A k-nearest neighbor centrifugal pump fault diagnosis method based on improved relief F weighting

Abstract

The application discloses a kind of k near neighbour centrifugal pump fault diagnosis methods based on improved ReliefF weighting provided by the patent, the vibration signal of centrifugal pump under different operating conditions is collected and is sliced after it, then the characteristic parameters in time domain and frequency domain are calculated, and the feature selection is carried out using extreme forest model, the weight of characteristic parameter is calculated using improved ReliefF algorithm, and then the fault type of the sample to be measured is predicted using k near neighbour algorithm weighted by Gaussian distance.The k near neighbour centrifugal pump fault diagnosis method based on improved ReliefF weighting proposed in the application can well realize the various fault diagnosis and classification of centrifugal pump and the system composed of it, improve the operation efficiency of algorithm, and ensure the reliability and stability of diagnostic result.

Description

Technical Field

[0001] This invention relates to the field of centrifugal pump condition monitoring and fault diagnosis, specifically to a fault diagnosis method for centrifugal pumps based on an improved ReliefF weighted method. Background Technology

[0002] Centrifugal pumps have been widely used in industries such as petrochemicals, metallurgy, machinery, and power. Due to the complex structure of centrifugal pumps, the relatively harsh working environment, and the influence of various random factors, the working system composed of motors, gearboxes, centrifugal pumps, etc. is prone to various failures. When the failure is serious, it will lead to the paralysis of the production system and endanger life and property safety. Therefore, it is very important to carry out fault diagnosis research on centrifugal pumps and the systems they are composed of.

[0003] Currently, in the field of centrifugal pump fault mode identification and diagnosis, the following methods are commonly used: common centrifugal pump fault diagnosis methods include mechanism diagnosis methods (such as noise detection diagnosis method, vibration detection method, etc.) and artificial intelligence diagnosis methods (such as neural network model, Naive Bayes model, etc.).

[0004] The k-nearest neighbor algorithm is also a widely used machine learning classification algorithm. Its basic idea is to find the k nearest neighbor samples of the sample to be classified in the feature space, and determine the fault type of the sample to be classified based on the fault category of most of the nearest neighbor samples.

[0005] The traditional k-nearest neighbor algorithm for pump fault diagnosis mainly includes: ① Feature signal: extracting time-frequency domain feature signals based on the vibration data of the pump's operation obtained by measurement; ② Distance calculation: calculating the distance from each text to be classified to all known samples, and obtaining the k nearest neighbor samples; ③ Fault discrimination: determining the fault category of the sample to be classified based on the principle of majority rule for the k nearest neighbor samples.

[0006] When the k-nearest neighbor algorithm is used for pump and motor fault diagnosis, it often faces the problem of large sample feature space dimensionality and large computational cost. However, not all features are actually useful for fault classification. At the same time, when the sample size of different labels is unbalanced, the majority of the k nearest neighbor samples are large samples that are far away, which affects the classification effect. In addition, the traditional k-nearest neighbor algorithm does not consider the different contributions of each dimension of features to fault classification, and there is still room for further improvement in the classification accuracy of the algorithm. Summary of the Invention

[0007] To overcome the shortcomings of the above technologies, this invention provides a fault diagnosis method for k-nearest neighbor centrifugal pumps based on an improved ReliefF weighted method.

[0008] The technical solution adopted by this invention to overcome its technical problems is as follows: This application proposes a fault diagnosis method for centrifugal pumps based on an improved ReliefF weighted k-nearest neighbor algorithm, applied to rotating machinery systems including at least centrifugal pumps. The method comprises: S1, acquiring vibration signals of the centrifugal pumps under different operating conditions of the rotating machinery system, wherein the operating conditions include normal operating conditions and several fault operating conditions; S2, slicing the acquired vibration signals, grouping them, and assigning fault category labels; S3, calculating time-domain and frequency-domain feature parameters from the acquired vibration signals, and obtaining a centrifugal pump vibration signal feature matrix based on the time-domain and frequency-domain feature parameters; S4, performing feature selection on the time-domain and frequency-domain feature parameters based on an extreme forest model, and dividing the test samples and training samples based on the feature matrix composed of the dimensionality-reduced feature parameters; S5, calculating the feature weights corresponding to the dimensionality-reduced feature parameters based on the improved ReliefF algorithm; and S6, classifying the test samples based on the feature weights and the k-nearest neighbor algorithm to obtain the fault type of the test samples.

[0009] Furthermore, the process of slicing and grouping the acquired vibration signals specifically includes: based on the centrifugal pump rotation frequency F... r and sampling frequency F s The collected vibration signal is sliced ​​to obtain several sets of data, where every n sampling points constitute one set of data, and n ≥ F. s / F r .

[0010] Furthermore, the centrifugal pump vibration signal feature matrix is ​​an a×b matrix, where a is the number of samples in the vibration signal data collected under each working state, and b is the total number of features of the time-domain feature parameters and frequency-domain feature parameters.

[0011] Furthermore, in the extreme forest model, the number of trees n_estimatorss is less than or equal to the total number of features b, the maximum number of features used by a single decision tree max_features is the square root of the total number of features b, and the evaluation metric is entropy.

[0012] Furthermore, step S3 specifically includes: calculating and ranking the Gini importance levels of the feature parameters based on the extreme forest model; reducing the dimensionality of the feature parameters by retaining several feature parameters with high importance levels; forming a feature matrix based on the dimensionality-reduced feature parameters; and dividing the feature matrix into training samples and test samples.

[0013] Furthermore, the calculation of feature weights corresponding to the dimensionality-reduced feature parameters based on the improved ReilefF algorithm specifically includes: S51, initializing the feature weight vector w as a zero vector; S52, randomly selecting a sample R from the training samples, and finding k nearest neighbor samples H from the sample set of the same type as sample R.j Find k nearest neighbor samples M from each class that is different from R. j Thus, based on formula (1), the feature weight w(A) of each feature parameter in sample R is obtained:

[0014]

[0015] In the formula, M j (C) represents the sample. The j-th nearest neighbor sample in the set, m represents the number of samplings, p represents the proportion of this class of samples in the set, and the function diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A. Its calculation formula is shown in (2).

[0016]

[0017] S53, after m samplings, calculate the average value of the feature weight vector w, reset the feature weights less than 0 in the mean of the feature weight vector to 0, and normalize the feature weight vector w.

[0018] Furthermore, the classification of the test samples based on feature weights and the k-nearest neighbor algorithm to obtain the fault type of the test samples specifically includes:

[0019] S61, calculate the Euclidean distance between each test sample and the training sample based on formula (3).

[0020]

[0021] Where R and N represent training samples and test samples, respectively, w is the feature weight vector, and x is the number of feature parameters after dimensionality reduction;

[0022] S62, sort the Euclidean distances and select the k sample points with the smallest distance to the sample points to be tested;

[0023] S63, Gaussian distance weighting of the k nearest neighbor samples is given by formula (4):

[0024]

[0025] In the formula, c = 10.

[0026] S64, determine the fault category of the sample under test, and the determination function g is formula (5).

[0027]

[0028] In the formula, k_label is the fault category label of the nearest neighbor sample, and L is the value of formula (6).

[0029]

[0030] S65. After obtaining the decision function value g of the sample to be tested, the fault category label class corresponding to the largest g value is the fault category of the sample to be tested, and the fault type of the sample to be tested is output.

[0031] Furthermore, it also includes diagnosing rotating machinery systems, specifically including: S71, acquiring vibration signals of centrifugal pumps collected during the fault state of rotating machinery, and grouping them into p groups of data with n sampling points as a set; S72, calculating the feature parameters of the p groups of data after dimensionality reduction, and using them as test samples; S73, calculating the Euclidean distance between the test samples and training samples, sorting the Euclidean distances, and selecting the k sample points with the smallest distance to the test sample points; S74, applying Gaussian distance weighting to the k sample points; S75, calculating the judgment function g of the fault sample; S76, the class label corresponding to the maximum value of the judgment function g is the fault category of the test sample.

[0032] The beneficial effects of this invention are:

[0033] 1. The embedded feature selection method of extreme forest is adopted to select features first. This process can effectively reduce the sample dimensionality and improve the speed and accuracy of model prediction.

[0034] 2. The improved ReliefF method is used to calculate the feature weights, which ensures the stability and rationality of the feature weight calculation results.

[0035] 3. The feature weights were normalized, which reduced the impact of large fluctuations in feature weights on classification performance and further improved the accuracy of model prediction.

[0036] 4. By using a Gaussian function to optimize the weights of samples at different distances, the error in fault classification caused by sample imbalance can be effectively overcome, thereby further improving the accuracy of model prediction. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the fault diagnosis method for k-nearest neighbor centrifugal pumps based on the improved ReliefF weighted method according to an embodiment of the present invention.

[0038] Figure 2 The data provided represent the fault diagnosis rate of the k-nearest neighbor centrifugal pump fault diagnosis method based on the improved ReliefF weighted method, as described in this embodiment of the invention. Detailed Implementation

[0039] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following are merely exemplary and do not limit the scope of protection of the present invention.

[0040] like Figure 1 The flowchart of a centrifugal pump fault diagnosis method based on improved ReliefF weighted k-nearest neighbor algorithm described in this embodiment includes: S1, acquiring vibration signals of the centrifugal pump under different operating conditions of the rotating machinery system, wherein the operating conditions include at least normal operating conditions and several fault operating conditions; S2, slicing the acquired vibration signals, grouping them, and assigning labels; S3, calculating time-domain feature parameters and frequency-domain feature parameters from the acquired vibration signals, and obtaining a centrifugal pump vibration signal feature matrix based on the time-domain feature parameters and frequency-domain feature parameters; S4, performing feature selection on the time-domain feature parameters and frequency-domain feature parameters based on an extreme forest model, and dividing the test samples and training samples based on the feature matrix composed of the dimensionality-reduced feature parameters; S5, calculating the feature weights corresponding to the feature parameters based on the improved ReliefF algorithm; S6, classifying the test samples based on the feature weights and the k-nearest neighbor algorithm to obtain the fault type of the test sample.

[0041] S1 obtains centrifugal pump vibration signals under different operating conditions of the rotating machinery system through an acceleration sensor of the rotating machinery system. The operating conditions of the rotating machinery include normal operating conditions and fault operating conditions. Fault operating conditions include the following types of faults: bearing faults, specifically including faults in the inner ring, outer ring, and rolling elements of the bearing, or bearing lubrication failure; rotor faults, including shaft bending, shaft cracking, rotor imbalance, and rotor misalignment; pump faults, including blade breakage; gear faults, including gear pitting, wear, and breakage; mechanical loosening; and mechanical abrasion.

[0042] In one embodiment of the present invention, the sampling frequency F of the accelerometer is... s The frequency F of the centrifugal pump used in the experiment was 51.2 kHz. r It is 25Hz.

[0043] S2, the collected vibration signal is sliced ​​into length segments, with each n sampling points forming a data set, where n ≥ F. s / F r .

[0044] In one embodiment of the present invention, 4096 sampling points are used as a set of data, and vibration signals under different working conditions are assigned category labels.

[0045] S3, calculate the time-domain characteristic parameters and frequency-domain characteristic parameters of the collected vibration signal, and obtain the centrifugal pump vibration signal characteristic matrix based on the time-domain characteristic parameters and frequency-domain characteristic parameters;

[0046] In one embodiment of the present invention, the time-domain parameters include a total of 36 parameters such as average value, standard deviation, root square amplitude, root mean square value, peak-to-peak value, skewness, and kurtosis, and the frequency-domain feature parameters include a total of 14 parameters. The time-domain feature parameters and the frequency-domain feature parameters are combined to form a centrifugal pump vibration signal feature matrix (a×b), where a is the number of groups of data for each working state, and b is the sum of the number of time-domain feature parameters and the number of frequency-domain feature parameters, i.e., the total number of features.

[0047] S4 uses an extreme forest model to select features from time-domain and frequency-domain features, and then uses the feature matrix composed of the reduced-dimensional features to divide the test samples and training samples.

[0048] In one embodiment of the present invention, the specific steps include:

[0049] S41, import the Extreme Forest model ExtraTreesClassifier function from Python's sklearn.ensemble module, the code is: from sklearn.ensemble importExtraTreesClassifier;

[0050] S42, the parameters selected in the extreme forest model are: the number of trees in the forest n_estimators, the evaluation metric is entropy, and the maximum number of features used in a single decision tree max_features. The larger n_estimators is, the better (n_estimators ≤ the total number of features b). The value of max_features is the square root of the total number of features b.

[0051] S43, calculate and sort the Gini importance levels of the feature parameters, retain the top x features with higher importance for subsequent feature inputs, and divide the training samples and test samples.

[0052] In one embodiment of the present invention, the number of trees n_estimators in the extreme forest model is 50, the maximum number of features used in a single decision tree max_features is 7, the features are sorted according to the Gini importance level, the 25 most important features are retained, and the feature matrix composed of the feature parameters after dimensionality reduction is divided into training samples and test samples, wherein the training samples account for 70%.

[0053] S5, calculate the weights w of different feature parameters according to the improved ReliefF algorithm.

[0054] In one embodiment of the present invention, when calculating the weight w, samples should be drawn evenly from all types of fault samples, with a sampling number of 5. After sampling, the average value of the feature weight vector is calculated, and the feature weights less than 0 are reset to 0. Then, the feature weights are normalized. Specifically, this includes:

[0055] S51, Initialize the feature weight vector w as a zero vector;

[0056] S52, randomly select a sample R from the training samples, and find k nearest neighbor samples H from the set of samples of the same type as sample R. j Find k nearest neighbor samples M from each class that is different from R. j Thus, based on formula (1), the feature weight w(A) of each feature parameter in sample R is obtained:

[0057]

[0058] In the formula, M j (C) represents the sample. The j-th nearest neighbor sample in the set, m represents the number of samplings, p represents the proportion of this class of samples in the set, and the function diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A. Its calculation formula is shown in (2).

[0059]

[0060] S53, after m samplings, calculate the average of the feature weight vector w, and reset the feature weights less than 0 in the mean of the feature weight vector to 0 during the calculation process. Also, normalize the feature weight vector w. Each sampling will yield a weight vector w, and the average of w is calculated after m samplings.

[0061] S6 uses the feature weight vector w as the feature weight for calculating the Euclidean distance using the k-nearest neighbor algorithm. The Euclidean distances are sorted, and the k nearest neighbor samples are selected. The category with the most frequent occurrences among the k nearest neighbors of the test sample is the fault type of the test sample, and the diagnostic result is output. Specifically, this includes:

[0062] S61, calculate the Euclidean distance between each test sample and the training sample based on formula (3).

[0063]

[0064] Where R and N represent training samples and test samples, respectively, w is the feature weight vector, and x is the number of feature parameters after dimensionality reduction;

[0065] S62, sort the Euclidean distances and select the k sample points with the smallest distance to the sample points to be tested;

[0066] S63, Gaussian distance weighting is performed on the k nearest neighbor samples according to formula (4):

[0067]

[0068] In the formula, c = 10, which is an empirical value, namely the standard deviation of the Gaussian function, representing the width of the Gaussian window.

[0069] S64, determine the fault category of the sample under test, and the determination function g is formula (5).

[0070]

[0071] In the formula, k_label is the fault category label of the nearest neighbor sample, and L is the value of formula (6).

[0072]

[0073] S65. After obtaining the decision function value g of the sample to be tested, the fault category label class corresponding to the largest g value is the fault category of the sample to be tested, and the fault type of the sample to be tested is output.

[0074] The reliability of the fault diagnosis method based on the improved ReliefF-weighted k-nearest neighbor centrifugal pump was verified by comparing the labels of the fault types predicted by the improved ReliefF-weighted k-nearest neighbor centrifugal pump fault diagnosis method with the previously assigned fault category labels.

[0075] The k-nearest neighbor algorithm based on the improved ReilefF weighted algorithm is optimized, including the selection of the number of samples m when calculating the weights in step S5 and the selection of the k value when calculating the sample distance in step S6. The selection of the k value depends on experience. A larger k value can reduce the impact of noise on the samples, but it will increase the ambiguity of the boundaries between categories; a smaller k value will affect the classification accuracy and amplify the interference of noisy data.

[0076] In this embodiment, the minimum number of samples for different fault types is 425, and the k-value is selected as 100 after model optimization. The fault diagnosis rate is as follows: Figure 2 As shown, the algorithm provided by this invention significantly improves the accuracy of fault diagnosis and the stability of multiple diagnosis results.

[0077] In one embodiment of the present invention, S7 is further included to diagnose the fault of the rotating machinery. If the rotating machinery system malfunctions during actual operation, the specific steps include the following steps to diagnose the fault.

[0078] S71, acquire the vibration signal of the centrifugal pump collected during operation, and divide it into p groups of data with n sampling points as a group.

[0079] S72, and calculate the feature parameters of p grouped data after dimensionality reduction, and form a feature matrix pxq as the fault sample, where q is the number of feature parameters after dimensionality reduction, which is known in step S43 of the above embodiment.

[0080] S73, calculate the Euclidean distance between p test samples and training samples, sort the Euclidean distances, and select the k sample points with the smallest distance to the test sample points, where k has been obtained empirically in step S6 and has been verified.

[0081] S74, Gaussian distance weighting is performed on k sample points according to formula (4).

[0082] S75, the judgment function g of the sample to be tested is calculated according to formulas (5) and (6).

[0083] S76. Based on the judgment function g of the sample to be tested, the class label corresponding to the largest g value is the fault category of the sample to be tested, and the fault type of the sample to be tested is output.

[0084] Existing k-nearest neighbor (kNN) methods for fault diagnosis do not consider the effects of different feature weights and uneven sample distribution. This invention obtains vibration signals from centrifugal pumps under different operating conditions, merges time-domain and frequency-domain feature parameters into a multi-dimensional feature parameter matrix, performs feature selection using the extreme forest method, calculates the weights of the feature parameters using an improved ReliefF algorithm, and finally classifies the test samples using an improved ReliefF-weighted kNN algorithm to obtain the fault type of the test samples. The method provided by this invention can effectively achieve fault diagnosis and classification of centrifugal pumps and their constituent systems, improves the algorithm's operating efficiency, and ensures the reliability and stability of the diagnostic results.

[0085] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0086] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A k-nearest neighbor centrifugal pump fault diagnosis method based on improved ReliefF weighting, applied to a rotating mechanical system comprising at least a centrifugal pump, characterized in that, The method comprises the following steps: S1, acquiring vibration signals of a centrifugal pump in different working states of a rotating mechanical system, wherein the working states include normal working state and several fault working states; S2, grouping and assigning fault category labels to the collected vibration signals after slicing processing; S3, calculating the collected vibration signals to obtain time domain characteristic parameters and frequency domain characteristic parameters, and obtaining a centrifugal pump vibration signal feature matrix based on the time domain characteristic parameters and the frequency domain characteristic parameters; S4, performing feature selection on the time domain characteristic parameters and the frequency domain characteristic parameters based on an extreme forest model, and dividing test samples and training samples based on a feature matrix composed of the reduced characteristic parameters; S5, calculating feature weights corresponding to the reduced characteristic parameters based on an improved ReilefF; S6, classifying the test samples based on the feature weights and a k-nearest neighbor algorithm, so as to obtain the fault type of the test samples. The collected vibration signals are grouped after slicing processing, and specifically comprising: based on the centrifugal pump rotation frequency F r and the sampling frequency F s The collected vibration signals are sliced to obtain a plurality of groups of data, wherein every n sampling points are taken as a group of data, wherein n≥F s / F r ; The centrifugal pump vibration signal feature matrix is an a×b matrix, a is the number of groupings of data of each working state, and b is the total number of features of the time domain characteristic parameters and the frequency domain characteristic parameters; In the extreme forest model, the number of trees n_estimators is less than or equal to the total number of features b, the maximum number of features used by a single decision tree max_features is the square root of the total number of features b, and the evaluation index is entropy entropy; Step S4 specifically comprises: calculating the Gini importance level of the characteristic parameters based on the extreme forest model and sorting; A number of characteristic parameters with high importance level are retained to reduce the characteristic parameters; A feature matrix is composed based on the reduced characteristic parameters; And the feature matrix is divided into training samples and test samples; The feature weights corresponding to the reduced characteristic parameters are calculated based on the improved ReilefF, specifically comprising: S51, initializing the feature weight vector w to a zero vector; S52, randomly select a sample R from the training samples, find k nearest neighbor samples H from the sample set of the same class as sample R j find k nearest neighbor samples M from each sample set of different classes as sample R j Thus, the feature weight w(A) of each feature parameter in sample R is obtained based on formula (1): （1） wherein M j (C) represents the sample C the jth nearest neighbor of class (R), m represents the number of samples, p represents the proportion of samples of this class in the sample set, and the function diff(A, R1, R2) represents the difference between samples R1 and R2 in feature A, the calculation formula of which is shown in (2). （2） S53, calculating the average value of the feature weight vector w after m times of sampling, setting the feature weights less than 0 in the average value of the feature weight vector to 0 during the calculation process, and normalizing the feature weight vector w.

2. The k-NN centrifugal pump fault diagnosis method based on improved ReliefF weighting according to claim 1, characterized in that, The test samples are classified based on the feature weights and the k-nearest neighbor algorithm, so as to obtain the fault type of the test samples, specifically comprising: S61, calculating the Euclidean distance between each test sample and the training sample based on formula (3), （3） Wherein, R and N respectively represent the training sample and the test sample, w is the feature weight vector, and x is the number of reduced characteristic parameters; S62, sorting the Euclidean distance and selecting the k nearest sample points with the smallest distance from the test sample point; S63, weighting the k nearest neighbor samples by Gaussian distance as formula (4): （4） Wherein, c=10, S64, determining the fault category of the test sample, and the determination function g is formula (5) （5） Wherein, k_label is the fault category label of the nearest neighbor sample, and L is formula (6) （6） S65, after obtaining the determination function value g of the test sample, the fault category label class corresponding to the maximum g value is the fault category of the test sample, and the fault type of the test sample is output.

3. The k-NN centrifugal pump fault diagnosis method based on improved ReliefF weighting according to any one of claims 1-2, characterized in that, The method further comprises diagnosing the rotating mechanical system, specifically comprising: S71, obtain the vibration signal of the centrifugal pump collected by the rotating machinery, and group the n sampling points as one group of data into p grouped data; S72, calculate the feature parameters of the p grouped data after dimension reduction, and take them as test samples; S73, calculate the Euclidean distance between the test sample and the training sample, sort the Euclidean distance, and select the k sample points with the smallest distance from the test sample point; S74, Gaussian distance weighting is performed on the k sample points; S75, calculate the decision function g of the test sample; S76, based on the maximum value of the decision function g and the corresponding class label class, the fault category of the test sample is obtained.

Images