Time-frequency fault diagnosis method based on frequency band filtering

Abstract

The invention relates to a time-frequency fault diagnosis method based on frequency band filtering. The method comprises the following steps: (1) acquiring a time-domain sound signal of electromechanical equipment through a sensor; (2) carrying out fast spectral kurtosis analysis on the time domain signal, determining the center frequency and bandwidth of an optimal band-pass filter, and carrying out frequency band filtering by using a Butterworth filter; (3) dividing the filtered signal into equal-length fragments, and generating a time-frequency graph through short-time Fourier transform (STFT); (4) constructing a two-path parallel convolutional neural network (CNN), processing the filtered one-dimensional time sequence signal in the first path, processing the two-dimensional time-frequency graph in the second path, extracting features respectively, and performing feature fusion through flattening and splicing; and (5) inputting the fused features into a Softmax classifier, and outputting a fault classification result. The accuracy rate of the method in the measured data set reaches 97.92%, and the method is obviously superior to a traditional method and is suitable for intelligent operation and maintenance of industrial equipment.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical equipment status monitoring and fault diagnosis, and specifically involves technologies such as frequency band filtering, short-time Fourier transform (STFT), and parallel convolutional neural network (CNN), and is suitable for non-invasive fault detection of electromechanical equipment. Background Art

[0002] Failure in electromechanical equipment can reduce its service life and potentially cause sudden downtime, leading to economic losses and safety hazards. Therefore, it is crucial to establish a practical algorithmic tool to accurately, quickly, and reliably diagnose the complex nonlinear correlation between raw signals and failure modes.

[0003] Traditional fault diagnosis methods typically include three steps: feature extraction, feature selection, and fault classification. Feature extraction is the most critical, directly impacting diagnostic accuracy. Traditional feature extraction methods include time domain, frequency domain, and time-frequency domain. Common time domain analysis methods include mean, variance, and kurtosis; common frequency domain analysis methods include envelope spectrum and frequency spectrum; and common time-frequency domain analysis methods include empirical mode decomposition, wavelet transform, short-time Fourier transform, and variational mode decomposition.

[0004] The above methods have low computational overhead, but require manual extraction of shallow signal features before classification. They rely on domain knowledge and parameter settings, and have a high risk of underfitting or overfitting, resulting in weak generalization performance and low accuracy, especially when facing large data sets.

[0005] In summary, the existing technology has the following deficiencies:

[0006] (1) Low diagnostic accuracy: Due to the complex operating environment of electromechanical equipment, fault information is submerged in other components, and the accuracy of direct diagnosis of full-band signals is low.

[0007] (2) Requires prior knowledge input: relies on expert knowledge and manually extracted features.

[0008] (3) Weak data processing capabilities: Traditional fault diagnosis methods, when faced with large data sets, rely on manual feature extraction and domain knowledge, making it difficult to mine deep-level features. They often suffer from underfitting or overfitting problems, resulting in decreased generalization performance and diagnostic accuracy.

[0009] For example, in the existing technology, traditional fault diagnosis methods rely on manual feature extraction (such as time domain mean, frequency domain envelope spectrum) and expert experience, which has the following defects:

[0010] 1. Low diagnostic accuracy: Fault characteristics are easily overwhelmed by noise in complex environments;

[0011] 2. Poor generalization ability: Shallow feature extraction can easily lead to overfitting or underfitting;

[0012] 3. Weak data processing capabilities: Difficulty in processing large data sets.

[0013] 4. Patent document CN110334886A discloses a diagnostic method based on FFT and CNN, but does not use frequency band filtering and does not combine time-frequency domain features; patent document CN119442129A proposes multi-domain feature fusion, but relies on the attention mechanism and has high computational complexity.

[0014] The present invention significantly improves accuracy and efficiency through frequency band filtering + dual-path CNN. Summary of the Invention

[0015] The present invention aims to propose a time-frequency fault diagnosis method based on frequency band filtering. In view of the problems that the operating environment of electromechanical equipment is complex, the fault features of collected signals are easily interfered and difficult to extract, and the generalization ability of conventional diagnostic methods is poor, the method uses frequency band filtering to find the frequency with the richest information, reduces the influence of noise and redundant information on feature extraction, mines the time-frequency domain features of the signal through short-time Fourier transform, and then uses a dual-path parallel convolutional neural network to extract features from the filtered time series and time-frequency graph respectively, reducing the dependence of manually extracted features on expert knowledge. Finally, a classifier is used to classify the features to achieve accurate identification of electromechanical equipment faults.

[0016] To achieve the above object, the technical solution of the present invention is: a time-frequency fault diagnosis method based on frequency band filtering, comprising the following steps:

[0017] (1) Collecting time-domain sound signals of electromechanical equipment through sensors;

[0018] (2) performing a fast spectral kurtosis analysis on the time domain signal to determine the center frequency and bandwidth of the optimal bandpass filter, and performing frequency band filtering using a Butterworth filter;

[0019] (3) Segment the filtered signal into segments of equal length and generate a time-frequency map through short-time Fourier transform (STFT);

[0020] (4) A two-way parallel convolutional neural network (CNN) is constructed. The first way processes the filtered one-dimensional time series signal, and the second way processes the two-dimensional time-frequency graph. The features are extracted from each way and then fused by flattening and splicing. (5) The fused features are input into the Softmax classifier to output the fault classification results.

[0021] Furthermore, in step (2), a frequency band with the largest kurtosis is selected as a bandpass filter parameter through fast spectral kurtosis analysis, and the center frequency of the frequency band is 2560 Hz and the bandwidth is 1024 Hz.

[0022] Furthermore, in step (3), after filtering, the signal is divided into segments of equal length, and the segmentation method is as follows:

[0023] x=[x 1：h , x h+1:2h ,...,x (s / h-1)h+1:n ]

[0024] Where h represents the length of the time window used for segmentation, and a signal is divided into s / h segments.

[0025] Furthermore, the window length of the signal segmentation is h=1024, the total signal length is n=32768, and the number of segments after segmentation is 32.

[0026] Furthermore, in step (4), the input of the first CNN is a one-dimensional time series of 1×1024, which passes through the first convolution layer with a convolution kernel of 6×1, the second convolution layer with a convolution kernel of 16×1, and the maximum pooling layer in sequence; the input of the second CNN is a time-frequency map of 3×32×32, which passes through the first convolution layer with a convolution kernel of 8×1, the second convolution layer with a convolution kernel of 16×1, and the maximum pooling layer in sequence.

[0027] Furthermore, in step (4), the one-dimensional convolutional neural network calculation formula used is as follows:

[0028]

[0029] in, is the output feature of the lth layer, is the input of the l-1 layer; is the weight of the convolution kernel, is the bias of the convolution kernel; f is the corresponding activation function;

[0030] The calculation formula of the maximum pooling used is:

[0031]

[0032] Where i, j are the row and column index values ​​of the output matrix z respectively; l, h are the row and column index values ​​within the pooling window respectively; r is the width of the pooling area.

[0033] Furthermore, in step (4), feature fusion is performed by concatenating the feature vectors output by the two CNNs into a single vector, inputting the vector into the fully connected layer, and then connecting the vector to the Softmax classifier.

[0034] Furthermore, in step (5), the loss function of the Softmax classifier is cross entropy loss, the optimizer uses Adam, and the learning rate is set to 0.0001, where:

[0035] The calculation formula of Softmax is:

[0036]

[0037] Where z i is the output of the i-th node, K is the number of all nodes, P(z i ) is the output probability value.

[0038] Furthermore, the method also includes performing data enhancement on the original signal, dividing the data set into a training set, a validation set, and a test set in an 8:1:1 ratio, with no overlap between the sets.

[0039] Furthermore, the method has an accuracy rate of ≥97.92% and a precision rate of ≥98% in electromechanical equipment fault diagnosis, wherein:

[0040] The accuracy calculation formula is:

[0041]

[0042] The calculation formula for accuracy is:

[0043]

[0044] Where TP represents the number of positive classes that are judged as positive, FP represents the number of negative classes that are judged as negative, FN represents the number of positive classes that are judged as negative, and TN represents the number of negative classes that are judged as negative.

[0045] The beneficial effects of the present invention are:

[0046] The present invention belongs to the field of mechanical equipment condition monitoring and fault diagnosis, and specifically involves technologies such as frequency band filtering, short-time Fourier transform, and parallel convolutional neural network. It is applicable to most electromechanical equipment and has the following beneficial effects compared with existing technologies:

[0047] (1) High-precision and high-efficiency diagnosis: This invention uses frequency band filtering technology to effectively identify the frequencies richest in information, reducing the interference of noise and redundant information on feature extraction. This method improves the accuracy of fault feature extraction in the complex operating environment of electromechanical equipment, thereby enhancing the accuracy and efficiency of fault diagnosis.

[0048] (2) No prior knowledge required: By using short-time Fourier transform to mine the time-frequency domain features of the signal and utilizing a dual-path parallel convolutional neural network to automatically extract features from the filtered time series and time-frequency graph, the reliance on expert knowledge during manual feature extraction is reduced. This automated feature extraction process reduces the requirements for parameter settings and domain knowledge, and improves the generalization and applicability of the diagnostic method.

[0049] (3) Powerful data processing capabilities: Traditional fault diagnosis methods, when faced with large data sets, rely too much on domain knowledge when extracting features manually. If the extracted features are not comprehensive enough, it is easy to make it difficult to identify the failure mode of the equipment. The present invention improves the processing capabilities of large data sets through automated feature extraction and classification processes, making the fault diagnosis method more robust, especially in complex and changing practical application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a fault diagnosis flow chart of the present invention;

[0051] Figure 2 is the result of fast spectral kurtosis analysis (the maximum kurtosis frequency band is located at the 4th layer);

[0052] Figure 3 It is a parallel convolutional neural network structure diagram;

[0053] Figure 4 is the diagnosis result confusion matrix. DETAILED DESCRIPTION

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] like Figure 1 As shown, the time-frequency fault diagnosis method based on frequency band filtering of the present invention has the following specific steps:

[0056] Step 1: Signal Acquisition

[0057] Acoustic sensors are used to collect sound signals generated during the operation of electromechanical equipment.

[0058] Step 2: Frequency band filtering

[0059] In the presence of noise and other interfering signals, fast spectral kurtosis analysis can determine optimal bandpass filter parameters, such as center frequency and bandwidth. By selecting the frequency band with the highest kurtosis for bandpass filtering, fault signals can be better highlighted during signal processing, improving the accuracy of fault diagnosis. Fast spectral kurtosis assists in fault diagnosis by studying the frequency domain characteristics of the signal, enhancing the diagnostic technology's resistance to interference and enabling more precise identification of fault characteristics. Its high accuracy, strong interference resistance, and wide range of applications make it highly valuable in engineering practice.

[0060] Step 3: Time-frequency feature extraction

[0061] The short-time Fourier transform (STFT) is a time-frequency analysis technique used to analyze time-varying, nonstationary signals. It converts a one-dimensional vibration signal into a matrix that contains characteristic information in the time and frequency domains, making it easier for convolutional neural networks to process. The STFT is based on the traditional Fourier transform. It first multiplies the signal by a window function and then performs a one-dimensional Fourier transform on the product. By shifting the window function and repeating this process, a series of spectra are generated, which are combined to form a time-frequency plot.

[0062] The formula for short-time Fourier transform is:

[0063]

[0064] Where x(t) is the time domain signal; h(t-ω) is the window function; e -jωt represents the complex exponential function, and dt represents the integration over time t.

[0065] Step 4: Feature Fusion

[0066] A convolutional neural network is a deep learning model that extracts features through convolution. It is primarily composed of convolutional layers, pooling layers, and fully connected layers. The model accepts two independent inputs: a one-dimensional time series signal and a time-frequency graph signal. The convolutional layer contains multiple convolution kernels, which regularly sweep over the elements of the input data, performing matrix multiplication and summing the elements and adding a bias. The pooling layer uses max pooling, which discards some low-weight feature data while retaining the highest-weight features, reducing computational complexity. After the max pooling layer, feature vectors from different channels are fused through flattening and feature fusion strategies, allowing the model to extract more comprehensive and accurate features. The fused vectors share a fully connected layer, which combines the features extracted from the connected nodes and introduces nonlinearity through activation functions, enabling the model to learn more complex function mappings.

[0067] Step 5: Troubleshooting

[0068] The extracted feature vector is input into a Softmax layer to generate the predicted probability of each category, obtain the final classification result, and realize the identification of electromechanical equipment faults.

[0069] Example 1:

[0070] Step 1: Signal acquisition: Place sensors on the electromechanical equipment and use data acquisition equipment to collect signals during operation.

[0071] Step 2: Frequency band filtering. Perform a fast spectral kurtosis analysis on the data collected in step 1. The results are shown in the attached Figure 2The maximum kurtosis occurs at the fourth layer, and the corresponding optimal frequency band is centered at 2560 Hz with a bandwidth of 1024 Hz. The Butterworth filter parameters are set, including the passband attenuation αp, the stopband attenuation αs, the passband upper cutoff frequency Ωp, and the stopband lower cutoff frequency Ωs. The Butterworth bandpass filter is used to filter the time domain signal, removing irrelevant or unimportant frequency components from the signal and retaining only the frequency range containing useful information, highlighting the useful signal and reducing the amount of calculation.

[0072] Step 3: Time-frequency feature extraction. After filtering, the signal is divided into segments of equal length. The segmentation method is as follows:

[0073] x=[x 1：h , x h+1:2h ,...,x (s / h-1)h+1：n ]

[0074] In the formula, h represents the length of the time window used for segmentation; a signal is divided into s / h segments. Here, n is 32768 and h is 1024. The entire dataset is divided in an 8:1:1 ratio, with no overlap between the training, validation, and test sets. This not only expands the sample size of the entire dataset but also reduces overfitting during network training.

[0075] The segmented segments are used for short-time Fourier transform to construct a sample time-frequency feature map dataset.

[0076] Step 4: Feature fusion. During the network model training process, the cross entropy loss function is used to quantify the gap between the model prediction value and the true value. The Adam optimizer is used and the learning rate is set to 0.0001. The parallel convolutional neural network structure is shown in the attached figure. Figure 3 As shown in the figure, for a one-dimensional time series, the input size is 1×1024. It first passes through the first convolution layer with 6 convolution kernels, a stride of 1, and a convolution kernel size of 6; then passes through a pooling layer with a size of 2 and a stride of 2; then passes through a second convolution layer with 16 convolution kernels, a stride of 1, and a convolution kernel size of 16; then passes through a pooling layer with a size of 2 and a stride of 2. For a time-frequency feature map, the input size is 3×32×32. It first passes through the first convolution layer with 8 convolution kernels, a stride of 1, and a convolution kernel size of 1×1; then passes through a pooling layer with a size of 2 and a stride of 2; then passes through a second convolution layer with 16 convolution kernels, a stride of 1, and a convolution kernel size of 1×1; then passes through a pooling layer with a size of 2 and a stride of 2.

[0077] The calculation formula of one-dimensional convolutional neural network is as follows:

[0078]

[0079] in, is the output feature of the lth layer, is the input of the l-1 layer; is the weight of the convolution kernel, is the bias of the convolution kernel; f is the corresponding activation function.

[0080] The pooling layer uses the maximum pooling strategy. Maximum pooling is to process the maximum value of neurons in the pooling area. The calculation formula of maximum pooling is:

[0081]

[0082] Where i, j are the row and column index values ​​of the output matrix z respectively; l, h are the row and column index values ​​within the pooling window respectively; r is the width of the pooling area.

[0083] The input of the model is a one-dimensional time series and a time-frequency graph, each of which passes through the corresponding convolution layer and pooling layer. A flattening and feature fusion strategy is adopted to merge and splice the feature vectors of different channels to achieve feature fusion of the two channels, which are jointly used as the input of the Softmax classifier.

[0084] Step 5: Fault diagnosis. Use the Softmax classifier as the last layer of the neural network and update the network weights using the gradient descent algorithm to convert the raw scores output by the previous layer of the network into a category probability distribution and make predictions based on the category with the highest probability. The Softmax calculation formula is:

[0085]

[0086] Where z i is the output of the i-th node, K is the number of all nodes, P(z i ) is the output probability value.

[0087] The effectiveness of the method proposed in this aspect is tested by accuracy and precision.

[0088] The accuracy calculation formula is:

[0089]

[0090] The calculation formula for accuracy is:

[0091]

[0092] Where TP represents the number of positive classes that are judged as positive, FP represents the number of negative classes that are judged as negative, FN represents the number of positive classes that are judged as negative, and TN represents the number of negative classes that are judged as negative.

[0093] The method of the present invention was used to conduct a fault diagnosis experiment on a certain electromechanical equipment measured data set, where the normal signal was marked as 0 and the fault signal was marked as 1. The results are shown in the attached figure. Figure 4 As shown, the diagnostic accuracy rate reaches 97.92%, proving that the present invention can well identify healthy devices and faulty devices.

[0094] Example 2:

[0095] 1. Signal acquisition: Install acoustic sensors at the bearings of electromechanical equipment, sampling at a frequency of 51.2kHz, collecting 32,768 signal points.

[0096] 2. Frequency band filtering:

[0097] Fast spectral kurtosis analysis ( Figure 2 ) Determine the center frequency to be 2560 Hz and the bandwidth to be 1024 Hz;

[0098] Design a 4th-order Butterworth bandpass filter with a passband of 2048-3072Hz and a stopband attenuation of ≥40dB.

[0099] 3. Time-frequency diagram generation:

[0100] The filtered signal is divided into 32 segments (h=1024), and each segment is subjected to STFT (Hanning window, overlap rate 50%) to generate a 3×32×32 time-frequency diagram.

[0101] 4. Dual-channel CNN construction:

[0102] One-dimensional path: input 1×1024, sequentially passes through the convolution kernel 6×1 (output 6×1019), pooling 2×1 (output 6×509), convolution kernel 16×1 (output 16×494), pooling 2×1 (output 16×247);

[0103] 2D path: Input 3×32×32, passes through convolution kernel 8×1×1 (output 8×32×32), pooling 2×2 (output 8×16×16), and convolution kernel 16×1×1 (output 16×8×8);

[0104] Feature fusion: Flatten the two outputs into 1×3936 and 1×1024, concatenate them and input them into the fully connected layer.

[0105] 5. Training and testing:

[0106] The dataset was split into 8:1:1, Adam optimizer (lr = 0.0001), and cross entropy loss;

[0107] Test set confusion matrix ( Figure 4 ) showed an accuracy of 97.92% and a precision of 98.3%.

[0108] The present invention proposes a time-frequency fault diagnosis method based on frequency band filtering. The technical innovations are as follows:

[0109] 1. Frequency band filtering: Determine the optimal frequency band through fast spectral kurtosis analysis to suppress noise interference;

[0110] 2. Time-frequency feature extraction: STFT converts the signal into a time-frequency graph, preserving the joint time-frequency information;

[0111] 3. Dual-channel CNN feature fusion: Parallel processing of one-dimensional time series and two-dimensional time-frequency maps, reducing human intervention;

[0112] 4. Softmax classification: The fused features are classified through a fully connected layer and the fault probability is output.

[0113] 5. Technical Results: Accuracy increased to 97.92%, more than 15% higher than traditional methods; supports large data set processing, shortens training time by 30%; and achieves end-to-end automated diagnosis without the need for expert knowledge.

Claims

1. A time-frequency fault diagnosis method based on frequency band filtering, characterized in that: The following steps are involved: (1) Collecting time-domain sound signals of electromechanical equipment through sensors; (2) performing a fast spectral kurtosis analysis on the time domain signal to determine the center frequency and bandwidth of the optimal bandpass filter, and performing frequency band filtering using a Butterworth filter; (3) Segment the filtered signal into segments of equal length and generate a time-frequency map through short-time Fourier transform (STFT); (4) Construct a two-way parallel convolutional neural network (CNN), where the first way processes the filtered one-dimensional time series signal and the second way processes the two-dimensional time-frequency graph. After extracting features from each way, they perform feature fusion by flattening and splicing. (5) Input the fused features into the Softmax classifier and output the fault classification results.

2. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In the step (2), a frequency band with the largest kurtosis is selected as a bandpass filter parameter through fast spectral kurtosis analysis, wherein the center frequency of the frequency band is 2560 Hz and the bandwidth is 1024 Hz.

3. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In step (3), after filtering, the signal is divided into segments of equal length, and the segmentation method is as follows: x=[x 1：h ,x h+1:2h ,...,x (s / h-1)h+1:n ] Where h represents the length of the time window used for segmentation, and a signal is divided into s / h segments.

4. The method according to claim 3, wherein: The window length of signal segmentation h=1024, the total signal length n=32768, and the number of segments after segmentation is 32.

5. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In the step (4), the input of the first CNN is a one-dimensional time series of 1×1024, which passes through the first convolution layer with a convolution kernel of 6×1, the second convolution layer with a convolution kernel of 16×1, and the maximum pooling layer in sequence; the input of the second CNN is a time-frequency map of 3×32×32, which passes through the first convolution layer with a convolution kernel of 8×1, the second convolution layer with a convolution kernel of 16×1, and the maximum pooling layer in sequence.

6. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In step (4), the one-dimensional convolutional neural network calculation formula used is as follows: in, is the output feature of the lth layer, is the input of the l-1 layer; is the weight of the convolution kernel, is the bias of the convolution kernel; f is the corresponding activation function; The calculation formula of the maximum pooling used is: Where i, j are the row and column index values ​​of the output matrix z respectively; l, h are the row and column index values ​​within the pooling window respectively; r is the width of the pooling area.

7. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In step (4), feature fusion is performed by concatenating the feature vectors output by the two CNNs into a single vector, inputting the vector into the fully connected layer, and then connecting the vector to the Softmax classifier.

8. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: In step (5), the loss function of the Softmax classifier is cross entropy loss, the optimizer uses Adam, and the learning rate is set to 0.0001, where: The calculation formula of Softmax is: Where z i is the output of the i-th node, K is the number of all nodes, P(z i ) is the output probability value.

9. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: The method also includes performing data enhancement on the original signal, dividing the data set into a training set, a validation set, and a test set in an 8:1:1 ratio, with no overlap between the sets.

10. The time-frequency fault diagnosis method based on frequency band filtering according to claim 1, characterized in that: The method has an accuracy rate of ≥97.92% and a precision rate of ≥98% in electromechanical equipment fault diagnosis, where: The accuracy calculation formula is: The calculation formula for accuracy is: Where TP represents the number of positive classes that are judged as positive, FP represents the number of negative classes that are judged as negative, FN represents the number of positive classes that are judged as negative, and TN represents the number of negative classes that are judged as negative.

Images