Centrifugal fan fault diagnosis method based on multi-dimensional information fusion

Abstract

The invention discloses a centrifugal fan fault diagnosis method based on multi-dimensional information fusion. The centrifugal fan fault diagnosis method comprises the steps that vibration data of a horizontal measuring point and vibration data of a vertical measuring point of a centrifugal fan are collected respectively; inputting the vibration data of the plurality of horizontal measuring points and the vibration data of one or more vertical measuring points into a pre-established neural network model for diagnosis and prediction; and the neural network model obtains feature information of the diagnosis data of the horizontal measuring point and the diagnosis data of the vertical measuring point by adopting different fusion modes, and predicts a diagnosis result based on the feature information obtained by fusion. Meanwhile, a plurality of sensors of one device are utilized, different fusion modes are adopted according to different sensor installation positions when sensor information is fused, correlation and specificity between information are fully considered, and the accuracy of a prediction result is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent fault diagnosis, and in particular to a centrifugal fan fault diagnosis method based on multi-dimensional information fusion. Background Art

[0002] The prior art generally uses a single sensor for fault diagnosis and monitoring of centrifugal fans. For example, a vibration sensor is generally provided. The centrifugal fan is judged to be faulty based on the comparison between the vibration amplitude collected by the vibration sensor and the set threshold. This fault judgment method can only make a diagnosis for obvious faults, and the diagnostic accuracy is often defective. Fault diagnosis research based on a single signal has poor reliability and stability, and poor fault tolerance. Once the sensor fails, the diagnosis will also stop. When the fault types are numerous and complex, the diagnosis or prediction effect obtained by one or several pieces of information is unreliable, and different faults may also manifest in the form of the same sign. Therefore, the prior art is improved in the direction of multiple sensors.

[0003] Multi-sensor information fusion can provide richer information, complement each other in time and space, and derive more effective information through the optimal combination of information. When dealing with multi-sensor information fusion, most methods directly input this information into the constructed deep learning fault diagnosis model and use the self-learning ability of deep learning for fault diagnosis. This method ignores the differences and expressiveness between sensors, does not consider the spatial correlation between multi-sensor information, and does not fully utilize the relevant complementarity between information. There is a certain amount of information loss, which affects the diagnosis effect. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a centrifugal fan fault diagnosis method based on multi-dimensional information fusion. At the same time, multiple sensors of a device are utilized, and different fusion methods are adopted when fusing sensor information according to the different installation positions of the sensors, fully considering the correlation and specificity between the information, and improving the accuracy of the prediction results.

[0005] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a centrifugal fan fault diagnosis method based on multi-dimensional information fusion, comprising: respectively collecting the vibration data of the horizontal measuring points and the vibration data of the vertical measuring points of the centrifugal fan; sending the vibration data of multiple horizontal measuring points and the vibration data of one or more vertical measuring points as input into a pre-established neural network model for diagnosis and prediction; the neural network model uses different fusion methods to obtain the characteristic information of the diagnostic data of the horizontal measuring points and the diagnostic data of the vertical measuring points and predicts the diagnosis results based on the fused characteristic information.

[0006] The structure of the pre-established neural network model includes a channel feature fusion module, a multi-level feature fusion module, a conv Block module, and a classifier; wherein the channel feature fusion module performs channel feature fusion on the vibration data of multiple sensors at the horizontal measuring points of the input centrifugal fan to obtain feature information corresponding to the horizontal vibration; the multi-level feature fusion module performs multi-level feature fusion on the vibration data collected by multiple sensors at the vertical measuring points to obtain feature information corresponding to the vertical vibration; the feature information corresponding to the horizontal vibration and the feature information corresponding to the vertical vibration are both input into the conv Block module for feature extraction and the extracted features are sent to the classifier for classification, and the classifier outputs the fault diagnosis result.

[0007] The channel feature fusion module performs fusion processing on the vibration data collected by the sensors at the horizontal measuring points, where the number of vibration sensors corresponding to the horizontal measuring points is three, and the vibration data at the three horizontal measuring points are tested respectively.

[0008] The data of three sensors are taken as a group, and the length of the vibration signal of each group of sensors is cut. The vibration sample data of the cut length corresponding to each sensor in each group is transformed into a grayscale image, and each vibration sample data generates a corresponding grayscale image sample; the grayscale images of the three sensors in each group are fused into a three-channel RGB image by using channel fusion.

[0009] Each sensor signal in each group is truncated to a length of 4096 to obtain n samples, each of which includes three sensor sample data. Each sensor sample data is transformed into a grayscale image, and each sensor sample data generates a grayscale image sample of size (1, 64, 64); the grayscale images corresponding to the three sensor sample data in each sample data are fused to form a three-channel RGB image sample.

[0010] The multi-level feature fusion module includes a first-level feature fusion branch, a second-level feature fusion branch, and a third-level feature fusion branch. The input data of each feature fusion branch are branch sub-features X2, X3, and X4 respectively. The output results of each feature fusion branch are multiplied to obtain multi-level features; among them, the first-level feature fusion branch, the second-level feature fusion branch, and the third-level feature fusion branch all include conv units, BN units, and ReLU units.

[0011] The vibration data collected by the sensor at the vertical measuring point is obtained by time series statistics, which is the input feature X1, X∈R L1x1xC , where the feature size of X1 is L1, and C represents the number of channels; the input feature X1 is split into three branch sub-features X2, X3, and X4; the size of each sub-feature is L2, L3, and L4, respectively, and L1 = L2+L3+L4.

[0012] When splitting the input feature X1, the settings of L2, L3, and L4 meet the following requirements: L2≠L3≠L4 and L2 <L3<L4,L2=2 n-3 ,L3=2 n-2 ,L4=2 n-1 , where n is a positive integer.

[0013] The pre-established neural network model is trained before fault diagnosis and prediction. A data set is established with the three sensor data of the horizontal measuring points and the sensor data of the vertical measuring points as input and the fault status of the centrifugal fan as output. The data set is divided into a training set and a test set. The established neural network model is trained and tested until the training end conditions are met, and the trained neural network model is obtained and used for automatic diagnosis of centrifugal fans.

[0014] The advantages of the present invention are: using multiple sensors of a device at the same time, taking different fusion methods according to the different sensor installation positions when fusing sensor information, and fully considering the correlation and specificity between information. Fully exploring the features in each segment of sensor signal, adopting multi-level feature information fusion, improving the expression ability of a small number of sample signals, and making the fault diagnosis prediction of centrifugal fans more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The following is a brief description of the contents expressed in the drawings of the present invention and the symbols in the drawings:

[0016] Figure 1 It is a structural diagram of the neural network model of the present invention;

[0017] Figure 2 This is a measurement point diagram of the centrifugal fan equipment of the present invention;

[0018] Figure 3 It is a schematic diagram comparing the prediction accuracy of the model of the present invention and the prediction accuracy of the prior art model.

[0019] The reference numerals in the figure are: 1, air outlet side; 2, motor side; 3, coupling. DETAILED DESCRIPTION

[0020] The specific implementation of the present invention will be further explained in detail below by describing the optimal embodiment with reference to the accompanying drawings.

[0021] The present application provides a centrifugal fan fault diagnosis method based on multi-dimensional information fusion, including: respectively collecting vibration data of horizontal measuring points and vibration data of vertical measuring points of the centrifugal fan; sending the vibration data of multiple horizontal measuring points and the vibration data of one or more vertical measuring points as input into a pre-established neural network model for diagnosis and prediction; the neural network model uses different fusion methods to obtain feature information of the diagnostic data of the horizontal measuring points and the diagnostic data of the vertical measuring points, and predicts the diagnostic results based on the fused feature information.

[0022] like Figure 2 As shown, the centrifugal fan consists of a motor side, a coupling and an air outlet side, wherein the motor side is the motor and its housing; the air outlet side has a spiral air inlet and outlet structure, including a housing, a shaft, an impeller, etc., which is mainly used for air outlet; the motor on the motor side is connected to the shaft on the air outlet side through a coupling, thereby realizing the drive control of the centrifugal fan. The measuring points set on the centrifugal fan include: three horizontal measuring points and one vertical measuring point are set on the centrifugal fan, wherein the horizontal measuring points are respectively set at the free end of the motor, the load end of the motor, and the side of the coupling; the vertical measuring points include the load end of the motor. Vibration sensors are respectively set at these measuring points to collect vibration signals of the centrifugal fan. These vibration signals are respectively sent to the pre-established neural network model for fault diagnosis and prediction, and the corresponding fault diagnosis results are output.

[0023] The neural network model consists of two parts: the establishment of the neural network model and the training and verification of the neural network model.

[0024] The establishment of the neural network model includes building a network model structure and defining its principles, specifically: the structure of the pre-established neural network model includes a channel feature fusion module, a multi-level feature fusion module, a conv Block module, and a classifier; the channel feature fusion module performs channel feature fusion on the vibration data of multiple sensors at the horizontal measuring points of the input centrifugal fan to obtain feature information corresponding to the horizontal vibration; the multi-level feature fusion module performs multi-level feature fusion on the vibration data collected by multiple sensors at the vertical measuring points to obtain feature information corresponding to the vertical vibration; the feature information corresponding to the horizontal vibration and the feature information corresponding to the vertical vibration are both input into the conv Block module for feature extraction and the extracted features are sent to the classifier for classification, and the classifier outputs the fault diagnosis results.

[0025] The channel feature fusion module performs fusion processing on the vibration data collected by the sensors at the horizontal measuring points, where the number of vibration sensors corresponding to the horizontal measuring points is three, and the vibration data at the three horizontal measuring points are tested respectively.

[0026] The data of three sensors are taken as a group, and the length of the vibration signal of each group of sensors is intercepted. The vibration sample data of the intercepted length corresponding to each sensor in each group is converted into a grayscale image, and each vibration sample data generates a corresponding grayscale image sample; the grayscale images of the three sensors in each group are fused into a three-channel RGB image by channel fusion. The vibration signal of the sensor is a series of vibration waveforms in time sequence, and the waveform amplitude is its vibration amplitude signal; each sensor corresponds to a waveform in time sequence, and the three sensors in each group of sensors correspond to three vibration data waveforms. Since the vibration data of each sensor is too long, it needs to be intercepted according to the size. After interception, n samples are obtained, and each sample number corresponds to the sub-sample data of three sensors. Each sub-sample data is converted into a grayscale image, so each sample data corresponds to three grayscale images, and the three grayscale images are fused to obtain an RGB channel color image as a feature input. That is, each sensor signal in each group is truncated with a length of 4096 to obtain n samples, each of which includes three sensor sample data. Each sensor sample data is transformed into a grayscale image, and each sensor sample data generates a grayscale image sample of size (1, 64, 64); the grayscale images corresponding to the three sensor sample data in each sample data are fused to form a three-channel RGB image sample.

[0027] The multi-level feature fusion module includes a first-level feature fusion branch, a second-level feature fusion branch, and a third-level feature fusion branch. The input data of each feature fusion branch are branch sub-features X2, X3, and X4 respectively. The output results of each feature fusion branch are multiplied to obtain multi-level features; among them, the first-level feature fusion branch, the second-level feature fusion branch, and the third-level feature fusion branch all include conv units, BN units, and ReLU units.

[0028] The vibration data collected by the sensor at the vertical measuring point is obtained by time series statistics, which is the input feature X1, X∈R L1x1xC , where the feature size of X1 is L1, and C represents the number of channels; the input feature X1 is split into three branch sub-features X2, X3, and X4; the size of each sub-feature is L2, L3, and L4, respectively, and L1 = L2+L3+L4.

[0029] When splitting the input feature X1, the settings of L2, L3, and L4 meet the following requirements: L2≠L3≠L4 and L2 <L3<L4,L2=2 n-3 ,L3=2 n-2 ,L4=2 n-1 , where n is a positive integer.

[0030] After the structure of the neural network model is built, it is necessary to train and verify the neural network model. After the training is completed and the verification meets the requirements, it can be applied to the actual engineering environment to perform fault diagnosis and prediction on the centrifugal fan working in real time. The training includes:

[0031] The pre-established neural network model is trained before fault diagnosis and prediction. A data set is established with the three sensor data of the horizontal measuring points and the sensor data of the vertical measuring points as input and the fault status of the centrifugal fan as output. The data set is divided into a training set and a test set. The established neural network model is trained and tested until the training end conditions are met, and the trained neural network model is obtained and used for automatic diagnosis of centrifugal fans.

[0032] Regarding the channel fusion module and multi-level feature fusion of the neural network model, which belong to the core of the network model of this application, this application introduces it in detail as follows:

[0033] 1. Channel feature fusion

[0034] 1. Collect multiple measurement data (i.e., multi-sensor signals) from the device. Specifically, collect data from the horizontal (H) measurement points in the device, and group three sensors together to facilitate later information fusion. The sampling frequency is 51.2kHZ, and the original signal is expressed as:

[0035] X = {x1:x2:x3}; each group of signals includes three sensor signals, and each sensor signal includes a vibration signal collected and arranged in time sequence; that is, each sensor signal is an array or a one-dimensional vector composed of a group of numbers; each group of signals includes three sensor signals;

[0036] 2. Each group of signals is cut into 4096 lengths to generate n samples. Each sample is converted into a grayscale image. Each sample will generate a grayscale image sample of (1, 64, 64).

[0037]

[0038] 3. The channel fusion method fuses the grayscale images of the three sensors into a three-channel RGB image sample, and fuses the sensor signals in the three horizontal directions to generate an image signal with rich signals, which is specifically expressed as (3, 64, 64).

[0039] 2. Multi-level feature fusion

[0040] 1. For sensor signals in the vertical (V) direction, a multi-level feature fusion method is used to extract more sufficient information from the data of a small number of sensors.

[0041] 2. Assume that the input feature is X1(X1∈R L1x1xC) Then the feature size of X1 is L1, where C represents the number of channels. X1 needs to be split into three branch sub-features X2, X3, and X4, with their sizes being L2, L3, and L4 respectively, and L1 = L2 + L3 + L4. It should be noted that the convolution kernel sizes of different convolutional layers are different, and these convolutional layers need to be assigned different features, so L2 ≠ L3 ≠ L4 and L2 < L3 < L4. X1 is a vector composed of data from vibration sensors at vertical measurement points arranged in time series.

[0042] 3. During the process of feature recognition in a deep neural network, the size of the feature is reduced layer by layer, and the reduction multiple is usually a power of 2. That is, L1 = L2 + L3 + L4 = 2 n (n ∈ N * , N * represents the set of positive integers) Since when the feature size is 2 n , the neural network can more easily recognize the feature, so the optimal effect is achieved when the values of L2, L3, and L4 are powers of 2. However, the function y = 2^x is a concave function in the coordinate system, and the sum of three different powers of 2 will not be a power of 2, that is, 2 a +2 b +2 c ≠2 n , a, b, c ∈ N * and a ≠ b ≠ c. Therefore, only two of the values of L2, L3, and L4 can be made to satisfy powers of 2. In addition, according to the characteristics of the exponential function, too large a difference in size between different features will result in one feature being too large or too small, which is not conducive to multi-scale feature learning. Therefore, the difference between L2, L3, and L4 should not be too large. It should be noted that 2 n satisfies the equation 2 n = 2 n-1 +2 n-2 +2 n-3 (n > 3)

[0043] According to the above size range and (3), the following method for setting L2, L3, and L4 is deduced:

[0044] L2 = 2 n-3 , L3 = 2 n-2 , L4 = 2 n-1

[0045] A multi-branch module for feature segmentation is established, and the feature will be divided into three branches, and feature extraction and fusion are performed through convolution kernels of different sizes to fully exert the sample feature expression.

[0046] In order to verify the accuracy of the prediction results of the model in this application, the research object is the centrifugal fan in the cement production process. According to the centrifugal fan failure cases and data provided by the diagnostic personnel, the failure caused by this equipment is an unbalanced failure. The failure data is the data of 4 measuring points and the vibration data of operation from March to May.

[0047]

[0048] The training process is also a fast convergence process, and the accuracy reaches 99.9%, which meets the diagnostic requirements. The test set that does not participate in the training is sent to the trained model for verification. Comparing several traditional model networks, it can be seen that the accuracy of the network model that extracts features by designing different feature fusion methods according to the installation positions of different sensors has certain advantages and meets the diagnostic requirements. At the same time, it can also be seen that the accuracy of the model in this scheme for centrifugal fan fault diagnosis after training is completed is more accurate than the existing technology.

[0049] Obviously, the specific implementation of the present invention is not limited to the above-mentioned methods. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, they are all within the protection scope of the present invention.

Claims

1. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion, characterized in that: include: The vibration data of the horizontal measuring points and the vibration data of the vertical measuring points of the centrifugal fan are collected respectively; the vibration data of multiple horizontal measuring points and the vibration data of one or more vertical measuring points are sent as input to a pre-established neural network model for diagnosis and prediction; the neural network model uses different fusion methods to obtain the characteristic information of the diagnostic data of the horizontal measuring points and the diagnostic data of the vertical measuring points, and predicts the diagnostic results based on the fused characteristic information.

2. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 1, characterized in that: The structure of the pre-established neural network model includes a channel feature fusion module, a multi-level feature fusion module, a conv Block module, and a classifier; The channel feature fusion module performs channel feature fusion on the vibration data of multiple sensors at the horizontal measuring points of the centrifugal fan to obtain the feature information corresponding to the horizontal vibration; the multi-level feature fusion module performs multi-level feature fusion on the vibration data collected by multiple sensors at the vertical measuring points to obtain the feature information corresponding to the vertical vibration; the feature information corresponding to the horizontal vibration and the feature information corresponding to the vertical vibration are both input into the conv Block module for feature extraction and the extracted features are sent to the classifier for classification, and the classifier outputs the fault diagnosis results.

3. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 2, characterized in that: The channel feature fusion module performs fusion processing on the vibration data collected by the sensors at the horizontal measuring points, where the number of vibration sensors corresponding to the horizontal measuring points is three, and the vibration data at the three horizontal measuring points are tested respectively.

4. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 3, characterized in that: The data of three sensors are taken as a group, and the length of the vibration signal of each group of sensors is cut. The vibration sample data of the cut length corresponding to each sensor in each group is transformed into a grayscale image, and each vibration sample data generates a corresponding grayscale image sample; the grayscale images of the three sensors in each group are fused into a three-channel RGB image by using channel fusion.

5. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 4, characterized in that: Each sensor signal in each group is intercepted with a length of 4096 to obtain n samples, each of which includes three sensor sample data. Each sensor sample data is transformed into a grayscale image, and each sensor sample data generates a grayscale image sample of size (1, 64, 64); The grayscale images corresponding to the three sensor sample data in each sample data are fused to form a three-channel RGB image sample.

6. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in any one of claims 1 to 5, characterized in that: The multi-level feature fusion module includes a first-level feature fusion branch, a second-level feature fusion branch, and a third-level feature fusion branch. The input data of each feature fusion branch are branch sub-features X2, X3, and X4, respectively. The output results of each feature fusion branch are multiplied to obtain a multi-level feature. The first-level feature fusion branch, the second-level feature fusion branch, and the third-level feature fusion branch all include conv units, BN units, and ReLU units.

7. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 6, characterized in that: The vibration data collected by the sensor at the vertical measuring point is obtained by time series statistics, which is the input feature X1, X∈R L1x1xC , where the feature size of X1 is L1, and C represents the number of channels; the input feature X1 is split into three branch sub-features X2, X3, and X4; the size of each sub-feature is L2, L3, and L4, respectively, and L1 = L2+L3+L4.

8. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion as claimed in claim 7, characterized in that: When splitting the input feature X1, the settings of L2, L3, and L4 meet the following requirements: L2≠L3≠L4 and L2 <L3<L4,L2=2 n-3 ,L3=2 n-2 ,L4=2 n-1 , where n is a positive integer.

9. A centrifugal fan fault diagnosis method based on multi-dimensional information fusion according to any one of claims 1 to 8, characterized in that: The pre-established neural network model is trained before fault diagnosis and prediction. A data set is established with the three sensor data of the horizontal measuring points and the sensor data of the vertical measuring points as input and the fault status of the centrifugal fan as output. The data set is divided into a training set and a test set. The established neural network model is trained and tested until the training end conditions are met, and the trained neural network model is obtained and used for automatic diagnosis of centrifugal fans.

Images