Multistage pump fault diagnosis method
The accuracy and efficiency of multi-stage pump fault diagnosis are improved by using adaptive weight vectors and frequency attention modules, which solves the problems of low efficiency and insufficient sensitivity in traditional methods and achieves efficient fault identification in strong noise environments.
Patent Information
- Application Number
- CN202511913152.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-02-06
AI Technical Summary
Existing multistage pump fault diagnosis methods rely on expert experience and are inefficient. Traditional deep learning methods are not sensitive enough to key fault feature frequencies and are easily drowned out by background noise and normal operating vibrations, resulting in low diagnostic accuracy and efficiency.
By calculating the structural parameters and real-time speed of a multi-stage pump, an adaptive weight vector is generated, a narrow-band frequency window is constructed, the time-frequency graph of key feature frequencies is enhanced, and a frequency attention module is introduced into the CNN model to improve the ability to identify key fault frequencies.
It significantly improves the accuracy and efficiency of multi-stage pump fault diagnosis, maintains high robustness under strong noise interference, and increases the recognition rate to 99.28%, which is 3.1% to 7.4% better than traditional methods.
Smart Images

Figure CN121479489A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of multistage pumps, and in particular to a method for diagnosing faults in multistage pumps. Background Technology
[0002] With the deepening of my country's new urbanization strategy and the comprehensive implementation of urban renewal initiatives, urban spatial forms continue to expand upwards and underground, with high-rise and super high-rise buildings and large communities becoming the mainstream form of urban living. This trend has placed unprecedentedly high standards and stringent requirements on the safety and security of urban secondary water supply systems. As the core power equipment of secondary water supply systems, the operational efficiency and health status of multistage centrifugal pumps are decisive factors in determining whether the water supply system can achieve the goals of "safety, stability, efficiency, and low carbon emissions."
[0003] However, current traditional vibration analysis relies on expert experience to extract features (such as FFT spectra), which is inefficient and highly subjective. Existing deep learning methods (such as CNNs that directly use the original signal or standard time-frequency signal input) are not sensitive enough to the key fault feature frequencies and are easily overwhelmed by strong background noise and normal operating vibrations. Summary of the Invention
[0004] To address the shortcomings of existing technologies, one of the objectives of this application is to provide a multi-stage pump fault diagnosis method that has the advantages of targeting key fault frequencies and improving diagnostic accuracy and efficiency.
[0005] The above-mentioned objective of this application is achieved through the following technical solution: A method for diagnosing faults in a multi-stage pump includes the following steps: Step S1: Collect and preprocess the pump vibration signal and / or sound signal; Step S2: Based on the pump's structural parameters and real-time speed, calculate the theoretical characteristic frequencies corresponding to different faults. Based on the calculated characteristic frequencies, construct an adaptive weight vector and generate samples using this weight. Step S3: Input the samples into the CNN model for training to obtain the results; Step S4: Output the classification results of the CNN.
[0006] In a preferred embodiment, this application can be further configured such that, in step S2, the adaptive weight vector is generated by setting a narrow-band frequency window centered on the feature frequency, where frequency components within the window are assigned high weights, while frequency components outside the window are assigned basic weights.
[0007] In a preferred embodiment, this application can be further configured as follows: in step S2, the signal is transformed according to the weight adaptive vector to obtain an initial time-frequency matrix, and then this matrix is multiplied by a two-dimensional weight mask constructed according to the weight vector to obtain an enhanced time-frequency map.
[0008] In a preferred example, this application can be further configured such that, in step S3, a frequency band attention module is placed before the CNN model, and the attention module adds specific weights to the specified frequency band.
[0009] In a preferred embodiment, this application can be further configured such that the narrowband frequency window in step S2 is located within a specified frequency band.
[0010] This application has the following advantages: Because the input image has been preprocessed and key features are highlighted, CNNs can converge faster: they do not need to painstakingly learn features from massive amounts of irrelevant information. After specialization, the attention of the CNN model is directed to the most discriminative region. Furthermore, the method described in this invention is insensitive to noise and load variations, exhibiting higher robustness than traditional fault identification methods and traditional CNN methods. Attached Figure Description
[0011] Figure 1 This is a schematic diagram of the weight adaptive curve of this application.
[0012] Figure 2 This is a comparison diagram of the standard CWT time-frequency diagram and the enhanced CWT time-frequency diagram of this invention.
[0013] Figure 3 This is the original diagnosis result.
[0014] Figure 4 This is the diagnostic result of this application. Detailed Implementation
[0015] The present application will be further described in detail below with reference to the accompanying drawings.
[0016] Reference Figures 1-4 The present application discloses a multi-stage pump fault diagnosis method, which includes the following steps: Step S1: Multi-source signal acquisition and preprocessing Vibration and / or sound signals are acquired by using vibration sensors and / or acoustic sensors installed in key parts of the multi-stage pump (such as bearing housings and pump casings). The acquired raw signals are preprocessed, including denoising, detrending, normalization, integration (merging the processed signals into a sample library), and sample naming.
[0017] Step S2: Calculation of Key Fault Feature Frequencies and Generation of Weight Vectors Based on the structural parameters of the multistage pump (selecting parameters according to the type of bearing: one or more of the following: inner ring speed, pitch diameter, number of rolling elements, rolling element diameter, contact angle) and real-time speed, calculate the theoretical characteristic frequencies corresponding to different faults, including but not limited to: shaft frequency, blade passage frequency, bearing outer ring fault frequency, bearing inner ring fault frequency, rolling element fault frequency, and cage fault frequency.
[0018] Based on the calculated theoretical feature frequencies, an adaptive weight vector is constructed. This vector is generated as follows: a narrow-band frequency window (e.g., f_feature ±Δf) is defined centered on these feature frequencies, where f_feature is the calculated frequency (this frequency can be BPFI, BPFO, or other frequencies), and ±Δf represents the narrowband width. Figure 1 This is manifested as (weight adjustments are added around fault frequencies such as BPFI, BPFO, or BSF). Frequency components within this window are assigned high weights (e.g., 1.5-5.0), while frequency components outside the window are assigned basic weights (e.g., 1.0). Different weight values can be assigned to multiple different characteristic frequencies. The signal is transformed according to the weighted adaptive vector, and the processed sample is obtained.
[0019] The acquired signal is subjected to a conventional CWT (or, in other embodiments, short-time Fourier transform, Hilbert-Huang transform, Mel-spectrum transform, etc.) to obtain an initial time-frequency matrix TFR_initial. This matrix is then multiplied by a two-dimensional weight mask constructed based on the weight vector (i.e., by covering the sample layer (time-frequency matrix) with a weight mask, the sample layer is emphasized), resulting in the enhanced time-frequency map TFR_enhanced. This mask is consistent along the time axis but varies along the frequency axis according to the weight vector.
[0020] Incorporate the following code when performing CWT transformations: # Let's assume we're interested in frequencies around BPFO (93Hz) bpfo = 93 First, obtain the amplitude of the time-frequency matrix of the CWT transform. TFR_standard = np.abs(coeffs) # Assuming the fault frequency we are interested in is 10Hz, construct an adaptive weight vector by creating a weight vector that is identical in shape to frequencies, with an initial value of 1 (base weight). center_freq = bpfo bandwidth = 10 weight_vector = np.ones_like(frequencies) #Add frequency indices near BPFO during CWT transformation and give these indices high weights (e.g., 2.5x enhancement). indices_to_enhance = np.where((frequencies>center_freq - bandwidth)&(frequencies <center_freq + bandwidth))[0] enhancement_factor=2.5 weight_vector[indices_to_enhance] = enhancement_factor # Create a weight mask and apply it to the entire time-frequency plot, expanding the one-dimensional weight vector into a two-dimensional mask with the same shape as the TFR, and weighting the original time-frequency plot to obtain the enhanced time-frequency plot. weight_mask=np.tile(weight_vector[:,np.newaxis],(1,TFR_standard.shape[1])) TFR_enhanced = TFR_standard * weight_mask After this step, in the generated time-frequency map TFR_enhanced, the time-frequency energy of key fault frequency regions such as BPFO and BPFI is significantly amplified and enhanced, while the energy of other irrelevant frequency bands is relatively weakened.
[0021] Step S3: Train the CNN model that also has a frequency attention filtering mechanism by inputting the samples. The samples are imported into a frequency-attention module CNN model for training. This involves introducing a frequency-band attention module into the CNN model, which is identical to the time-frequency energy amplification region (i.e., the operation to enhance the time-frequency map) in preprocessing; the rest remains the same as the original CNN model. Since this method doesn't modify the subsequent CNN model but effectively increases the training weights for different frequency bands during training, it is applicable to all CNN models, such as 2D-CNN or 1D-CNN.
[0022] By defining a specific frequency band, assuming the frequency band is within the range of low1 Hz to high1 Hz. def frequency_attention_block(input_tensor, frequency_bands=(low1,high1)): #Frequency Attention Module #frequency_bands is set to the list of frequency ranges that need to be given more attention [(low1, high1), (low2, high2), ...] # Global average pooling acquisition frequency dimension gap = layers.GlobalAveragePooling2D()(input_tensor) # Frequency Attention Weights attention_weights=layers.Dense(input_tensor.shape[-1],activation='sigmoid')(gap)attention_weights=layers.Reshape((1,1,input_tensor.shape[-1]))(attention_weights) # Specify frequency bands and manually increase the weight of these regions. You can also add manual weight adjustment logic to assign higher weights to specific frequency channels. if frequency_bands is not (low1, high1): pass output = layers.Multiply()([input_tensor, attention_weights]) return output The enhanced time-frequency map TFR_enhanced is used as input and fed into a convolutional neural network (CNN model) with a frequency attention module for training and recognition.
[0023] Because the input image has been preprocessed and key features are highlighted, CNNs can converge faster: they do not need to painstakingly learn features from massive amounts of irrelevant information. After specialization, the attention of the CNN model is directed to the most discriminative region. Furthermore, the method in this application is insensitive to noise and load variations, exhibiting higher robustness than traditional fault identification methods and traditional CNN methods.
[0024] Step S4: Output of diagnostic results Output the CNN classification results (fault type and probability) to the user interface or monitoring system.
[0025] Taking a multistage pump as an example, a pump bearing failure state was artificially created by means of destruction. Vibration signals were collected under normal conditions and 24 failure states. The 24 failure states include 3 types of failures (inner ring failure, outer ring failure, and roller failure), each type with 4 different dimensions (0.2mm, 0.4mm, 0.6mm, and 0.8mm), totaling 12 individual failures, and 3 combined failure states (inner + outer, inner + roller, and outer + roller), also with 4 different dimensions, totaling 12 combined failures. A total of 25 signals were tested.
[0026] The actual environment of the test bench is an open, non-anechoic environment with a lot of signal interference.
[0027] First, the fault characteristic frequencies of the installed machinery were identified, and the fault characteristic frequencies BPFI, BPFO, BSF, and BPF were calculated, yielding results of 225Hz, 161Hz, 282Hz, and 290Hz. Incorporate the following code when performing CWT transformations: # Introducing 4 fault characteristic frequencies bpfi = 225 bpfo = 161 bsf = 282 bpf = 290 TFR_standard = np.abs(coeffs) # Construct a weight vector. Create a weight vector that is identical in shape to frequencies, with an initial value of 1 (base weight). fault_frequencies = { 'BPFI': {'freq': bpfi, 'bandwidth': 15, 'enhancement': 3.0, 'color': 'red'}, 'BPFO': {'freq': bpfo, 'bandwidth': 12, 'enhancement': 3.5, 'color': 'blue'}, 'BSF': {'freq': bsf, 'bandwidth': 10, 'enhancement': 2.5, 'color':'green'}, 'FTF': {'freq': bpf, 'bandwidth': 8, 'enhancement': 2.0, 'color': 'orange'}} #Add frequency indices near BPFO during CWT transformation and assign high weights to these indices, such as 3.0, 3.5, 2.5, 2.0, etc.
[0028] indices = np.where((frequencies>center_freq - bandwidth / 2)&(frequencies <center_freq + bandwidth / 2))[0] enhancement_factor = params['enhancement'] weight_vector[indices] = enhancement print(f" {fault_name}: {center_freq}Hz ±{bandwidth / 2}Hz ->enhancement} times") # Create a weight mask and apply it to the entire time-frequency plot. Weight the original time-frequency plot according to frequency to obtain the enhanced time-frequency plot. weight_mask = np.tile(weight_vector[:, np.newaxis], (1, TFR_standard.shape[1])) TFR_enhanced = TFR_standard * weight_mask ... Use the generated images as training samples and add the following code before the original CNN model: ... #Introducing the attention-focusing section def frequency_attention_block(input_tensor, frequency_bands=(220,230),(155, 165),(275, 295)): #Set the frequency range requiring increased attention to [(220, 230), (155, 165), (275, 295)] gap = layers.GlobalAveragePooling2D()(input_tensor) # Frequency Attention Weights attention_weights=layers.Dense(input_tensor.shape[-1], activation='sigmoid')(gap) attention_weights=layers.Reshape((1,1,input_tensor.shape[1]))(attention_weights) # Specify frequency bands and manually increase the weight of these regions. You can also add manual weight adjustment logic to assign higher weights to specific frequency channels. if frequency_bands is not (220, 230),(155, 165),(275, 295): pass output = layers.Multiply()([input_tensor, attention_weights]) return output Subsequently, training was performed according to the original training process, and the recognition results of the conventional method were compared with those of the method described in this application. First, a conventional signal acquisition and preprocessing process was used, and a non-specialized convolutional neural network was employed for fault identification, achieving an average fault identification rate of 92.19%. However, using the method described in this application, after adding signal attention frequencies of 48Hz, 161Hz, 226Hz, and 282Hz, and adding a frequency attention module before the original CNN model and training it, the average fault identification rate increased to 99.28%. The attention mechanism designed for specific fault characteristic frequencies in this invention outperforms general attention methods by approximately 3.1% to 7.4% in average recognition rate, fully demonstrating its excellent robustness and reliability even under strong noise interference.
[0029] The implementation principle of this embodiment is as follows: after signal preprocessing, key features are highlighted, thus enabling the CNN to converge faster: it eliminates the need for arduous feature learning from massive amounts of irrelevant information. After specialization, the attention of the CNN model is guided to the most discriminative region. Furthermore, the method described in this invention is insensitive to noise and load variations, exhibiting higher robustness than traditional fault identification methods and traditional CNN methods.
[0030] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A method for diagnosing faults in a multi-stage pump, characterized in that: Includes the following steps: Step S1: Collect and preprocess the pump vibration signal and / or sound signal; Step S2: Based on the pump's structural parameters and real-time speed, calculate the theoretical characteristic frequencies corresponding to different faults. Based on the calculated characteristic frequencies, construct an adaptive weight vector and generate samples using this weight. Step S3: Input the samples into the CNN model for training to obtain the results.
2. The multi-stage pump fault diagnosis method according to claim 1, characterized in that: In step S2, the adaptive weight vector is generated as follows: a narrow frequency window is set with the feature frequency as the center. The frequency components within the window are given high weights, while the frequency components outside the window are given basic weights.
3. The multi-stage pump fault diagnosis method according to claim 2, characterized in that: In step S2, the signal is transformed according to the weight adaptive vector to obtain the initial time-frequency matrix. Then, this matrix is multiplied by a two-dimensional weight mask constructed based on the weight vector to obtain the enhanced time-frequency map.
4. The multi-stage pump fault diagnosis method according to claim 3, characterized in that: In step S3, a frequency band attention module is added before the CNN model, and the attention module adds specific weights to the specified frequency band.
5. The multi-stage pump fault diagnosis method according to claim 4, characterized in that: The narrowband frequency window in step S2 is located within the specified frequency band.
Citation Information
Patent Citations
Rotary equipment fault diagnosis method and device
CN116595346A
Bearing interpretable fault diagnosis method based on digital-analog collaborative federated learning
CN121145594A
