Bucket elevator fault diagnosis system and diagnosis method

Abstract

The invention discloses a bucket elevator fault diagnosis system and method, and the method comprises the steps: firstly transmitting data collected by a vibration sensor and a temperature measurement camera of an information sensing layer to a data processing layer, and carrying out the cleaning of the original data collected by the sensor and the camera through the data processing layer, and forming an image data set; and training an IDRN-BiGRU network fused and integrated by a BiGRU and an IDRN based on an improved residual neural network by using the image data set to obtain a fault diagnosis model and realize automatic diagnosis and classification of faults. Compared with simple time domain or frequency domain conversion, fault information is more sufficient; and a deep learning algorithm is adopted, and fault identification and classification are performed on different fault maps, so that compared with traditional mechanism analysis, the method is more efficient, and intelligent monitoring can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment intelligent operation and maintenance, and in particular to a bucket elevator fault diagnosis system and method. Background Art

[0002] Bucket elevators are widely used in foundries. Since the equipment operates in a dusty environment all year round and some are located underground, inspections are difficult. Traditional inspection methods also make it difficult to detect potential or early failures. Unplanned downtime caused by equipment failures can cause serious economic losses to the factory. Traditional mechanical fault diagnosis methods require professionals to analyze the health of the equipment based on the monitored vibration and temperature changes, which requires high professional skills of practitioners.

[0003] The Chinese invention patent with the authorization announcement number CN113432877B discloses a complex rotating mechanical equipment fault diagnosis method based on visual feature fusion. The method adopted by the patent is a complex rotating mechanical equipment fault diagnosis method based on visual feature fusion, which solves the problems of incomplete fault diagnosis sample data, difficulty in fault feature characterization and extraction, and poor fault diagnosis ability in the existing diagnosis method. In terms of fault feature proposal, the patent obtains a one-dimensional spectrum image after analyzing and processing the original vibration signal, and the simple spectrum image lacks time domain information, which will lose some time series features, resulting in incomplete fault feature information extraction. That is, the existing fault diagnosis method has the following problems: the traditional method performs mechanism analysis on the original vibration signal of the machine, which has high professional requirements for equipment operation and maintenance personnel, and the signal analysis calculation is complex and inefficient; the one-dimensional spectrum image used in the existing public technology lacks time series information, and the fault feature information extraction is not complete; the fault image that only uses vibration is relatively simple in fault feature extraction, and does not fully consider factors such as time series and temperature changes. Summary of the invention

[0004] In order to overcome the above problems existing in the prior art, the present invention proposes a bucket elevator fault diagnosis system and monitoring method.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a bucket elevator fault diagnosis method, comprising: Step 1, collecting the vibration signal of the bearing through the vibration sensor, and collecting the heating state image of the head wheel and the tail wheel of the hoist through the temperature measurement camera; Step 2, denoising, filtering, and normalizing the vibration signal collected in step 1; Step 3, converting the vibration signal obtained in step 2 into a time-frequency domain, converting the one-dimensional vibration signal into a two-dimensional time-frequency domain spectrum; Step 4, the two-dimensional time-frequency domain spectrum obtained in step 3 and the fever state image collected in step 1 together form an image data set; Step 5, constructing a bidirectional gated recurrent unit BiGRU and an IDRN-BiGRU network based on an improved residual neural network IDRN fusion integration, and training the IDRN-BiGRU network with the image data set obtained in step 4 to obtain a fault diagnosis model; Step 6, input the vibration signal to be diagnosed after being processed in steps 2 and 3 into the fault diagnosis model obtained in step 5 to obtain the fault type.

[0006] In the above-mentioned bucket elevator fault diagnosis method, the time-frequency domain conversion in step 3 uses a time sliding analysis window to truncate the non-stationary signal, the original signal is decomposed into each short-time signal, and the spectrum of each short-time signal is subjected to Fourier transform.

[0007] In the above-mentioned bucket elevator fault diagnosis method, in step 5, the IDRN-BiGRU network first uses the residual neural network IDRN to avoid network degradation and gradient vanishing problems, and then uses the advantages of the bidirectional gated recurrent unit BiGRU in extracting temporal features, and the fully connected layer and the Flatten layer in the traditional CNN network architecture are replaced by a global pooling layer.

[0008] In the above-mentioned bucket elevator fault diagnosis method, the residual neural network adds a Dropout layer to the residual block; the bidirectional gated recurrent unit is composed of an update gate and reset gate constitute, control Enter degree, The larger the value, the The more information there is; control Enter degree, The smaller the value, The less information an entry has.

[0009] The above-mentioned method for diagnosing faults of a bucket elevator, and The calculation formula is: ; ; ℎ t ̃ =tanh[ W ℎt cat( r t ℎ t−1 , x t )] ; ; ; ; y t =[ ℎ t ⃗ , ℎ t ⃖ ] ; In the formula is the input value at the current moment, , , , represented as a weight matrix, cat() indicates that the eigenvectors are connected; Represents sigmoid; represents tensor product; Indicates the status information of the previous moment; Indicates the current instant status information; Represents the output value of the current GRU unit, through the update gate Implemented the forgetting and remembering steps; Indicates forward State and backward The states are connected together and output through a bidirectional hidden layer, which extracts more feature information.

[0010] A bucket elevator fault diagnosis system is based on the above-mentioned bucket elevator fault diagnosis method, including an information perception layer, a data processing layer, and a data application layer. The information perception layer is provided with a vibration sensor and a temperature measurement camera. The information perception layer is used to collect data information and transmit the data information to the data processing layer; the data processing layer cleans the data information to form an image data set; the data application layer includes a server and a fault diagnosis model, and the fault diagnosis model is trained through the image data set.

[0011] In the above-mentioned bucket elevator fault diagnosis system, the vibration sensor realizes the collection of vibration data of the monitoring point, and the temperature measurement camera collects the heating state images of the elevator head wheel and tail wheel motors.

[0012] The beneficial effects of the present invention are as follows: (1) by converting the vibration signal of the mechanical equipment into time-frequency domain, a spectrum integrating the time domain and the frequency domain can be obtained, which provides more complete fault information than the conversion of the time domain or the frequency domain alone; (2) The deep learning algorithm is used to identify and classify different fault maps. Compared with traditional mechanism analysis, it is more efficient and can realize intelligent monitoring.

[0013] (3) With the accumulation of long-term data and iteration of algorithm models, predictive maintenance of equipment can be achieved, which can avoid unplanned equipment downtime, ensure equipment safety, and reduce economic losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a fault diagnosis flow chart of the present invention; Figure 2 It is the normal vibration signal analysis spectrum, where (a) is the time domain diagram, (b) is the frequency domain diagram, (c) is the time-frequency domain diagram, and (d) is the top view of the time-frequency domain analysis; Figure 3 1 is a vibration time-frequency domain image under normal and fault conditions in an embodiment of the present invention, wherein (a) is a normal spectrum, (b) is a fault spectrum of the outer ring of a motor bearing, (c) is a fault spectrum of the roller of a motor bearing, and (d) is a fault spectrum of the inner ring of a motor bearing; Figure 4 It is a residual block structure diagram of the prior art; Figure 5 : is a residual block structure diagram of the present invention; Figure 6 It is the BiGRU structure diagram of the present invention; Figure 7 It is the IDRN-BiGRU network structure diagram of the present invention; Figure 8 Schematic diagram of the monitoring system of the present invention. DETAILED DESCRIPTION

[0015] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0016] This embodiment discloses a method for diagnosing a bucket elevator fault. The fault diagnosis flow chart is as follows: Figure 1 As shown, including: Step 1: collect the vibration signal of the bearing through the vibration sensor, and collect the heating state image of the head wheel and tail wheel of the hoist through the temperature measurement camera.

[0017] (1) Selection of vibration signal acquisition equipment High-precision vibration velocity sensors are used to ensure accurate capture of bearing vibration signals under different working conditions. The sensors have good sensitivity and frequency response characteristics and are suitable for various complex working conditions.

[0018] (2) Determine the sensor installation location Install the vibration velocity sensor on the rotating part of the machine, such as the bearing seat. Use a magnetic base or other fixing method to ensure that the sensor is in close contact with the measured surface to avoid looseness or displacement that affects the measurement accuracy.

[0019] (3) Data collection system construction Construct a data acquisition system, including sensors, high-speed data acquisition cards and computers. The vibration signal is collected in real time through the data acquisition card and transmitted to the computer for storage and processing to ensure the integrity and real-time nature of the data.

[0020] Step 2, denoising, filtering and normalizing the vibration signal collected in step 1.

[0021] (1) Denoising: The vibration signal is denoised using a digital high-pass filter to remove high-frequency noise and interference in the signal.

[0022] The formula for a first-order digital high-pass filter is: ; Where s is a complex frequency domain variable and ωc is the cutoff frequency.

[0023] This transfer function represents a single-pole system with a stopband that slopes down at -20dB / decade. The cutoff frequency is given by: ; Where, fc is the cutoff frequency and fs is the sampling frequency.

[0024] The above transfer function is converted into the digital domain through bilinear transformation to obtain the difference equation: ; Where α and β are filter coefficients, calculated as follows: ; Where Ts is the sampling period.

[0025] The design of a first-order digital high-pass filter can be achieved by selecting a suitable cutoff frequency and sampling frequency, and then calculating the filter coefficient according to the above formula to ultimately achieve high-pass filtering of the signal.

[0026] According to the characteristics of mechanical vibration fault signal and noise characteristics, select appropriate filter parameters. For example, for low-frequency interference, a high-pass filter can be used to suppress it.

[0027] (2) Filtering: After removing the noise, the vibration signal is further filtered to smooth the signal and highlight the fault characteristics.

[0028] Median filtering is a nonlinear filtering method that eliminates spikes and noise by taking the median of the current sampling point and its adjacent sampling points as output. The formula is: ; Where k is half of the window size and n is the number of filter points. For sudden signal or spike pulse, median filtering can effectively remove these interferences while retaining the useful features of the signal.

[0029] (3) Normalization processing: The filtered vibration signal is normalized to adjust the amplitude range of the signal to a uniform standard range. Common normalization methods include maximum and minimum value normalization.

[0030] Min-Max Scaling is a commonly used data preprocessing method used to convert data of different scales and dimensions into the same range. This method maps the minimum value of the data to 0 and the maximum value to 1 (or a user-specified range) through linear transformation. The following are the implementation steps: a. Calculate the minimum and maximum values: Find the minimum and maximum values ​​in the data set.

[0031] b. Apply normalization formula: Normalize each data point using the above formula.

[0032] Assume data set , its minimum value is min(X) and its maximum value is max(X). The formula for normalizing the maximum and minimum values ​​is: ; Among them, xi′ is the normalized data point.

[0033] Data offset: Normalize the data to other ranges (such as [0,1]), which can be adjusted based on the above formula. For example, to normalize to the range of [a,b], use the following formula: .

[0034] Step 3, converting the vibration signal obtained in step 2 into time-frequency domain, and converting the one-dimensional vibration signal into a two-dimensional time-frequency domain spectrum.

[0035] Because the focus of time domain analysis is the analysis of the signal in time and signal amplitude, it is difficult to distinguish whether it is a fault by the change of vibration amplitude under variable industrial control, while frequency domain analysis can reveal the characteristics of the signal in the frequency domain at a deeper level, but lacks information in the time dimension. Simply relying on the power spectrum expressed in the frequency domain, it is impossible to see the relationship between a certain frequency component and time, nor the change of the component over time, and it is impossible to take into account the localized nature of the signal in both the time and frequency domains. The time-frequency diagram has the advantages of time domain and frequency domain analysis, and it is a very effective method to extract bearing fault characteristics in a subdivided time unit.

[0036] The specific method is to set the number of sampling points or sampling time, and model the frequency, amplitude and other information in the unit in the three-coordinate space. By setting the observation angle, you can see three-dimensional images at different angles. In order to obtain better image effects, this paper selects the top view of the time-frequency domain image. The top view can more completely reflect the fault characteristics, such as Figure 2 The time domain, frequency domain, time-frequency domain, and time-frequency domain top view under normal signals are shown. The time-frequency domain top view image contains the relevant physical characteristics of the motor bearing in both the time domain and frequency domain.

[0037] Linear time-frequency analysis and nonlinear time-frequency analysis are two commonly used methods for time-frequency domain analysis. The most important method in linear time-frequency analysis is short-time Fourier transform. This method uses a time sliding analysis window to truncate non-stationary signals. The original signal is decomposed into short-time signals, and then the spectrum of each short-time signal is transformed using Fourier transform.

[0038] Let x(t) be the signal to be analyzed and the analysis window be g(t). Then the short-time Fourier transform of x(t) is defined as: ; The above formula is a continuous short-time Fourier transform (STFT), however, in practical applications, most signals are discrete signals. The short-time Fourier transform of a discrete signal x(n) is called a discrete short-time Fourier transform, and is defined as follows: ; Where is the real window function. So the discrete STFT can be regarded as the Fourier transform of the sequence x(m)y(nm).

[0039] like Figure 3 As shown in the figure, the normal signal and the bearing fault signal are processed by short-time Fourier transform to obtain different spectra. It can be seen from the figure that the spectrum under fault is obviously different from that under normal condition.

[0040] In step 4, the two-dimensional time-frequency domain spectrum obtained in step 3 and the fever state image collected in step 1 together constitute an image data set.

[0041] Step 5, construct a bidirectional gated recurrent unit BiGRU and an IDRN-BiGRU network based on the fusion integration of the improved residual neural network IDRN. The network structure is as follows: Figure 7 As shown, the IDRN-BiGRU network is trained using the image dataset obtained in step 4 to obtain a fault diagnosis model.

[0042] IDRN-BiGRU network. It is proposed to combine the advantages of IDRN and BiGRU. First, the residual neural network IDRN is used to avoid network degradation and gradient vanishing problems. Then, the bidirectional gated recurrent unit BiGRU is used to extract the advantages of timing features. In order to retain the timing features of the signal to the greatest extent, the fully connected layer and the Flatten layer in the traditional CNN network architecture are replaced by the global pooling layer. This achieves the purpose of shortening the feature extraction time, focusing on timing issues, and improving the accuracy of fault diagnosis.

[0043] This embodiment proposes an improved deep residual network (IDRN) to improve the feature extraction capability and convergence speed of the deep residual network. The Dropout layer is added to the residual block to improve the diagnostic efficiency and effectively alleviate overfitting, and to achieve end-to-end fault diagnosis. Finally, it is compared with some commonly used network models. The results show that the method proposed in this article is effective and feasible.

[0044] In order to avoid the overfitting of the network, the structure of the residual block is improved. Figure 4 The figure shows the standard residual block structure before improvement. Figure 5 Shown is the improved residual block structure.

[0045] The structure of the bidirectional gated recurrent unit BiGRU is as follows Figure 6 As shown, the bidirectional gated recurrent unit consists of an update gate and reset gate constitute, control Enter degree, The larger the value, the The more information there is; control Enter degree, The smaller the value, The less information an entry has.

[0046] and The calculation formula is: ; ; ℎ t ̃ =tanh[ W ℎt cat( r t ℎ t−1 , x t )] ; ; ; ; y t =[ ℎ t ⃗ , ℎ t ⃖ ] ; In the formula is the input value at the current moment, , , , represented as a weight matrix, cat() indicates that the eigenvectors are connected; Represents sigmoid; represents tensor product; Indicates the status information of the previous moment; Indicates the current instant status information; Represents the output value of the current GRU unit, through the update gate Implemented the forgetting and remembering steps; Indicates forward State and backward The states are connected together and output through a bidirectional hidden layer, which extracts more feature information.

[0047] In this embodiment, the specific process of training the IDRN-BiGRU network is as follows: first, the vibration signal is converted into the time-frequency domain to form a spectrum, and the infrared temperature measurement camera is used to obtain the infrared imaging of the motor, and a data set is constructed. According to the analysis of the motor vibration signal, there is a large amount of noise and redundancy in the signal. To solve this problem, it is very necessary to perform one-hot encoding on the label data in the data processing stage. The pre-processed data is segmented into 1,000 samples of label data, which are distributed as follows: 700 training sets, 200 validation sets, and 100 test sets. In the model training stage, the network parameters need to be initialized first, and then the network model is pre-trained using the training set. After the iterative training is completed, the validation set data is input into the model to verify the generalization ability, and the optimization parameters are adjusted at the same time. In the forward propagation process, the data is first subjected to multi-layer feature extraction through the improved residual block, and then the extracted features are linearized and input into the two-layer BiGRU network as time series features. The time series information in the feature data is fully mined and the classification results are output. In the back-propagation process of model verification, the output and label types are compared, the loss is calculated, and the back-propagation algorithm is used to update and adjust the weight value of each layer until the iteration of the training set and the verification set data is completed, and then the network model is tested for fault diagnosis using the test set.

[0048] Step 6, input the vibration signal to be diagnosed after being processed in steps 2 and 3 into the fault diagnosis model obtained in step 5 to obtain the fault type.

[0049] Based on the above detection method, this embodiment also discloses a bucket elevator fault diagnosis system, such as Figure 8 As shown, it includes an information perception layer, a data processing layer, and a data application layer. The information perception layer is provided with a vibration sensor and a temperature measurement camera. The information perception layer is used to collect data information and transmit the data information to the data processing layer; the data processing layer cleans the data information to form an image data set; the data application layer includes a server and a fault diagnosis model, and the fault diagnosis model is trained through the image data set.

[0050] The vibration sensor collects vibration data of the monitoring point, and the temperature measurement camera collects heating state images of the head wheel and tail wheel motors of the hoist.

[0051] Because the working environment of the hoist is dusty, the temperature measuring camera lens is easily blocked by dust. In this embodiment, an automatic dust scraper is also designed. A brush for periodically scraping dust is added in front of the temperature measuring camera lens. The system can set the interval time for automatic dust scraping.

[0052] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A bucket elevator fault diagnosis method, characterized in that: include: Step 1, collecting the vibration signal of the bearing through the vibration sensor, and collecting the heating state image of the head wheel and tail wheel motor of the lifting machine through the temperature measurement camera; Step 2, denoising, filtering, and normalizing the vibration signal collected in step 1; Step 3, converting the vibration signal obtained in step 2 into a time-frequency domain, converting the one-dimensional vibration signal into a two-dimensional time-frequency domain spectrum; Step 4, the two-dimensional time-frequency domain spectrum obtained in step 3 and the fever state image collected in step 1 together form an image data set; Step 5, constructing a bidirectional gated recurrent unit BiGRU and an IDRN-BiGRU network based on an improved residual neural network IDRN fusion integration, and training the IDRN-BiGRU network with the image data set obtained in step 4 to obtain a fault diagnosis model; Step 6, input the vibration signal to be diagnosed after being processed in steps 2 and 3 into the fault diagnosis model obtained in step 5 to obtain the fault type.

2. A bucket elevator fault diagnosis method according to claim 1, characterized in that: In the step 3, the time-frequency domain conversion uses a time sliding analysis window to truncate the non-stationary signal, the original signal is decomposed into each short-time signal, and the spectrum of each short-time signal is subjected to Fourier transform.

3. A bucket elevator fault diagnosis method according to claim 1, characterized in that: In step 5, the IDRN-BiGRU network first uses the residual neural network IDRN to avoid network degradation and gradient vanishing problems, and then uses the advantages of the bidirectional gated recurrent unit BiGRU in extracting temporal features. The fully connected layer and the Flatten layer in the traditional CNN network architecture are replaced by a global pooling layer.

4. A bucket elevator fault diagnosis method according to claim 3, characterized in that: The residual neural network adds a Dropout layer to the residual block; the bidirectional gated recurrent unit is composed of an update gate and reset gate constitute, control Enter degree, The larger the value, the The more information there is; control Enter degree, The smaller the value, The less information an entry has.

5. A bucket elevator fault diagnosis method according to claim 4, characterized in that: Said and The calculation formula is: ; ; ; ; ; ; ; In the formula is the input value at the current moment; , , Represented as a weight matrix, cat() indicates that the eigenvectors are connected; Represents sigmoid; represents tensor product; Indicates the status information of the previous moment; Indicates the current instant status information; Represents the output value of the current GRU unit, through the update gate Implemented the forgetting and remembering steps; Indicates forward State and backward The states are connected together through a bidirectional hidden layer output.

6. A bucket elevator fault diagnosis system, characterized in that: A bucket elevator fault diagnosis method based on any one of claims 1 to 5, comprising an information perception layer, a data processing layer, and a data application layer, wherein the information perception layer is provided with a vibration sensor and a temperature measurement camera, and the information perception layer is used to collect data information and transmit the data information to the data processing layer; the data processing layer cleans the data information to form an image data set; the data application layer includes a server and a fault diagnosis model, and the fault diagnosis model is trained using the image data set.

7. A bucket elevator fault diagnosis system according to claim 6, characterized in that: The vibration sensor collects vibration data of the monitoring point, and the temperature measurement camera collects heating state images of the head wheel and the tail wheel of the hoist.

Images