A bearing fault diagnosis method and system based on dynamic pattern fusion

Abstract

The present invention belongs to the field of fault diagnosis. The present invention provides a bearing fault diagnosis method and system based on dynamic pattern fusion. The present invention uses a deterministic learning algorithm to extract local dynamic information of bearing vibration signals, and obtains a more comprehensive and in-depth dynamic knowledge expression for each category through cluster analysis and knowledge fusion, and performs dynamic pattern matching based on the minimum residual principle to achieve bearing fault diagnosis. The present invention improves the generalization ability of the model to new working conditions by integrating the dynamic information of different samples, and shows significant performance improvement in the accuracy, robustness, interpretability and adaptability of fault detection to complex working conditions.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and in particular relates to a bearing fault diagnosis method and system based on dynamic mode fusion. Background Art

[0002] In the modern industrial system, bearings are key components in rotating machinery and their operating status is directly related to the stability, safety and efficiency of the entire system. If bearing failures are not discovered and handled in a timely manner, they often lead to equipment shutdown, damage, and even serious accidents, causing huge economic losses and safety hazards. Therefore, the development of efficient and accurate bearing fault diagnosis technology is of immeasurable value in ensuring the healthy operation of equipment, preventing major accidents, and improving production efficiency.

[0003] At present, the methods of bearing fault diagnosis are mainly divided into traditional signal processing methods and data-driven intelligent diagnosis methods. Traditional signal processing methods mainly include time domain analysis, frequency domain analysis and time-frequency analysis. These methods can reflect the operating status of the bearing in real time, but are easily interfered by environmental noise and rely on professional signal processing knowledge and experience. They have certain limitations in the extraction and analysis of fault features. Data-driven intelligent diagnosis methods, such as support vector machines and deep neural networks, can automatically learn and extract fault features from large amounts of data, improving the accuracy and efficiency of diagnosis. These methods reduce the reliance on domain expertise and show significant advantages in processing massive data. They have been widely used in bearing fault diagnosis.

[0004] However, although data-driven methods have shown good performance in fault diagnosis, they still face challenges in practical applications. First, the collection of bearing fault data under actual working conditions is easily restricted, resulting in insufficient training data and difficulty in building an effective diagnostic model. Second, existing diagnostic methods have limited ability to extract fault features under complex working conditions, especially in harsh environments such as variable working conditions, heavy loads, high speeds, and strong noise. The generalization and real-time performance of the model are poor, and misdiagnosis and missed diagnosis are prone to occur. In addition, the model obtained based solely on data training is usually regarded as a "black box" model, and its results are difficult to be convincing, so improving the interpretability of the model remains a challenge. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a bearing fault diagnosis method and system based on dynamic pattern fusion. The present invention uses a deterministic learning algorithm to extract the local dynamic information of the bearing vibration signal, and obtains a more comprehensive and in-depth dynamic knowledge expression for each category through cluster analysis and knowledge fusion, and further performs dynamic pattern matching based on the minimum residual principle to achieve bearing fault diagnosis. By integrating the dynamic information of different samples, the present invention improves the generalization ability of the model to new working conditions, and shows significant performance improvements in the accuracy, robustness, interpretability and adaptability of fault detection to complex working conditions.

[0006] According to some embodiments, the present invention adopts the following technical solutions:

[0007] A bearing fault diagnosis method based on dynamic pattern fusion includes the following steps:

[0008] Obtain vibration signal data of different fault types at multiple measuring points and under various working conditions of the bearing, and pre-process the data;

[0009] Use deterministic learning algorithms to perform local accurate dynamic modeling on each preprocessed vibration signal, extract the inherent dynamic information of the data, express it in the form of constant weights of the neural network, and form a dynamic pattern library;

[0010] For different dynamic patterns in the pattern library, a dynamic similarity measurement matrix is ​​constructed, and dynamic patterns with similar dynamic characteristics are grouped into the same cluster through cluster analysis to obtain different clusters;

[0011] For several dynamic modes in the same cluster, the weight fusion algorithm is used to perform dynamic mode fusion;

[0012] The vibration signal to be diagnosed is dynamically compared with the fused dynamic pattern library, and dynamic pattern matching is performed based on the minimum residual principle to obtain the fault diagnosis result.

[0013] As an optional implementation, the multiple measuring points are arranged at different positions of the bearing;

[0014] Multiple operating conditions include different operating environments of the faulty bearing, including operating environments with different speeds and / or loads;

[0015] The different fault types include bearing inner race fault, outer race fault, ball fault and several fault types among no fault.

[0016] As an optional implementation, the preprocessing process includes using a high-gain observer to expand the original single-dimensional data into two dimensions; performing band-pass filtering on the expanded two-dimensional data; and performing maximum value normalization processing on the band-pass filtered data.

[0017] As an optional implementation, the process of using a deterministic learning algorithm to perform local accurate dynamic modeling on each pre-processed vibration signal includes:

[0018] The two-dimensional vibration signal forms a quasi-periodic / regression trajectory in the phase space to meet the continuous excitation condition, and the RBF neural network along the state trajectory is used to realize the dynamics of the unknown system. Modeling:

[0019] ;

[0020] in, is the unknown dynamics of the nonlinear system, is the weight of the RBF neural network after training convergence, is the regression vector of the RBF neural network, is the modeling error;

[0021] All vibration signals in the data set are analyzed in turn. Perform kinetic modeling to obtain A dynamic pattern library of neural network weights ,in, Represents the number of dynamic modes.

[0022] As an optional implementation method, in the process of using a deterministic learning algorithm to perform local accurate dynamic modeling on each preprocessed vibration signal, vibration signals at different measuring points and under different working conditions are modeled separately, each vibration signal corresponds to a neural network constant weight, and a dynamic pattern library is formed.

[0023] As an optional implementation, the process of constructing a dynamic similarity measurement matrix includes:

[0024] Dynamic pattern library based on vibration signal , construct a set of dynamic estimators:

[0025] ;

[0026] in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, is the vibration signal sampling period, is the RBF neural network regression vector;

[0027] The original vibration signal corresponding to one of the dynamic modes is input into the dynamic estimator to generate the dynamic recognition residual of the dynamic mode relative to other dynamic modes in the mode library, and the recognition residual is calculated. The norm is used as the dynamic distance to measure the similarity between dynamic patterns;

[0028] After calculating the kinetic distances between all dynamic patterns, they are formed into a matrix form, and cluster analysis is performed based on the kinetic similarity matrix.

[0029] As an optional implementation, the process of grouping dynamic patterns with similar dynamic characteristics into the same cluster by cluster analysis includes:

[0030] S1: Select a group of samples as the initial cluster center points;

[0031] S2: Assign the remaining samples to the cluster where the nearest cluster center is located;

[0032] S3: Calculate the sum of the kinetic distances from each sample in each cluster to the remaining samples, and select the sample with the smallest sum of kinetic distances as the new center point of the cluster;

[0033] S4: Calculate the sum of the kinetic distances of all clusters , used to evaluate clustering quality;

[0034] S5: Repeat steps S2-S4 until the sum of the kinetic distances It becomes stable, or the center point no longer changes, or a pre-specified number of iterations is reached.

[0035] As an optional implementation, the process of performing dynamic mode fusion using a weight fusion algorithm includes:

[0036] For all the similar / close dynamic patterns in each cluster, solve the neural network weights that can fuse all the similar / close dynamic patterns, and fuse the network weights The following equation is satisfied:

[0037] ;

[0038] in, Indicates The weight of the neural network after cluster fusion, Indicates The first Dynamic mode, It is The number of dynamic modes in a cluster, Indicates The first The vibration signal corresponding to the dynamic mode is the input of the neural network. Represents the RBF neural network regression vector corresponding to the dynamic mode.

[0039] Solve the above equation using the least squares method based on subspace constraints:

[0040] ;

[0041] in, , , , subscript Represents the subspace corresponding to the regressor vector composed of neurons with high excitation levels.

[0042] As an optional implementation, the vibration signal to be diagnosed is compared with the fused dynamic pattern library through a dynamic estimator, wherein the dynamic estimator is constructed through the fused dynamic pattern library:

[0043] ;

[0044] in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, After fusion Dynamic mode weights, is the sampling period of the vibration signal to be diagnosed, is the RBF neural network regression vector.

[0045] The vibration signal to be diagnosed is used as input, and the dynamic estimator is used to generate the dynamic identification residual of the vibration signal relative to all dynamic modes in the mode library.

[0046] A bearing fault diagnosis system based on dynamic pattern fusion, comprising:

[0047] A signal acquisition module is configured to acquire vibration signal data of different fault types at multiple measuring points of the bearing under multiple working conditions, and pre-process the data;

[0048] The dynamic pattern library building module is configured to use a deterministic learning algorithm to perform local accurate dynamic modeling on each preprocessed vibration signal, extract the inherent dynamic information of the data, express it in the form of constant weights of a neural network, and form a dynamic pattern library;

[0049] The cluster analysis module is configured to construct a dynamic similarity measurement matrix for different dynamic patterns in the pattern library, and classify dynamic patterns with similar dynamic characteristics into the same cluster through cluster analysis to obtain different clusters;

[0050] The dynamic mode fusion module is configured to perform dynamic mode fusion using a weight fusion algorithm for a number of dynamic modes in the same cluster;

[0051] The fault diagnosis module is configured to dynamically compare the vibration signal to be diagnosed with the fused dynamic pattern library, perform dynamic pattern matching based on the minimum residual principle, and obtain a fault diagnosis result.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention uses a deterministic learning algorithm to mine the intrinsic system dynamics of bearing signals, thereby extracting fault dynamic features that are more sensitive and discriminative than traditional methods. The dynamic features can be accurately extracted through limited sample data, avoiding dependence on large-scale training data, and are particularly suitable for small sample scenarios such as difficult data collection.

[0054] 2. The cluster analysis and pattern fusion method based on dynamic information proposed in this invention integrates the dynamic information of different samples to form a more comprehensive and in-depth dynamic knowledge expression. The dynamic modeling scope of the existing deterministic learning is expanded from the local space to the global space, realizing the establishment of a large-scale effective dynamic model, enhancing the generalization ability of the model to new working conditions, and significantly improving the accuracy, robustness and adaptability of fault detection to complex working conditions.

[0055] 3. The present invention extracts dynamic features from vibration signals and implements fault detection by measuring dynamic differences. This strategy is consistent with the existing conclusion that "system failures originate from the change of dynamic characteristics from healthy state to fault state" and has significant advantages in the interpretability of diagnostic decisions.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0058] Figure 1 is a flow chart of a method in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of dynamic mode fusion in an embodiment of the present invention;

[0060] Figure 3 4 is a system structure block diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0062] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0063] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0064] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0065] First, some necessary terminology explanations:

[0066] Deterministic learning: Deterministic learning theory is a new method for machine learning of dynamic systems that can achieve locally accurate dynamic modeling with physical interpretability of time series data. Deterministic learning theory includes the following key elements:

[0067] (1) Using a localized radial basis function (RBF) neural network;

[0068] (2) For periodic / regression trajectories, the subvectors of the RBF neural network satisfy the continuous excitation condition;

[0069] (3) Accurately model the local dynamics of the system along the periodic / regression trajectory and store it in the form of constant weights of the neural network.

[0070] Embodiment 1

[0071] like Figure 1 As shown, this embodiment provides a bearing fault diagnosis method based on dynamic pattern fusion, comprising the following steps:

[0072] Step (1): Obtain vibration signal data of different fault types at multiple measuring points of the bearing under multiple working conditions, and preprocess the data.

[0073] The vibration signal data of different fault types under multiple measuring points and multiple working conditions are single-dimensional signal data collected by the accelerometer sensor. Among them, multiple measuring points are the arrangement points of the sensor, such as the outer casing and base of the mechanical equipment; multiple working conditions are to change the operating environment of the faulty bearing, such as variable speed, variable load, etc.; different fault types include bearing inner ring fault, outer ring fault, ball fault, no fault, etc.

[0074] The data preprocessing specifically includes: using a high-gain observer to expand the original single-dimensional data into two dimensions; performing band-pass filtering on the expanded two-dimensional data; and performing maximum value normalization processing on the band-pass filtered data.

[0075] Step (2): Use a deterministic learning algorithm to perform local accurate dynamic modeling on the preprocessed vibration signal, extract the inherent dynamic information of the data, and express it in the form of constant weights of the neural network. Among them, the vibration signals at different measurement points and under different working conditions are modeled separately, that is, each vibration signal corresponds to a neural network constant weight, and form a dynamic pattern library.

[0076] The vibration signal can be regarded as a nonlinear signal generated by a dynamic physical process. This dynamic physical process can be described by a class of nonlinear systems, usually represented by ordinary differential equations:

[0077] (1)

[0078] in, is the state variable of the nonlinear system; is the unknown dynamics of the nonlinear system. The real vibration signal obtained by the sensor is a continuous variable The time series obtained by sampling.

[0079] Furthermore, the local dynamics modeling of the vibration signal is carried out based on the deterministic learning algorithm. Among them, the two-dimensional vibration signal can form a quasi-periodic / regression trajectory in the phase space, thereby satisfying the continuous excitation condition, and the RBF neural network along the state trajectory is used to realize the dynamics of the unknown system. Modeling:

[0080] (2)

[0081] in, is the weight of the RBF neural network after training convergence, is the regression vector of the RBF neural network, is the modeling error.

[0082] Furthermore, all vibration signals in the data set are analyzed in turn. Perform kinetic modeling to obtain A dynamic pattern library of neural network weights .in, Represents the number of dynamic modes.

[0083] Step (3): For different dynamic modes in the mode library, a dynamic similarity measurement matrix is ​​constructed. Dynamic modes with similar dynamic characteristics are grouped into the same group (cluster) through cluster analysis, and different clusters are obtained.

[0084] The dynamic similarity measurement matrix is ​​constructed to measure the similarity between dynamic patterns. In the qualitative analysis of nonlinear dynamic systems, it is believed that the similarity between two dynamic behaviors depends on the topological equivalence of the two dynamic systems. Therefore, the similarity of dynamic patterns depends on the similarity of the dynamic patterns' intrinsic system dynamics.

[0085] Specifically, the process of calculating the similarity between dynamic patterns is as follows:

[0086] First, based on the dynamic pattern library of vibration signals obtained in step (2), , construct a set of dynamic estimators:

[0087] (3)

[0088] in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, is the vibration signal sampling period, is the RBF neural network regression vector.

[0089] Furthermore, by inputting the original vibration signal corresponding to one of the dynamic modes into this group of dynamic estimators, the dynamic identification residual of the dynamic mode relative to other dynamic modes in the mode library can be generated:

[0090] (4)

[0091] in, Indicates The dynamic mode is relative to The residuals of the dynamic identification of the dynamic modes, Indicates The dynamic mode is relative to The dynamic estimator output for each dynamic mode.

[0092] Furthermore, the identification residual is calculated The norm is used as the dynamic distance to measure the similarity between dynamic patterns:

[0093] (5)

[0094] in, Mean The range of the norm.

[0095] Furthermore, after calculating the dynamic distances between all dynamic modes, a matrix form is formed:

[0096] (6)

[0097] in, Represents the first Dynamic mode and The dynamic distances between the dynamic modes are calculated, and the dynamic similarity matrix composed of the dynamic distances can be further used for subsequent clustering analysis.

[0098] It is worth noting that the similarity measurement matrix It is an asymmetric matrix. From the above steps, we can see that the dynamic mode To dynamic mode The kinetic distance is the comparison of the vibration signal With vibration signal Dynamic modeling (i.e., dynamic mode ), while dynamic mode To dynamic mode The kinetic distance is the comparison of the vibration signal With vibration signal Dynamic modeling (i.e., dynamic mode ),so .

[0099] Since the similarity measure matrix Due to the asymmetry of , the present invention designs a special strategy to achieve clustering, that is, along the matrix Fixed row or column direction to find the dynamic distance.

[0100] The dynamic modes with similar dynamic characteristics are classified into the same cluster by cluster analysis. The specific process is as follows:

[0101] S1: Select a group of samples as the initial cluster center points;

[0102] S2: Assign the remaining samples to the cluster where the nearest cluster center is located, that is, for the matrix , based on the rows where the remaining samples are located, find the column element where the central sample is located. The column number where the minimum value is located is the nearest cluster center to be found;

[0103] S3: Calculate the sum of the kinetic distances from each sample in each cluster to the rest of the samples, and select the sample with the smallest sum of kinetic distances as the new center of the cluster, that is, for the distance matrix , extract the rows and columns where the cluster members are located to form a new square matrix, calculate the sum of each column, and take the column item corresponding to the minimum value as the new cluster center point;

[0104] S4: Calculate the sum of the kinetic distances of all clusters , used to evaluate clustering quality;

[0105] S5: Repeat steps S2-S4 until It becomes stable, or the center point no longer changes, or a pre-specified number of iterations is reached.

[0106] Of course, in other embodiments, the term "cluster" may also be changed to "group", and the meaning thereof does not change significantly, and the essential content is the same.

[0107] It is worth noting that the clustering method based on dynamic differences proposed in the present invention measures the intrinsic relationship between data from the perspective of the intrinsic dynamics of the data rather than the surface differences, which is significantly different from the traditional clustering method based on similarity metrics such as Euclidean distance and DTW distance.

[0108] After clustering, the dynamic pattern library Divided into several clusters ,in, and Relationship satisfaction ,and .

[0109] Step (4): For several dynamic modes in the same cluster, a weight fusion algorithm is used to perform dynamic mode fusion, that is, several close / similar dynamic modes are fused into a more experienced (stronger generalization) neural network. By integrating local dynamic knowledge, large-scale dynamic modeling is achieved to obtain a more comprehensive dynamic knowledge expression.

[0110] The weight fusion algorithm is to obtain a fusion network through the knowledge learned from all independent signals (state trajectories). The fusion network stores the union of the dynamics knowledge along each independent trajectory, such as Figure 2 As shown, it should be noted that Figure 2 It is numerically simulated and has no actual unit. The present invention proposes a knowledge fusion algorithm based on the least square method under subspace constraints.

[0111] Specifically, for all the close / similar dynamic patterns in each cluster, the neural network weights that can fuse all the close / similar dynamic patterns are solved. Fusion network weights The following equation is satisfied:

[0112] (7)

[0113] in, Indicates The weight of the neural network after cluster fusion, Indicates The first Dynamic mode, It is The number of dynamic modes in a cluster.

[0114] Furthermore, the least square method based on subspace constraints is used to solve the above equation:

[0115] (8)

[0116] in, represents the RBF neural network regression vector, Indicates The diagonal matrix composed of all dynamic mode weights in the clusters, Represents the RBF neural network regression subvector, subscript Represents the subspace corresponding to the regressor vector composed of neurons with high excitation levels.

[0117] It is worth noting that if the least squares method is used directly, it is difficult to guarantee the characteristics of the local RBF network, and the consequence is that neurons far away from the trajectory (unactivated neurons) will obtain relatively large weights. Therefore, the present invention improves the least squares method by adding spatial constraints, and solves the weight of equation (7) in the subspace corresponding to the activated neurons instead of the entire space.

[0118] After fusion, the dynamic pattern library Refactored to The fused pattern library realizes the expansion from local modeling to large-scale modeling and has more comprehensive dynamic expression capabilities.

[0119] Step (5): The vibration signal to be diagnosed is compared with the fused dynamic pattern library through the dynamic estimator, and dynamic pattern matching is performed based on the minimum residual principle to obtain the fault diagnosis result.

[0120] The dynamic estimator is constructed by integrating the dynamic pattern library:

[0121] (9)

[0122] in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, After fusion, Dynamic mode weights, is the sampling period of the vibration signal to be diagnosed, is the RBF neural network regression vector.

[0123] Furthermore, the vibration signal to be diagnosed is input into the group of dynamic estimators to generate the dynamic identification residual of the vibration signal relative to all dynamic modes in the mode library. The identification residual reflects the dynamic difference between modes, and similar or close modes will result in small identification residuals.

[0124] Furthermore, dynamic pattern matching is performed based on the minimum residual principle. It is considered that the dynamic pattern corresponding to the minimum residual is most similar to the vibration signal to be diagnosed, and the final fault diagnosis result is obtained.

[0125] It is worth mentioning that the dynamic estimator constructed based on the fused weights enhances the model's generalization ability to new working conditions and significantly improves the accuracy, robustness and adaptability of fault detection to complex working conditions.

[0126] Embodiment 2

[0127] like Figure 3 As shown, this embodiment provides a bearing fault diagnosis system based on dynamic pattern fusion, including the following modules:

[0128] The signal acquisition module is configured to acquire the single-dimensional vibration signal data of the bearing, wherein the vibration signal data is acquired by a plurality of accelerometer sensors arranged on the outer casing and the base of the mechanical equipment.

[0129] The data preprocessing module is configured to expand the single-dimensional vibration signal into a two-dimensional vibration signal by using an observer, and perform filtering and normalization processing on the expanded signal.

[0130] The dynamic mode library building module is configured to use a deterministic learning algorithm to establish an intrinsic dynamic model of the bearing vibration signal to form a dynamic mode library containing rich dynamic modes.

[0131] The clustering analysis module is configured to divide the dynamic patterns in the dynamic pattern library into different clusters according to dynamic similarity through clustering analysis.

[0132] The dynamic mode fusion module is configured to use a weighted fusion algorithm to fuse several local dynamic modes in the same cluster into a large-scale dynamic mode to obtain a more comprehensive dynamic knowledge expression of the bearing vibration signal.

[0133] The fault diagnosis module is configured to compare the vibration signal to be diagnosed with the fused dynamic pattern library through the dynamic estimator to obtain a fault diagnosis result.

[0134] The modules involved in the above embodiment 2 correspond to the method in embodiment 1. For the specific implementation method, please refer to the relevant description part of embodiment 1.

[0135] It can be understood that the above-mentioned units / modules can be separately or completely combined into one or several other units / modules, or one (some) of the units can be further divided into multiple functionally smaller units, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application.

[0136] In addition, the above modules of the system are divided based on logical functions. In practical applications or in some embodiments, the function of one module can also be implemented by multiple modules, or the functions of multiple modules can be implemented by one module. For example, the signal acquisition module and the data preprocessing module of the present embodiment can be combined into one signal acquisition module.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on dynamic pattern fusion, characterized in that: The following steps are involved: Obtain vibration signal data of different fault types at multiple measuring points and under various working conditions of the bearing, and pre-process the data; Use deterministic learning algorithms to perform local accurate dynamic modeling on each preprocessed vibration signal, extract the inherent dynamic information of the data, express it in the form of constant weights of the neural network, and form a dynamic pattern library; For different dynamic patterns in the pattern library, a dynamic similarity measurement matrix is ​​constructed, and dynamic patterns with similar dynamic characteristics are grouped into the same cluster through cluster analysis to obtain different clusters; For several dynamic modes in the same cluster, the weight fusion algorithm is used to perform dynamic mode fusion; Dynamically compare the vibration signal to be diagnosed with the fused dynamic pattern library, perform dynamic pattern matching based on the minimum residual principle, and obtain the fault diagnosis result; The process of dynamic mode fusion using the weight fusion algorithm includes: For all the close / similar dynamic modes in each cluster, solve the neural network weights that can fuse all the close / similar dynamic modes, and fuse the network weights The following equation is satisfied: ； in, Indicates The weight of the neural network after cluster fusion, Indicates The first Dynamic mode, It is The number of dynamic modes in a cluster, Indicates The first The vibration signal corresponding to the dynamic mode is the input of the neural network. Represents the RBF neural network regression vector corresponding to the dynamic mode; The above equation is solved using the least squares method based on subspace constraints to obtain the fused neural network weights: ； in, , , , subscript Represents the subspace corresponding to the regressor vector composed of neurons with high excitation levels.

2. A bearing fault diagnosis method based on dynamic pattern fusion as claimed in claim 1, characterized in that: The multiple measuring points are arranged at different positions of the bearing; Multiple operating conditions include different operating environments of the faulty bearing, including operating environments with different speeds and / or loads; The different fault types include bearing inner race fault, outer race fault, ball fault and several fault types among no fault.

3. A bearing fault diagnosis method based on dynamic pattern fusion as claimed in claim 1, characterized in that: The preprocessing includes using a high-gain observer to expand the original single-dimensional data into two dimensions; performing band-pass filtering on the expanded two-dimensional data; and performing maximum value normalization processing on the band-pass filtered data.

4. A bearing fault diagnosis method based on dynamic pattern fusion as claimed in claim 1, characterized in that: The process of using the deterministic learning algorithm to perform local accurate dynamic modeling on each preprocessed vibration signal includes: The two-dimensional vibration signal forms a quasi-periodic / regression trajectory in the phase space to meet the continuous excitation condition, and the RBF neural network along the state trajectory is used to realize the dynamics of the unknown system. Modeling: ； in, is the unknown dynamics of the nonlinear system, is the weight of the RBF neural network after training convergence, is the regression vector of the RBF neural network, is the modeling error; All vibration signals in the data set are analyzed in turn. Perform kinetic modeling to obtain A dynamic pattern library of neural network weights ,in, Represents the number of dynamic modes.

5. The bearing fault diagnosis method based on dynamic pattern fusion according to claim 1, characterized in that: In the process of using the deterministic learning algorithm to perform local accurate dynamic modeling on each preprocessed vibration signal, the vibration signals at different measuring points and under different working conditions are modeled separately. Each vibration signal corresponds to a neural network constant weight and constitutes a dynamic pattern library.

6. The bearing fault diagnosis method based on dynamic pattern fusion according to claim 1, characterized in that: The process of constructing the dynamic similarity measure matrix includes: Dynamic pattern library based on vibration signal , construct a set of dynamic estimators: ； in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, is the vibration signal sampling period; The original vibration signal corresponding to one of the dynamic modes is input into the dynamic estimator to generate the dynamic recognition residual of the dynamic mode relative to other dynamic modes in the mode library, and the recognition residual is calculated. The norm is used as the dynamic distance to measure the similarity between dynamic patterns; After calculating the kinetic distances between all dynamic patterns, they are formed into a matrix form, and cluster analysis is performed based on the kinetic similarity matrix.

7. The bearing fault diagnosis method based on dynamic pattern fusion according to claim 1 is characterized in that: The process of grouping dynamic patterns with similar dynamic characteristics into the same cluster through cluster analysis includes: S1: Select a group of samples as the initial cluster center points; S2: Assign the remaining samples to the cluster where the nearest cluster center is located; S3: Calculate the sum of the kinetic distances from each sample in each cluster to the remaining samples, and select the sample with the smallest sum of kinetic distances as the new center point of the cluster; S4: Calculate the sum of the kinetic distances of all clusters , used to evaluate clustering quality; S5: Repeat steps S2-S4 until the sum of the kinetic distances It becomes stable, or the center point no longer changes, or a pre-specified number of iterations is reached.

8. The bearing fault diagnosis method based on dynamic pattern fusion according to claim 1, characterized in that: The vibration signal to be diagnosed is compared with the fused dynamic pattern library through a dynamic estimator, wherein the dynamic estimator is constructed through the fused dynamic pattern library: ； in, represents the state of the dynamic estimator at the previous moment, represents the input of the dynamic estimator at the previous moment, is the dynamic estimator gain, After fusion, Dynamic mode weights, is the sampling period of the vibration signal to be diagnosed, is the RBF neural network regression vector; The vibration signal to be diagnosed is used as input, and the dynamic estimator is used to generate the dynamic identification residual of the vibration signal relative to all dynamic modes in the mode library.

9. A bearing fault diagnosis system based on dynamic pattern fusion, applying the method described in any one of claims 1 to 8, characterized in that: include: A signal acquisition module is configured to acquire vibration signal data of different fault types at multiple measuring points of the bearing under multiple working conditions, and pre-process the data; The dynamic pattern library building module is configured to use a deterministic learning algorithm to perform local accurate dynamic modeling on each pre-processed vibration signal, extract the inherent dynamic information of the data, express it in the form of constant weights of a neural network, and form a dynamic pattern library; The cluster analysis module is configured to construct a dynamic similarity measurement matrix for different dynamic patterns in the pattern library, and classify dynamic patterns with similar dynamic characteristics into the same cluster through cluster analysis to obtain different clusters; The dynamic mode fusion module is configured to perform dynamic mode fusion using a weight fusion algorithm for a number of dynamic modes in the same cluster; The fault diagnosis module is configured to dynamically compare the vibration signal to be diagnosed with the fused dynamic pattern library, perform dynamic pattern matching based on the minimum residual principle, and obtain a fault diagnosis result.

Images