotating machine vibration signal with high accuracy, and can obtain a high-accuracy fault diagnosis result. The single-hidden-layer neural network is used, the training process is fast and efficient, and compared with most existing deep learning fault diagnosis methods, the resource occupation is small. In the recognition stage, a parallel algorithm is used, and online learning is supported, which can quickly detect the occurrence of faults. In the fault recognition stage, an integrated method is used to process the identification errors generated by different patterns of the same category, so that the diagnosis result is more reliable and accurate.

[0131] Embodiment two:

[0132] The embodiment two of the present application provides a rotating machinery fault diagnosis system based on high-precision deterministic learning, comprising:

[0133] a data acquisition module configured to acquire vibration signal data of the rotating machinery, and to pre-process the vibration signal data;

[0134] a feature extraction module configured to extract vibration signal dynamics information of the pre-processed vibration signal data by using a sampling deterministic learning theory fused with a Runge-Kutta method, to obtain a vibration signal dynamics mode to be diagnosed; wherein the vibration signal dynamics information extraction process is as follows: neural network pre-training is performed by using the sampling deterministic learning theory, system dynamics identification is performed on the pre-processed vibration signal data, and a constant radial basis function neural network expression of the internal dynamics mode of the vibration signal data is obtained; the Runge-Kutta approximation model is used to reconstruct the sampling deterministic learning theory, the pre-trained neural network model is trained again by using the sampling deterministic learning theory fused with the Runge-Kutta method, the weights of the neural network after the re-training are processed, and the vibration signal dynamics mode to be diagnosed is obtained.

[0135] a fault diagnosis module configured to compare the vibration signal dynamics mode to be diagnosed with dynamics modes in a preset mode library through a dynamic estimator, to obtain a fault diagnosis result.

[0136] The steps and the method of the embodiment two are corresponding to the steps and the method of the embodiment one, and the specific implementation can be understood with reference to the related description of the embodiment one.

[0137] Those skilled in the art should understand that each module or each step of the present application described above can be realized by using a general computer device, and alternatively, they can be realized by using program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be respectively manufactured into each integrated circuit module, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0138] The specific embodiments of the present application are described above in combination with the accompanying drawings, but the description is not a limitation on the protection scope of the present application, and those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A rotating machine fault diagnosis method based on high-precision deterministic learning, characterized by, The method comprises the following steps: obtaining vibration signal data of a rotating machine, and preprocessing the vibration signal data; extracting vibration signal dynamics information from the preprocessed vibration signal data by using a sampling-determined learning theory fused with a Runge-Kutta method, to obtain a vibration signal dynamics mode to be diagnosed; wherein the vibration signal dynamics information extraction process is as follows: pre-training a neural network by using the sampling-determined learning theory, and performing system dynamics identification on the preprocessed vibration signal data, to obtain a constant radial basis function neural network expression of an internal dynamics mode of the vibration signal data; reconstructing the sampling-determined learning theory by using a Runge-Kutta approximation model, and re-training the pre-trained neural network model by using the sampling-determined learning theory fused with the Runge-Kutta method, and the specific steps are as follows: expressing the vibration signal by using the Runge-Kutta approximation model, and the formula is as follows: ； wherein denotes the state at the kth time instant, T denotes the sampling time, , , , ; reconstructing a neural network identifier with the structure of the Runge-Kutta method, and the formula is as follows: ； wherein, , , , , denotes the state of the dynamic neural network identifier, denotes the identifier gain to be designed, denotes the RBF neural network for learning the system dynamics, denotes the weights of the RBF neural network to be estimated, denotes the regression vector whose dimension m is the number of neuron nodes of the RBF neural network; re-training the neural network by using the pre-trained weight value of the internal dynamics mode of the vibration signal data as the initial weight; processing the weight of the re-trained neural network, to obtain the vibration signal dynamics mode to be diagnosed; comparing the vibration signal dynamics mode to be diagnosed with a dynamics mode in a preset mode library through a dynamic estimator, to obtain a fault diagnosis result.

2. The high-precision determination learning-based rotating machine fault diagnosis method according to claim 1, characterized by, The specific steps of preprocessing the vibration signal data are as follows: performing high-gain observation on the vibration signal data, and expanding the signal dimension; performing normalization processing on each dimension of the vibration signal. 3.The high-precision determination learning-based rotating machinery fault diagnosis method of claim 1, wherein, The specific steps of pre-training the neural network by using the sampling-determined learning theory are as follows: expressing the rotating machine vibration signal by using an Euler approximation model, and constructing a dynamic neural network identifier; under the condition of meeting the continuous excitation, updating the dynamic neural network identifier by using a parameter updating algorithm, so that the neural network parameters converge into a small neighborhood of the ideal true value; the unknown dynamics of the updated dynamic neural network identifier along the vibration signal trajectory can be locally accurately approximated in the form of a constant neural network.

4. The high-precision determination learning-based rotating machine fault diagnosis method according to claim 3, characterized by, The neural network is a radial basis function neural network.

5. The high-precision determination learning-based rotating machine fault diagnosis method according to claim 3, characterized by, The parameter updating algorithm is designed according to the Lyapunov stability theory.

6. The high-precision determination learning-based rotating machine fault diagnosis method of claim 1, wherein, The construction steps of the preset mode library are as follows: saving all the dynamics modes of the training rotating machine vibration signals in the form of constant neural network weight values, to form a mode library.

7. The high-precision determination learning-based rotating machine fault diagnosis method according to claim 6, characterized by, The construction process of the dynamic estimator is as follows: reconstructing the constant neural network of the vibration signal by considering the constant neural network weight value vector saved in the mode library; building a dynamic estimator for each set of dynamics mode weight vectors in combination with the constructed neural network; each set of sampling data dynamic estimator represents a kind of vibration signal dynamics mode in the mode library. 8.The high-precision determination learning-based rotating machinery fault diagnosis method of claim 7, wherein, The specific steps of comparing the vibration signal dynamics mode to be diagnosed with the dynamics mode in the preset mode library through the dynamic estimator, to obtain the fault diagnosis result, are as follows: inputting the vibration signal to be diagnosed into the dynamic estimator in a parallel manner, comparing the generated estimation mode with the mode to be diagnosed, and generating an identification error corresponding to different dynamics modes in the vibration signal mode library; calculating the norm of the identification error, and performing mode integration on the identification error norm of the same category of dynamics modes; According to the integrated identification error norm, the most similar signal dynamics mode in the mode library is determined, and a fault diagnosis result is obtained.

9. A rotating machine fault diagnosis system based on high-precision determination learning, characterized by, Comprise: The data acquisition module is configured to acquire vibration signal data of the rotating machinery, and to pre-process the vibration signal data; The feature extraction module is configured to extract vibration signal dynamics information from the pre-processed vibration signal data by using a sampling determination learning theory fused with a Runge-Kutta method, to obtain a vibration signal dynamics mode to be diagnosed; wherein the vibration signal dynamics information extraction process is as follows: neural network pre-training is performed by using the sampling determination learning theory, system dynamics identification is performed on the pre-processed vibration signal data, and a constant radial basis function neural network expression of the internal dynamics mode of the vibration signal data is obtained; the Runge-Kutta approximate model is used to reconstruct the sampling determination learning theory, the pre-trained neural network model is trained again by using the sampling determination learning theory fused with the Runge-Kutta method, and the specific steps are as follows: The vibration signal is expressed by using the Runge-Kutta approximate model, and the formula is as follows: ； wherein denotes the state at the kth time instant, T denotes the sampling time, , , , ; The neural network identifier with the structure of the Runge-Kutta method is reconstructed, and the formula is as follows: ； wherein, , , , , denotes the state of the dynamic neural network identifier, denotes the identifier gain to be designed, denotes the RBF neural network for learning the system dynamics, denotes the weights of the RBF neural network to be estimated, denotes the regression vector with dimension m being the number of neuron nodes of the RBF neural network; The weight value of the internal dynamics mode of the vibration signal data obtained by pre-training is used as the initial weight to train the neural network again; The weight of the neural network after re-training is processed, and the vibration signal dynamics mode to be diagnosed is obtained; The fault diagnosis module is configured to compare the vibration signal dynamics mode to be diagnosed with the dynamics mode in the preset mode library through a dynamic estimator, to obtain a fault diagnosis result.

Images