An intelligent fault diagnosis method and system

Abstract

The application discloses an intelligent fault diagnosis method and system, and relates to the technical field of artificial intelligence.The method comprises the following steps: obtaining vibration signals of different rotors, labeling the vibration signals, and constituting an original data set; performing feature extraction on the original data set to obtain a feature data set; processing the feature data set by using a feature selection evolution algorithm, and optimizing the processed data set by using a particle swarm optimization algorithm to obtain a key feature combination for fault diagnosis; and identifying and classifying a fault to be identified according to the key feature combination by using a K nearest neighbor classification method, so as to determine a fault category.The application can realize fault diagnosis with higher classification accuracy and fewer selected fault features.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent fault diagnosis method and system. Background Technology

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

[0003] Fault diagnosis is a crucial step in equipment health management, playing a vital role in ensuring the availability and safety of mechanical systems and preventing dangerous situations. Traditional fault diagnosis methods rely on human learning from experience data, using the correlation between fault sources and symptoms to determine if a fault has occurred. However, this approach limits fault detection to existing knowledge and cannot guarantee a certain level of accuracy. With the rapid development and widespread application of artificial intelligence, engineers have actively explored intelligent fault diagnosis. Intelligent fault diagnosis simulates the reasoning process of human thought. By effectively acquiring, transmitting, and processing diagnostic information, it can mimic human experts, using flexible diagnostic strategies to make intelligent judgments and decisions regarding the operating status and faults of monitored objects. Currently, signal-based intelligent fault diagnosis methods are receiving widespread attention due to their wide applicability and strong anti-interference capabilities.

[0004] Fault diagnosis is essentially a classification task. Its purpose is to classify fault categories based on the information presented by fault characteristics. The main steps include feature extraction, feature selection, and fault classification. Measured vibration signals are typically nonlinear and non-stationary, and useful fault feature information is often drowned out by heavy background noise. Therefore, advanced signal processing techniques are essential for effective feature extraction. Furthermore, with advancements in data collection techniques and the increasing complexity of fault problems, fault diagnosis involves more and more attributes, making computationally prohibitive. This data often contains many noisy, irrelevant, and redundant features, which severely impact the accuracy of fault diagnosis. Manually selecting sensitive features based on specific fault diagnosis problems is not portable to new problems. Therefore, there is an urgent need for effective feature selection methods to automatically find useful fault features.

[0005] In fault classification, the quantity and quality of input features significantly impact the accuracy of the classification. Therefore, selecting an appropriate subset of features as input is crucial. For irrelevant and redundant features, feature selection is an important data preprocessing technique that addresses these challenges by choosing relevant and complementary subsets of fault features. Feature selection methods integrate search techniques with classifiers, thereby linking collected data to the machine's health status. Although some feature selection methods have shown promising diagnostic performance, they still have limitations. For example, some implicit knowledge among fault features is not fully explored, the computational cost is too high due to the large number of fault features involved, and the algorithm is prone to getting trapped in local optima. Therefore, achieving accurate and low-cost intelligent fault diagnosis remains a pressing issue for current technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide an intelligent fault diagnosis method and system that can achieve higher classification accuracy and a smaller number of selected fault features for fault diagnosis.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] The first aspect of this invention provides an intelligent fault diagnosis method, comprising the following steps:

[0009] Vibration signals under different rotor conditions are acquired, and the vibration signals are labeled to form the original dataset;

[0010] Feature extraction is performed on the original dataset to obtain the feature dataset;

[0011] The feature dataset is processed using a feature selection evolutionary algorithm, and the processed dataset is then optimized using a particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0012] The K-nearest neighbor classification method is used to identify and classify faults based on key feature combinations, thereby determining the fault category.

[0013] Furthermore, the vibration signal includes vibration signals of four rotor states: normal, unbalanced, misaligned, and frictional.

[0014] Furthermore, features were extracted from the original dataset using sliding window, fast Fourier transform, and wavelet transform methods, respectively.

[0015] Furthermore, the feature dataset includes 35-dimensional features, with each column representing a fault feature, each row representing a sample, and corresponding to a class label.

[0016] Furthermore, the specific steps for processing the feature dataset using the feature selection evolutionary algorithm are as follows:

[0017] Set the parameters for the feature selection evolutionary algorithm;

[0018] Calculate the feature ReliefF weight values ​​based on the set parameters;

[0019] The inflection point selection strategy is used to select relevant features according to the feature ReliefF weight value to form the initial population;

[0020] The initial population is optimized using the particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0021] Furthermore, the specific steps for calculating the feature ReliefF weight values ​​based on the set parameters are as follows:

[0022] Normalize the weight values ​​of all features in the feature dataset;

[0023] Sort in descending order based on the feature weight values.

[0024] Furthermore, the specific steps for selecting relevant features based on the feature ReliefF weight values ​​using the inflection point selection strategy are as follows:

[0025] Arrange the features in descending order of their weights to obtain a curve about the feature weights;

[0026] Connecting the maximum and minimum weight points of the curve yields a limit line;

[0027] The point at which the distance from the limit line to the curve is the inflection point;

[0028] Features with weights greater than the inflection point are assigned to the strongly correlated feature subset, while the remaining features are assigned to the weakly correlated feature subset.

[0029] Furthermore, the initial probability is calculated based on all relevant features;

[0030] The initial population is obtained by selecting feature values ​​based on the comparison results between relevant features and initialization probabilities.

[0031] Furthermore, the specific steps for optimizing the initial population using the particle swarm optimization algorithm are as follows:

[0032] The feature dataset is divided into a training set and a test set;

[0033] Evaluate the particles and set optimization parameters;

[0034] Update the particle velocity and position based on feature correlation;

[0035] The particles in the update are evaluated by changing the size of the training set until the global optimal position is obtained as the final solution.

[0036] A second aspect of the present invention provides an intelligent fault diagnosis system, comprising:

[0037] The data acquisition module is configured to acquire vibration signals under different rotor conditions, label the vibration signals, and form the original dataset.

[0038] The feature extraction module is configured to extract features from the original dataset to obtain a feature dataset.

[0039] The feature selection module is configured to process the feature dataset using a feature selection evolutionary algorithm, and then optimize the processed dataset using a particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0040] The fault diagnosis module is configured to use the K-nearest neighbor classification method to identify and classify the fault to be identified based on the combination of key features, and determine the fault category.

[0041] The above one or more technical solutions have the following beneficial effects:

[0042] This invention discloses an intelligent fault diagnosis method and system that effectively improves the quality of the initial population and increases the depth of population analysis by utilizing the correlation between features and class labels. This invention fully develops the correlation knowledge between fault features and fault states, as well as among fault features themselves, thus providing good prior knowledge for fault diagnosis. Furthermore, to alleviate the problem of the population getting trapped in local optima, it effectively utilizes the influence of training set size on the classification model to help adjust the search direction of the population, thereby identifying key fault feature combinations for fault diagnosis. Based on key fault feature combinations from different experimental platforms, the fault states of mechanical systems can be classified, achieving high classification accuracy and significantly reducing the cost of fault diagnosis. Compared with existing feature selection methods, the fault diagnosis method of this invention achieves higher classification accuracy and selects fewer fault features, making it widely applicable in intelligent fault diagnosis.

[0043] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0044] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0045] Figure 1This is a flowchart of the intelligent fault diagnosis method in Embodiment 1 of the present invention. Detailed Implementation

[0046] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0047] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] Example 1:

[0049] Embodiment 1 of the present invention provides an intelligent fault diagnosis method, such as Figure 1 As shown, it includes the following steps:

[0050] Step 1: Obtain vibration signals under different rotor conditions, label the vibration signals, and form the original dataset.

[0051] Step 2: Extract features from the original dataset to obtain the feature dataset.

[0052] Step 3: The feature dataset is processed using the feature selection evolution algorithm, and the processed dataset is then optimized using the particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0053] Step 4: Use the K-nearest neighbor classification method to identify and classify the faults to be identified based on the combination of key features, and determine the fault category.

[0054] In step 1, the experimental rotating machinery system for signal acquisition consists of a rotor test bench, a speed controller, a front-end processor, and a computer. The rotor test bench is driven by a DC motor with a rated current of 1.95A and a maximum output power of 148W, controlled by a DH5600 speed controller. The rotor, supported by bearing supports on both sides, has a diameter of 10 mm and a length of 850 mm, including two shafts connected by couplings and two mass disks with a diameter of 75 mm. Friction screw housings for contact friction experiments and sensor mounting are mounted on the workbench. The data acquisition device uses a GTS3-TG series test simulator. The time resolution of the frequency measurement input and output channels is 8 ns. The frequency output channel has a frequency resolution of 0.00002 Hz near 50 Hz, with an error of no more than ±0.0005 Hz. The signals are acquired by two eddy current sensors fixed on sensor supports, transmitted to the front-end processor for amplification and filtering, and finally transmitted to the computer for analysis and storage.

[0055] Specifically, vibration signals were collected for four rotor states: normal, unbalanced, misaligned, and frictional. An unbalanced rotor was simulated by screwing a 2g mass block into a threaded hole on the edge of the mass disk. Misalignment was achieved by changing the relative position of the two shafts at the coupling, while contact friction was achieved by screwing a friction screw into its housing to contact the rotating shaft. During signal acquisition, the rotor speed was set to 1200 r / min, the sampling frequency to 2048 Hz, and the sampling length to 1 s. Five sets of tests were conducted under each rotor state (normal, contact friction, unbalanced, and misaligned), resulting in a total of 180 samples as the raw dataset. Each row of the raw dataset corresponds to a class label, representing a fault state, and each column contains the signal data collected under that fault state.

[0056] In step 2, features are extracted from the original dataset using sliding window, fast Fourier transform, and wavelet transform methods, respectively.

[0057] In this embodiment, to reduce noise and interference in the signal, the original signal is first divided into several segments using a sliding window. The signal length is 2400, the window length is set to 500, and the window interval is set to 150, resulting in 11 windowed signals. The average value of each window is taken as the time-domain feature of the signal, ultimately yielding 11-dimensional time-domain features. Because some signals are difficult to exhibit obvious features in the time domain but are more pronounced in the frequency domain, Fast Fourier Transform (FFT) and Wavelet Transform are used to perform frequency domain analysis on the original signal. FFT can extract the spectrum of a signal for spectral analysis. FFT is performed on the signal extracted by each sliding window, and the variance of the resulting spectrum is used as one of the frequency domain features, ultimately yielding 11-dimensional frequency domain features. Wavelet analysis is a time-frequency domain analysis method that considers information from both the time and frequency domains. When extracting features, dB5 (the fifth-level Daubechies wavelet) was used as the mother wavelet for noise removal. The original signal was decomposed into 6 layers of signals, resulting in a better filtering effect. The energy spectral density and average value of each layer of signal were taken as time-frequency domain feature values, and finally 12-dimensional time-frequency domain features were obtained.

[0058] The original signal set was processed using three extraction methods to obtain 35-dimensional features. In the final dataset, each column represents a fault feature, and each row represents a sample with a corresponding class label.

[0059] In step 3, the specific steps for processing the feature dataset using the feature selection evolutionary algorithm are as follows:

[0060] Step 3.1: Set the parameters of the feature selection evolutionary algorithm.

[0061] Set the parameters for the feature selection evolutionary algorithm: population size is denoted as N, and N is set to 100; maximum number of iterations is denoted as Max_iter, and Max_iter is set to 100; fault feature dimension is denoted as D, where D is the number of columns in the dataset.

[0062] Step 3.2: Calculate the feature ReliefF weight values ​​based on the set parameters. The more important the feature, the larger its feature weight.

[0063] Step 3.2.1: Normalize the weight values ​​of all features in the feature dataset.

[0064] Step 3.2.2: Sort the features in descending order based on their weight values.

[0065] Step 3.3: Use the inflection point selection strategy to select relevant features according to the feature ReliefF weight value, divide them into two groups: strongly relevant features and weakly relevant features, and construct the initial population.

[0066] Step 3.3.1: Arrange the features in descending order of their weights to obtain a curve about the feature weights.

[0067] Step 3.3.2: Connect the maximum weight point and the minimum weight point of the curve to obtain the limit line L.

[0068] Step 3.3.3: The point at which the distance from the limit line L to the curve is the inflection point.

[0069] Step 3.3.4: Assign features with weights greater than the inflection point to the strongly correlated feature subset, and assign the remaining features to the weakly correlated feature subset.

[0070] Step 3.3.5: Calculate the initial probability based on all relevant features.

[0071] The initial probability P is calculated according to formula (1):

[0072]

[0073] The sum of the weights of all strongly correlated features is denoted as W. s The sum of the weights of all weakly correlated features is denoted as W. p .

[0074] Step 3.3.6: Select feature values ​​according to the comparison results of relevant features and initialization probabilities to obtain the initial population.

[0075] In this embodiment, an N x D matrix is ​​randomly constructed, where each element is either 1 or 0, and each column represents a feature. For strongly correlated features, if the random value is less than P, the corresponding element is set to 1; otherwise, it is set to 0. For weakly correlated features, if the random value is greater than P, the corresponding element is set to 1; otherwise, it is set to 0. Each row of the matrix corresponds to an individual, and the element value in each row represents the position of that individual, thus yielding N individuals, forming the initial population.

[0076] Step 3.4: Optimize the initial population using the particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0077] Step 3.4.1: Divide the feature dataset into a training set and a test set. Specifically, divide the feature dataset into a training set (70%) and a test set (30%). The training set is used for model training, and the test set is used to derive the final accuracy value.

[0078] Step 3.4.2: Evaluate the particles and set optimization parameters.

[0079] In this embodiment, the classification error rate of each particle in the initial population is first obtained by using 5-fold cross-validation and K-nearest neighbor classification, where the distance value K in K-nearest neighbor classification is 5; then the selected feature ratio of each particle in the initial population is calculated, where the selected feature ratio of each particle is equal to the ratio of the number of elements with a value of 1 in that particle to D; the fitness function of each particle is constructed using the classification error rate and the selected feature ratio, as shown in formula (2):

[0080] Fitness=α*ErrorRate+(1-α)*NSF(2).

[0081] Where Fitness represents the particle's fitness value, α is the weight, which is greater than 0 and less than 1, set to 0.9, * is the multiplication operator, ErrorRate is the particle's classification error rate, and NSF is the ratio of selected features to the particle's weight.

[0082] The position corresponding to the particle with the smallest fitness value is set as the global optimal position Xgb, and the current position of each particle is set as its individual optimal position Xpb. The corresponding fitness values ​​are also set as the individual optimal pbest and the global optimal gbest.

[0083] In this embodiment, during the feature selection process for fault diagnosis, the fitness value of a particle is composed of the fault classification error rate and the fault feature selection ratio. The optimal position represents which fault features were selected from the particles with the smallest fitness value, i.e., a combination of 1s and 0s of length D.

[0084] Set the iteration variable t and initialize t to 1; initialize the number of times the population gets stuck in gbest (noimprove) to 0; initialize the number of times the population gets stuck in local optima (NI) to 0.

[0085] Step 3.4.3: Update the particle velocity and position based on feature correlation.

[0086] In this embodiment, the t-th iteration is performed, and the specific iteration process is as follows: Figure 1 As shown:

[0087] The inertia weight w of the particle swarm optimization algorithm is updated using a linear decreasing method:

[0088]

[0089] Here, iter and Max_iter represent the current iteration number and the maximum iteration number, respectively.

[0090] Update the velocity of each particle in the population: Update the velocity of the particles using formula (4);

[0091]

[0092] in, Let represent the current velocity of the i-th particle. This represents the velocity of the i-th particle at the previous moment; Let represent the optimal position of the i-th particle. This represents the global optimal position of the current population; c1 and c2 represent the acceleration coefficients, for example, set to 1.49445; r1 and r2 are both random numbers in the range [0,1]; d represents the d-th dimension vector of the particle.

[0093] Calculate the position update probability TF for each particle in the population:

[0094]

[0095] in, denoted as sigmond transfer function; λ is a weighted value in the range [0,1], set to 0.9; W d is the weight value of the d-th dimension feature; the value of the θ-th dimension in the range [0,1] is set to 0.5.

[0096] Update the position of each particle:

[0097]

[0098] in, This represents the position of the i-th particle. This indicates the position of the i-th particle at the next moment.

[0099] Step 3.4.4: Evaluate the updating particles by changing the size of the training set until the global optimal position is obtained as the final solution.

[0100] In this embodiment, it is checked whether the population has fallen into a local optimum. If NI is less than 1, it indicates that the population has not fallen into a local optimum, and the classification model is trained using a training set of 70% size. If NI is greater than 1, it indicates that the population has fallen into a local optimum. When the population falls into a local optimum again, the size of the training set is changed. If the training set size was 70% when the population last fell into a local optimum, then when it falls into a local optimum again, 10% of the samples are extracted from the 70% training set, and the classification model is trained using only 60% of the training set. If the training set size was 60% when the population last fell into a local optimum, then when it falls into a local optimum again, the classification model is trained using 70% of the training set.

[0101] In this embodiment, when the training set is reduced, a certain number of samples are extracted from each class in the training set according to a certain ratio; when the training set is expanded, the previous samples are put back in.

[0102] The fitness value of each particle is calculated using formula (7);

[0103] Fitness=α*ErrorRate+(1-α)*NSF (7).

[0104] Update the individual best position Xpb and the global best position Xgb of the population, and also update the corresponding fitness values ​​to individual best pbest and global best gbest.

[0105] Check if the gbest value in the t-th iteration is equal to the gbest value in the (t-1)-th iteration. If they are equal, increment the noimprove value by 1; otherwise, set the noimprove value to 0. When the noimprove value is greater than or equal to 8, it indicates that the population has fallen into a local optimum once, and increment the NI value by 1.

[0106] Determine if the current value of t is equal to 100. If not, increment the current value of t by 1 and update the value of t. Then return to steps 3-7 for the next iteration. If it is equal, the iteration terminates.

[0107] Output the global optimal position Xgb as the final solution; match the meaning of column D in the final solution with that of column D in the dataset, and the features corresponding to all columns of 1 in the final solution in the dataset constitute the key fault feature combination for fault diagnosis of the experimental platform.

[0108] In step 4, when a fault occurs again, the type of fault in the experimental platform is detected. Fault features are extracted from the signal using the feature extraction method, and key fault features of the experimental platform are selected. Then, based on the key fault feature data, the known K-nearest neighbor classification technique is used to obtain the classification label corresponding to the fault features from the public dataset corresponding to the experimental platform. In the K-nearest neighbor classification technique, the distance value K is 5. The classification label obtained at this time is the fault state of the experimental platform.

[0109] This embodiment discloses an intelligent fault diagnosis method, proposing a novel feature selection algorithm by utilizing the correlation between fault features and fault states. This effectively determines which fault features are most predictive of the target variable, thus aiding the algorithm in effective feature selection. The final output combination of key fault features can efficiently classify faults in mechanical systems. The advantages are not only high classification accuracy but also significantly reduced fault diagnosis costs.

[0110] To verify the performance of the intelligent fault diagnosis method of this invention, fault diagnosis was first performed on the WHU dataset using the method. The feature selection evolutionary algorithm parameters in this fault diagnosis method were set as follows: population size N was set to 100, and the number of iterations Max_iter was set to 100. In the table, CR(%) represents the classification accuracy of the feature subset. Furthermore, the effect was compared with that of using all fault features for identification and classification; the specific comparison data is shown in Table 1.

[0111] Table 1. Comparison of the effectiveness of the method of the present invention and the method using all fault features for identification and classification.

[0112]

[0113] The comparison results show that the fault diagnosis method proposed in this embodiment can achieve higher classification accuracy. This indicates that the fault diagnosis method of the present invention can successfully reduce the dimensionality of the fault diagnosis dataset, filter out a large number of redundant fault features that are not helpful for fault diagnosis, and can correctly classify faults by taking only a small portion of the fault features.

[0114] Example 2:

[0115] Embodiment 2 of the present invention provides an intelligent fault diagnosis system, comprising:

[0116] The data acquisition module is configured to acquire vibration signals under different rotor conditions, label the vibration signals, and form the original dataset.

[0117] The feature extraction module is configured to extract features from the original dataset to obtain a feature dataset.

[0118] The feature selection module is configured to process the feature dataset using a feature selection evolutionary algorithm, and then optimize the processed dataset using a particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis.

[0119] The fault diagnosis module is configured to use the K-nearest neighbor classification method to identify and classify the fault to be identified based on the combination of key features, and determine the fault category.

[0120] The steps and methods involved in the above embodiment two correspond to those in embodiment one. For specific implementation details, please refer to the relevant description section of embodiment one.

[0121] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0122] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An intelligent fault diagnosis method, characterized in that, Includes the following steps: Vibration signals under different rotor conditions were acquired, and the vibration signals were labeled to form the original dataset; Feature extraction is performed on the original dataset to obtain the feature dataset; The feature dataset is processed using a feature selection evolutionary algorithm. The specific steps are as follows: set the parameters of the feature selection evolutionary algorithm; calculate the feature ReliefF weight value according to the set parameters; and select relevant features according to the feature ReliefF weight value using the inflection point selection strategy to form the initial population. The specific steps for selecting relevant features based on their ReliefF weight values ​​to form the initial population using the inflection point selection strategy are as follows: Features are arranged in descending order of their weight values, resulting in a curve representing the feature weight values. Connecting the points with the maximum and minimum weight values ​​on the curve yields a limit line. The point with the largest distance from the limit line to the curve is the inflection point. Features with weight values ​​greater than the inflection point are assigned to a strongly correlated feature subset, while the remaining features are assigned to a weakly correlated feature subset. Initialization probabilities are calculated based on all relevant features. Feature values ​​are then selected based on the comparison between relevant features and initialization probabilities to obtain the initial population. Finally, the initial population is optimized using a particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis. The specific steps for optimizing the initial population using the particle swarm optimization algorithm are as follows: Divide the feature dataset into training and testing sets; evaluate the particles and set optimization parameters; update the particle velocity and position based on feature correlation; evaluate the updating particles by changing the size of the training set; check if the population has fallen into a local optimum; if so, change the size of the training set when the population falls into a local optimum again; when shrinking the training set, extract a certain number of samples from each class according to a ratio; when expanding the training set, reintroduce the previous samples until the global optimum is obtained as the final solution. The K-nearest neighbor classification method is used to identify and classify faults based on key feature combinations, thereby determining the fault category.

2. The intelligent fault diagnosis method as described in claim 1, characterized in that, The vibration signals include vibration signals for four rotor states: normal, unbalanced, misaligned, and frictional.

3. The intelligent fault diagnosis method as described in claim 1, characterized in that, Feature extraction was performed on the original dataset using sliding window, fast Fourier transform, and wavelet transform methods, respectively.

4. The intelligent fault diagnosis method as described in claim 1, characterized in that, The feature dataset includes 35-dimensional features, with each column representing a fault feature, each row representing a sample, and corresponding to a class label.

5. The intelligent fault diagnosis method as described in claim 1, characterized in that, The specific steps for calculating the feature ReliefF weight values ​​based on the set parameters are as follows: Normalize the weight values ​​of all features in the feature dataset; Sort in descending order based on the feature weight values.

6. An intelligent fault diagnosis system, characterized in that, The intelligent fault diagnosis method as described in any one of claims 1-5 includes: The data acquisition module is configured to acquire vibration signals under different rotor conditions, label the vibration signals, and form the original dataset; The feature extraction module is configured to extract features from the original dataset to obtain a feature dataset. The feature selection module is configured to process the feature dataset using a feature selection evolutionary algorithm, and then optimize the processed dataset using a particle swarm optimization algorithm to obtain the key feature combination for fault diagnosis. The fault diagnosis module is configured to use the K-nearest neighbor classification method to identify and classify the fault to be identified based on the combination of key features, and determine the fault category.

Images