A method for diagnosing bearing corrosion based on regional convolutional networks

Abstract

This invention proposes a method for diagnosing bearing corrosion based on a regional convolutional network. First, stray magnetic flux signals are acquired using a coil flux sensor. These signals are then processed and decomposed into axial and radial components. Power spectra are extracted from the axial and radial components to generate a continuous spectrum image. This continuous spectrum image is then processed, and a cropped and adjusted image is used as input to identify potential regions for diagnosing bearing corrosion using a selective search algorithm. The identified potential regions are then input into a CNN neural network, which extracts shallow features for diagnosing bearing corrosion. These shallow features are used to classify the degree of bearing corrosion, resulting in the final diagnostic result.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault diagnosis technology, and in particular relates to a method for diagnosing bearing corrosion based on regional convolutional networks. Background Technology

[0002] With the advancement of science and technology, industrial equipment has become more complex and precise. Since industrial equipment plays a crucial role in actual production, any malfunction can severely impact the industrial production process. Furthermore, statistics show that most motor failures are caused by bearing electrolytic corrosion. Therefore, research on the degree of corrosion in motor bearings is of great significance.

[0003] Previous research on bearing fault diagnosis has mainly focused on the acquisition and feature extraction of vibration or current signals. Currently, many studies have systematically summarized recent research on using vibration or current signals for rolling bearing fault diagnosis. However, on the one hand, adding vibration sensors significantly increases actual production costs. On the other hand, especially under certain special operating conditions, some harmonics in the current spectrum can be amplified. For these reasons, it is necessary to select signals less affected by operating conditions, such as methods for detecting bearing galvanic corrosion using stray magnetic flux signals from motors. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for diagnosing bearing corrosion based on regional convolutional networks, which can solve the problems of low diagnostic accuracy caused by the installation limitations and high cost of vibration sensors, as well as harmonic amplification of current signals under special working conditions.

[0005] Technical Solution: The objective of this invention can be achieved through the following technical solution: A method for diagnosing bearing corrosion based on region convolutional networks, comprising the following steps:

[0006] Step 1: Acquire stray magnetic flux signals using a coil flux sensor;

[0007] Step 2: Extract the power spectrum of the acquired stray magnetic flux signal to obtain the continuous spectrum image, which is the binary histogram accumulated in the power frequency space.

[0008] Step 3: The continuous spectrum image from Step 2 is cropped and adjusted, and used as input to a region convolutional neural network to diagnose the degree of corrosion of the motor bearing.

[0009] Preferably, in step 2, the power spectrum of the acquired stray magnetic flux signal is extracted, and the power spectrum extraction expression is:

[0010] X(f)=[X1(f)X2(f)…X n (f)]

[0011]

[0012] In the formula: X m (f) represents the Discrete Fourier Transform (DTF) of the window data centered at time mR; x(n) is the input signal at time n, g(n) is the window function, and R is the number of samples between subsequent DFTs.

[0013] Preferably, in step 2, the acquired stray magnetic flux signal is processed and decomposed into axial and radial components, and the power spectrum of the axial and radial components is extracted and a continuous spectrum image is plotted.

[0014] The continuous spectrum image is generated in two dimensions with frequency as the horizontal axis and power spectrum as the vertical axis. The continuous spectrum image is used to characterize the correlation between bearing corrosion state and magnetic flux signal changes.

[0015] Preferably, in step 3, the process of diagnosing the corrosion degree of the motor bearing using a region convolutional network includes:

[0016] Step 3-1: Process the continuous spectrum image. Using the cropped and adjusted continuous spectrum image as input, use a selective search algorithm to identify potential areas for bearing corrosion diagnosis.

[0017] Step 3-2: Input the potential regions identified in Step 3-1 into the CNN neural network, use the CNN neural network to extract shallow representations for bearing corrosion degree diagnosis, use the shallow features to classify the bearing corrosion degree, and obtain the diagnosis results;

[0018] Preferably, the sustained spectrum image described in step 3-1 represents the relationship between frequency and power, and for each frequency band of the power frequency pair, it represents the percentage of time it exists in the signal evolution. The redder the color in the sustained spectrum image, the longer the specific power frequency band persists in the signal.

[0019] Preferably, in step 3-1, the selective search algorithm includes: investigating all edge regions attributable to different degrees of corrosion and having peak amplitudes, and classifying them as anchor points or bounding boxes. A regression task (using non-maximum suppression) is performed on the input continuous spectrum image to generate a set of potential regions using the bounding boxes;

[0020] Preferably, in step 3-2, the use of a CNN neural network to extract shallow representations for judging the degree of bearing corrosion, and the use of shallow features to classify the degree of bearing corrosion, includes inputting a 32×32 region proposal into a primary two-dimensional convolutional block to extract a fixed-length feature vector. These high-level features are input into two fully connected layers and then into a softmax layer used as a classifier.

[0021] Preferably, the softmax layer used as a classifier is a core component of the CNN classification task, achieving efficient multi-classification decision-making through probabilistic output and cross-entropy loss optimization.

[0022] Preferably, in the softmax layer, the process of probabilistic output includes:

[0023] The raw scores output by the last layer (fully connected layer) of the neural network are transformed into a probability distribution such that the probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1.

[0024]

[0025] In the formula, For each score z i The purpose of indexation is to amplify the impact of high scores, where K is the number of categories;

[0026] Preferably, in the softmax layer, the cross-entropy loss is optimized to measure the probability distribution predicted by the model, p = [p1, p2, ..., p...]. K The true label distribution y = [y1, y2, ..., y] is the same as the true label distribution. K Differences between:

[0027] Loss for a single sample:

[0028]

[0029] In the formula, K is the number of categories, p i Let y be the probability distribution predicted by the model. i This represents the true label distribution;

[0030] Loss on batch samples:

[0031]

[0032] In the formula, K is the number of categories, N is the number of samples, and p i Let y be the probability distribution predicted by the model. i This represents the true label distribution;

[0033] Beneficial effects:

[0034] Compared with the prior art, the present invention has at least the following technical effects:

[0035] 1. This invention diagnoses the degree of corrosion of motor bearings by using stray magnetic flux signals from the motor. The stray magnetic flux signals are not easily affected by special working conditions, avoiding the problems of installation limitations and high costs of vibration sensors, as well as the low diagnostic accuracy caused by harmonic amplification of current signals under special working conditions.

[0036] 2. To address the timing feature identification problem of stray magnetic flux signals, the applied regional convolutional network can first filter the signals using a selective search algorithm, and then use a simple convolutional neural network to classify the degree of bearing electrical corrosion, which greatly reduces the network size. Attached Figure Description

[0037] Figure 1 This is an overall framework diagram of the present invention.

[0038] Figure 2 This is a flowchart for diagnosing the degree of corrosion according to the present invention.

[0039] Figure 3 This is a continuous spectrum image used for fault diagnosis. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0041] like Figure 1 The diagram shown illustrates the overall framework of a bearing corrosion diagnosis method based on a region convolutional network provided by this invention. First, stray magnetic flux signals are acquired using a coil flux sensor. These signals are then processed and decomposed into axial and radial components. Power spectra are extracted from the axial and radial components to generate a continuous spectrum image. The continuous spectrum image is then processed, and a cropped and adjusted continuous spectrum (PS) image is used as input. A selective search algorithm is employed to identify potential regions for assessing bearing corrosion levels. These identified potential regions are then input into a CNN neural network, which extracts shallow features for assessing bearing corrosion levels. These shallow features are used to classify the bearing corrosion levels, resulting in the final diagnostic result. The specific implementation steps are as follows:

[0042] Step 1: Acquire stray magnetic flux signals using a coil flux sensor;

[0043] The stray magnetic flux signal collected by the coil flux sensor during motor operation is collected. The coil flux sensor has 1000 turns, an inner diameter of 65mm, and a wire diameter of 0.2mm.

[0044] The initial signal acquired by this method is obtained by a coil flux sensor, which needs to be placed on the frame of the motor under test.

[0045] Step 2: Extract the power spectrum of the acquired stray magnetic flux signal to obtain the continuous spectrum image, which is the binary histogram accumulated in the power frequency space.

[0046] Due to the specific characteristics of the acquired signal, the stray magnetic flux signal needs to be decomposed into axial and radial components. Power spectrum extraction is then performed on the axial and radial components, and a continuous spectrum image is plotted as the input signal for the corrosion degree diagnostic model. The power spectrum extraction expression for the acquired stray magnetic flux signal is as follows:

[0047] X(f)=[X1(f)X2(f)…X n (f)]

[0048]

[0049] In the formula: X m (f) represents the Discrete Fourier Transform (DTF) of the window data centered at time mR; x(n) is the input signal at time n, g(n) is the window function, and R is the number of samples between subsequent DFTs.

[0050] Step 3: The continuous spectrum image from Step 2 is cropped and adjusted, and used as input to a region convolutional neural network to diagnose the degree of corrosion of the motor bearing.

[0051] Corrosion degree diagnosis process: such as Figure 2 The specific process for corrosion degree diagnosis is illustrated below: The power spectrum of the acquired stray magnetic flux signal is extracted to obtain a persistent spectrum image. The image is then preprocessed and fed into a region convolutional neural network (CNN). This network first uses a selective search algorithm to identify potential regions. Subsequently, the image data is input into a CNN model based on these potential regions. This model further analyzes the image and extracts key features. The model calculates the classification probability using a softmax function and continuously updates its parameters during training until convergence. Finally, the model provides a diagnosis based on the classification results, completing the image analysis and recognition task and obtaining the final diagnostic result.

[0052] Specifically: Region neural networks include selective search algorithms and CNN neural networks based on latent regions. The steps of the selective search algorithm are explained below.

[0053] Step 1. Perform preliminary spectral analysis and segmentation on the characteristics of candidate corrosion levels;

[0054] Step 2. Investigate all edge regions attributable to different degrees of corrosion and with peak amplitude, and classify them as anchor points or bounding boxes;

[0055] Step 3. Perform a regression task (using non-maximum suppression) on the input continuous spectrum image to generate a set of potential regions using bounding boxes;

[0056] Step 4. Check the termination condition, i.e., determine whether the number of regions detected up to the current iteration has reached the upper limit. If not, calculate an additional potential region and insert it into the hierarchical queue;

[0057] Step 5. Input the identified potential regions into the CNN neural network.

[0058] The CNN neural network takes 32×32 region proposals as input to a primary 2D convolutional block to extract fixed-length feature vectors. A batch normalization layer is introduced before the first convolutional layer to ensure dimensionality consistency. The three convolutional layers have 128, 64, and 32 filters, respectively. These high-level features are then fed into two fully connected layers, each containing 128 and 64 neurons, stacked together. The entire adaptive feature learning process is performed within these fully connected layers, followed by a softmax layer used as a classifier. The standard ReLU activation function is used throughout the network. The output of the ReLU activation is represented as follows:

[0059] f(x) = max(0,x)

[0060] When the input is positive: the output equals the input, and the derivative is 1; when the input is negative: the output is 0, and the derivative is 0.

[0061] The softmax layer, used as a classifier, is a core component of CNN classification tasks. It achieves efficient multi-class decision-making through probabilistic output and cross-entropy loss optimization.

[0062] For the input vector z = [z1, z2, ..., z...] K (K is the number of categories), the softmax output is:

[0063]

[0064] In the softmax layer, the process of probabilistic output includes: converting the raw score output by the last layer (fully connected layer) of the neural network into a probability distribution, such that the probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1.

[0065]

[0066] In the formula, For each score z i The purpose of indexation is to amplify the impact of high scores, where K is the number of categories;

[0067] In the softmax layer, cross-entropy loss is optimized to measure the probability distribution predicted by the model, p = [p1, p2, ..., p]. K The true label distribution y = [y1, y2, ..., y] is the same as the true label distribution.K Differences between:

[0068] Loss for a single sample:

[0069]

[0070] In the formula, K is the number of categories, p i Let y be the probability distribution predicted by the model. i This represents the true label distribution;

[0071] Loss on batch samples:

[0072]

[0073] In the formula, K is the number of categories, N is the number of samples, and p i Let y be the probability distribution predicted by the model. i This represents the true label distribution;

[0074] To accurately evaluate the performance of the proposed region convolutional neural network model, the model was trained using both region convolutional neural network (R-CNN) and CNN neural network in the experiment.

[0075] Figure 3 The image shows a sustained spectrum for fault diagnosis, with frequency on the horizontal axis and power spectrum on the vertical axis. Certain regions are highlighted in the image; as the fault worsens, components in these regions remain in the signal for longer periods and with higher amplitudes. Furthermore, the sustained spectrum image shows certain differences in each case. These features allow region-based convolutional neural networks to distinguish the degree of bearing corrosion using the sustained spectrum image.

[0076] The corrosion level was divided into three grades: healthy state and two different corrosion levels. Grade 1 corrosion was the damage level of a brand-new bearing after 30 hours of operation in a motor, and Grade 2 corrosion was the damage level of a brand-new bearing after 60 hours of operation in a motor. Each corrosion level has a different topological structure, which is its main characteristic.

[0077] Stray magnetic flux signals from the coil flux sensor were extracted using a sampling frequency of 20kHz. Based on different voltages and load levels (percentage of rated load), the following different operating conditions were designed.

[0078] Operating condition type:

[0079] Operating conditions Voltage (V) Load level Operating Condition 1 100 Unloaded Operating Condition 2 100 50％ Operating Condition 3 100 75％ Operating Condition 4 100 100％ Operating Condition 5 200 Unloaded Operating Condition 6 200 50％ Operating Condition 7 200 75％ Operating Condition 8 200 100％ Operating Condition 9 300 Unloaded Operating Condition 10 300 50％ Operating Condition 11 300 75％ Operating Condition 12 300 100％

[0080] Diagnostic accuracy of each model under 100V conditions:

[0081]

[0082]

[0083] Diagnostic accuracy of each model under 200V conditions:

[0084] Operating conditions R-CNN CNN 5 95.1％ 63.2％ 6 96.4％ 51.5％ 7 96.9％ 58.1％ 8 97.2％ 53.7％

[0085] Diagnostic accuracy of each model under 300V conditions:

[0086] Operating conditions R-CNN CNN 9 96.3％ 67.4％ 10 97.1％ 64.9％ 11 97.5％ 60.2％ 12 98.6％ 55.8％

[0087] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. The scope of the present invention is defined only by the appended claims.

Claims

1. A method for diagnosing bearing corrosion based on region convolutional networks, characterized in that, Includes the following steps: Step 1: Acquire stray magnetic flux signals using a coil flux sensor; Step 2: Extract the power spectrum of the acquired stray magnetic flux signal to obtain the continuous spectrum image, which is the binary histogram accumulated in the power frequency space. Step 3: The continuous spectrum image from Step 2 is cropped and adjusted, and used as input to a region convolutional neural network, which includes a selective search algorithm and a CNN neural network. The region convolutional neural network is used to diagnose the degree of corrosion of the motor bearing. In step 2, the acquired stray magnetic flux signal is processed and decomposed into axial and radial components, and the power spectrum of the axial and radial components is extracted and a continuous spectrum image is plotted. The continuous spectrum image is generated in two dimensions with frequency as the horizontal axis and power spectrum as the vertical axis. The continuous spectrum image is used to characterize the correlation between bearing corrosion state and magnetic flux signal changes. Step 3, the process of diagnosing the corrosion level of the motor bearing using a region convolutional neural network, includes: Step 3-1: Process the continuous spectrum image. Using the cropped and adjusted continuous spectrum image as input, use a selective search algorithm to identify potential areas for bearing corrosion diagnosis. Step 3-2: Input the potential regions identified in Step 3-1 into the CNN neural network, use the CNN neural network to extract shallow representations for judging the degree of bearing corrosion, use the shallow features to classify the degree of bearing corrosion, and obtain the diagnostic results. In step 3-2, the shallow representations used to determine the degree of bearing corrosion are extracted using a CNN neural network, and the degree of bearing corrosion is classified using the shallow features. This includes inputting a 32×32 region proposal into a primary two-dimensional convolutional block to extract a fixed-length feature vector. These high-level features are input into two fully connected layers and then into a softmax layer that serves as a classifier.

2. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 1, characterized in that, In step 2, the power spectrum of the acquired stray magnetic flux signal is extracted. The power spectrum extraction expression is: ； ； In the formula: For time The Discrete Fourier Transform (DTF) of the center window data; For a moment The input signal, It's a window function. It's the follow-up. The number of samples between.

3. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 1, characterized in that: The sustained spectrum image described in step 3-1 represents the relationship between frequency and power, and each frequency band of the power frequency pair represents the percentage of time it exists in the signal evolution. The redder the color in the sustained spectrum image, the longer the specific power frequency band lasts in the signal.

4. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 1, characterized in that, In step 3-1, the selective search algorithm includes: investigating all edge regions with peak amplitudes attributable to different degrees of corrosion and classifying them as anchor points or bounding boxes; performing a regression task on the input continuous spectrum image using non-maximum suppression and generating a set of potential regions using bounding boxes; the non-maximum suppression loU>=0.5 is performed as follows: (1) the regions are arranged from largest to smallest score; (2) redundancy is eliminated, and all regions with loU>=0.5 with the region with the largest score are eliminated; (3) the region with the largest score is retained, and the remaining regions are used as a new candidate set.

5. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 1, characterized in that, The softmax layer, used as a classifier, is a core component of CNN classification tasks. It achieves efficient multi-class classification decisions through probabilistic output and cross-entropy loss optimization.

6. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 5, characterized in that, In the softmax layer, the process of probabilistic output includes: The raw scores output by the last layer of the neural network, the fully connected layer, are converted into a probability distribution such that the probability value of each category is between 0 and 1, and the sum of the probabilities of all categories is 1. ； In the formula, For each score Exponentialization is used to amplify the impact of high scores. This represents the number of categories.

7. The method for diagnosing bearing corrosion based on a region convolutional network according to claim 6, characterized in that, In the softmax layer, the cross-entropy loss optimization measures the probability distribution predicted by the model. Distribution of real labels Differences between them: Loss for a single sample: ； In the formula, For the number of categories, The probability distribution predicted by the model. This represents the true label distribution; Loss on batch samples: ； In the formula, For the number of categories, For the number of samples, The probability distribution predicted by the model. This represents the true label distribution.

Images