Centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved EfficientNet-b0-TL

Abstract

This invention addresses the limitations of single signal sources in containing limited fault information, failing to comprehensively reflect the operating status of centrifugal fire pumps, and the low accuracy of fault diagnosis in centrifugal fire pumps due to the difficulty of dynamically and adaptively differentiating feature weights during fault diagnosis using convolutional neural networks (CNNs). To address these issues, a fault diagnosis method for centrifugal fire pumps based on multi-source information fusion and an improved EfficientNet-b0-TL model is proposed. Vibration signals from centrifugal fire pumps are collected from different angles and locations using multiple sensors, and the signals are cleaned, aligned, and denoised. A generalized S-transform is used to perform time-frequency transformation on the denoised signals, and time-frequency features are extracted and fused. A CBAM module is introduced to improve the EfficientNet-b0-TL model, constructing a fault diagnosis model. This model is then applied to actual centrifugal fire pump experimental data, and the results show that this method has high fault diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of centrifugal fire pump fault diagnosis technology, specifically a centrifugal fire pump fault diagnosis method involving multi-source signal fusion, wavelet threshold denoising, generalized S-time-frequency transform, and deep learning. Background Technology

[0002] As the core power equipment of modern building fire protection systems, the reliability of fire pumps is directly related to public safety and the safety of life and property. Centrifugal fire pumps, as key equipment in fire water supply systems, are widely used in fire protection systems of high-rise buildings, petrochemical plants, and large shopping malls. Centrifugal fire pumps, as key equipment in fire water supply systems, have complex structures and are in a static state for a long time. At the moment of startup, the operating conditions change suddenly, which can easily lead to various mechanical failures, reducing their function or even causing failure, resulting in the paralysis of the entire fire protection system, causing significant property damage and casualties. Therefore, it is of great significance to carry out fault diagnosis and operational status monitoring of centrifugal fire pumps.

[0003] Centrifugal fire pumps are mainly composed of pump casing, impeller, bearing, shaft, seals, and connectors. Traditional fault diagnosis of centrifugal fire pumps is mainly based on a single vibration signal source. During the operation of a centrifugal fire pump, there is mutual friction between components, which causes the vibration signal to be coupled multiple times and nonlinearized. There is coupling and aliasing between its features. If a single signal source is used for fault diagnosis, the accuracy is low. Traditional neural networks use static features for weight allocation in the fault diagnosis process, which causes problems such as lack of network adaptability and performance degradation. Summary of the Invention

[0004] This invention addresses the problems of incomplete representation of fault features from a single signal source and low fault diagnosis accuracy caused by the inability of the fault diagnosis model to adaptively and finely optimize and allocate weights during fault diagnosis of centrifugal fire pumps. It proposes a fault diagnosis method for centrifugal fire pumps based on multi-source information fusion and an improved EfficientNet-b0-TL model.

[0005] Its unique feature lies in the simultaneous use of centrifugal fire pump vibration signals and drive motor vibration signals as fault diagnosis signals for centrifugal fire pumps. After generalized S-time-frequency transformation, the time-frequency features are fused from multiple sources, thereby changing the way the fault features of centrifugal fire pumps are determined. Furthermore, the CBAM attention module is added to the EfficientNet-b0-TL model, enabling the model to have both adaptive and refined feature optimization and weight allocation capabilities.

[0006] The method includes the following steps: Step 1: Data Acquisition and Preprocessing. A first accelerometer is installed on the casing of the centrifugal fire pump, and a second accelerometer is installed on the casing of the drive motor. Vibration signals of the centrifugal fire pump under fault-free operation and under different fault conditions, as well as motor vibration signals, are collected simultaneously. This step further includes data cleaning, alignment, and signal denoising to ensure data consistency, eliminate background noise, and improve data quality.

[0007] Step 2: Extraction of time-frequency features from multi-source signals. The denoised signal obtained in Step 1 is segmented without overlap using a sliding window with a step size and window length of 1024 data points. Then, the generalized S-transform is used to perform time-frequency transformation on the segmented denoised signal, transforming the one-dimensional vibration signal into a two-dimensional time-frequency map, and extracting the time-frequency features of different signals.

[0008] Step 3: Feature Fusion and Sample Splitting. The two-dimensional time-frequency images obtained in Step 2 are subjected to feature fusion using a data-level fusion strategy. The fused samples are then randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0009] Step 4: Building and improving the fault diagnosis model Load the pre-trained EfficientNet-b0 model, and replace the global average pooling layer, fully connected layer, and output layer with pre-trained weights and parameters in the original pre-trained EfficientNet-b0 model with completely new global average pooling layer, fully connected layer, and output layer without pre-trained weights and parameters. Adjust the number of output neurons in the output layer after the replacement to meet the fault diagnosis task, thus building the transfer learning model EfficientNet-b0-TL. Then, insert a CBAM attention module after the 7th MBConv module and before the global average pooling layer in the EfficientNet-b0-TL model to improve the EfficientNet-b0-TL model. The improved EfficientNet-b0-TL model contains a total of 7 MBConv modules with pre-trained weights, 1 CBAM attention module with weights and parameters but no pre-training, 1 global average pooling layer without pre-trained weights and parameters, 1 fully connected layer without pre-trained weights and parameters, and 1 output layer without pre-trained weights and parameters. Step 5: Model Training The improved EfficientNet-b0-TL model was trained using the training sample set partitioned in step 3. During training, a strategy of freezing the weights and parameters of shallow networks was adopted. The parameters of the seven MBConv modules with pre-trained weights were frozen, so that the weights and parameters of the seven MBConv modules with pre-trained weights were not updated during model training. During model training, only the weights and parameters of the inserted CBAM attention module, as well as the weights and parameters of the global average pooling layer, fully connected layer, and output layer without pre-training weights and parameters, were updated. The weights were dynamically adjusted through the backpropagation algorithm to accelerate the model convergence speed and improve the model's fault diagnosis capability. The model performance was verified and monitored using a validation set to avoid the risk of overfitting. Step 6: Model Testing and Evaluation. The model trained in Step 5 is evaluated using the test set defined in Step 3 to test its ability to identify unknown faults. One or more evaluation metrics, such as accuracy, recall, precision, and F1 score, are used to assess the model's capabilities. If the evaluation results do not meet expectations, the model is adjusted or the process is repeated in Step 1 to collect more data. After processing in Steps 2 and 3, the model is retrained in Step 5 until the evaluation results meet expectations. Then, the model is used to diagnose the centrifugal fire pump to be tested.

[0010] As a preferred embodiment, in step 1, the first acceleration sensor is vertically arranged on the centrifugal fire pump housing, and the second acceleration sensor is vertically arranged on the drive motor housing. The multi-source signal synchronous acquisition refers to the synchronous acquisition of the drive motor vibration signal and the centrifugal fire pump vibration signal under the same time reference.

[0011] As a preferred embodiment, the operating status of the centrifugal fire pump includes a fault-free state, a bearing inner ring fault, a bearing outer ring fault, a loose foundation, a shaft imbalance fault, and a combined bearing fault.

[0012] As a preferred embodiment, the specific process of denoising the acquired vibration signal in step 1 is as follows: By selecting appropriate wavelet basis functions and decomposition levels, the cleaned and aligned signal is decomposed using the Maximum Overlap Discrete Wavelet Transform (MODWT) to obtain approximation coefficients and detail coefficients. A suitable threshold and threshold function are then selected to perform threshold filtering on the signal detail coefficients. Finally, the filtered detail coefficients and approximation coefficients are reconstructed using the inverse MODWT transform to obtain the denoised signal.

[0013] As a preferred embodiment, in step 2, the one-dimensional vibration signal is first segmented without overlap using a sliding window with a step size and a window length of 1024 data points. Then, the segmented one-dimensional vibration signal is transformed into a two-dimensional time-frequency graph using a generalized S-transform to simultaneously characterize the fault features in the time and frequency domains, thereby enhancing the expressive power of the sample features. The one-dimensional vibration signal includes centrifugal fire pump vibration signals and drive motor vibration signals under different fault labels. The two-dimensional time-frequency graph includes two-dimensional time-frequency graphs of centrifugal fire pump vibration signals and drive motor vibration signals under different fault labels.

[0014] As a preferred embodiment, in step 3, the feature fusion adopts a data-level feature fusion strategy, which splices together the two-dimensional time-frequency diagrams of the centrifugal fire pump vibration signal and the drive motor vibration signal obtained in step 2 with the same fault label to complete the time-frequency feature fusion. Compared with the prior art, the beneficial effects of this invention are as follows: 1. During the operation of a centrifugal fire pump, the vibration signal exhibits significant multi-source excitation superposition and nonlinear modulation characteristics due to the combined effects of multiple excitation sources such as impeller rotation, bearing wear, water pressure pulsation, and coupling misalignment. This causes the effective fault information in a single measuring point vibration signal to be masked by structural resonance and background noise during transmission, resulting in problems such as incomplete fault features contained in a single signal source, low effective information density, and ambiguous marginal judgment of fault features. This invention adopts a multi-source information fusion strategy, using different signal sources to reflect different physical mechanisms of the system. By synchronously acquiring the vibration signal from the drive motor side and the vibration signal from the centrifugal fire pump body, a cross-component vibration response information fusion framework is constructed, enabling fault excitation at the power source end and the actuator end to be integrated. Different propagation positions at the ends are co-expressed, enabling multi-node measurement of the vibration transmission path. This effectively enhances the separability and stability of fault features in the feature space. Simultaneously, through the complementary fusion of multi-source vibration information, the dependence of a single measurement point on the sensitivity of local structures is reduced. Different signal sources are used to reflect different physical mechanisms of the system, suppressing feature distortion caused by structural transmission differences. This solves the problem that a single signal can only characterize the local physical features of the system, resulting in incomplete fault features, low effective information density, and ambiguous judgment of fault feature boundaries. It improves the detection capability of early weak faults and the adaptability across operating conditions, thereby significantly improving the accuracy of fault identification and the reliability of system diagnosis. It has high engineering application value and promotion significance.

[0015] 2. To address the issue of traditional neural networks using static features for weight allocation during fault diagnosis, which leads to a lack of network adaptability and performance degradation, this invention introduces the CBAM attention module to improve the EfficientNet-b0-TL model. Utilizing the spatial and channel attention mechanisms of the CBAM module, the model adaptively and precisely optimizes features and allocates weights. This invention inserts the CBAM attention module after the last MBConv module in the EfficientNet-b0-TL model, rather than replacing the original SE module within MBConv. This improvement method does not alter the original network residual structure and depthwise separable convolutional computation path. This invention achieves joint spatial and channel enhancement of high-level semantic features. Compared with the improvement method of directly replacing the SE module, this invention maintains the stability of the original network structure and the compatibility of pre-trained weights. In the model improvement, the pre-trained parameters are directly transferred, which improves the model overfitting problem faced by small samples in actual engineering problems, improves the stability of model convergence, avoids excessive interference to low-level features, enhances the ability to preserve weak fault features, and the parameter growth of the model is controllable with small incremental computational complexity. It has the advantage of lightweight deployment and can generate the spatial attention map of the final layer of the model, ensuring the interpretability of the fault diagnosis process. Under the premise of ensuring lightweight, small sample size and stability, this invention achieves enhanced feature expression ability and weight allocation. 3. This invention first performs a generalized S-time-frequency transform on multi-source signals, mapping non-stationary signals into a two-dimensional time-frequency image with adaptive time-frequency resolution. While preserving impact characteristics and energy distribution characteristics, it achieves a unified expression of time-frequency structural information between different signal sources. Subsequently, a data-level fusion strategy is used to stitch together the multi-source two-dimensional time-frequency images, constructing a composite feature input tensor containing multi-physics information, thereby enhancing the correlation expression capability of cross-signal source information. Based on this, the fused time-frequency image is input into the EfficientNet-b0-TL model, and output in the last MBConv module. By embedding a CBAM attention module at the edge, high-level semantic features are adaptively weighted and modulated sequentially in the channel and spatial dimensions to enhance the feature response of key fault regions and sensitive frequency bands, while suppressing redundant background information. This technical solution achieves full expression and discrimination enhancement of multi-source time-frequency information without destroying the original lightweight network structure and residual mapping path, enhances the feature response capability of key fault regions, avoids excessive interference from low-level texture details, and improves the diagnostic accuracy, stability, and engineering interpretability of the model. It has significant technical effects such as small incremental computational complexity, strong structural compatibility, and high deployment adaptability. Attached Figure Description

[0016] Figure 1A schematic diagram of the improved EfficientNet-b0-TL model structure with the introduction of the CBAM attention mechanism; Figure 2 Flowchart for fault diagnosis of centrifugal fire pump; Figure 3(a) shows the vibration signal of the fire pump; Figure 3(b) shows the vibration signal of motors with different faults. Figure 4(a) shows the vibration time-frequency diagram of the drive motor with a loose foundation fault; Figure 4(b) shows the vibration time-frequency diagram of the fire pump with a loose foundation fault; Figure 4(c) shows the vibration time-frequency diagram of the drive motor without faults; Figure 4(d) shows the vibration time-frequency diagram of the fire pump without faults; Figure 4(e) shows the vibration time-frequency diagram of the drive motor with a shaft imbalance fault; Figure 4(f) shows the vibration time-frequency diagram of the fire pump with a shaft imbalance fault; Figure 4(g) shows the vibration time-frequency diagram of the drive motor with a combined bearing fault; Figure 4(h) shows the vibration time-frequency diagram of the fire pump with a combined bearing fault; Figure 4(i) shows the vibration time-frequency diagram of the drive motor with an inner bearing ring fault; Figure 4(j) shows the vibration time-frequency diagram of the fire pump with an inner bearing ring fault; Figure 4(k) shows the vibration time-frequency diagram of the drive motor with an outer bearing ring fault; Figure 4(l) shows the vibration time-frequency diagram of the fire pump with an outer bearing ring fault. Figure 5(a) shows the model confusion matrix; Figure 5(b) shows the model training loss and validation accuracy curves; Figure 5(c) shows the model data distribution. Detailed Implementation

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0018] This invention provides a fault diagnosis method for centrifugal fire pumps based on multi-source information fusion and an improved EfficientNet-b0-TL. This method can well meet the fault diagnosis requirements of centrifugal fire pumps. The specific steps are as follows: Step 1: Data Acquisition and Preprocessing. A first accelerometer is installed on the casing of the centrifugal fire pump, and a second accelerometer is installed on the casing of the drive motor. Vibration signals of the centrifugal fire pump under fault-free operation and under different fault conditions, as well as vibration signals of the drive motor, are acquired simultaneously. This step further includes data cleaning, alignment, and signal denoising to ensure data consistency, eliminate background noise, and improve data quality, as shown in Figures 3(a) and 3(b). Signal denoising uses the Maximum Overlap Discrete Wavelet Transform (MODWT) algorithm. f ( t )conduct j Layer decomposition, to obtain j Layer detail factor and approximation coefficients .

[0019] In the formula, For wavelet filters; It is a low-pass filter; l For filter length, l =0,1,2······ L ; For periodic extension boundary conditions; This is a scale-dependent time-shifted signal; By selecting a suitable threshold and threshold function, threshold filtering is performed on the signal detail coefficients. The filtered detail coefficients and approximate coefficients are then reconstructed using the maximum overlap discrete wavelet transform (MODWT) inverse transform to obtain the denoised signal. The formula for the maximum overlap discrete wavelet transform (MODWT) inverse transform is as follows: ;

[0020] In the formula, f ( t () represents the reconstructed signal after denoising. For scale-dependent detail coefficient translation; This is a shift of the scale-dependent approximation coefficients.

[0021] Step 2: Time-frequency feature extraction of multi-source signals. The denoised signal obtained in Step 1 is segmented without overlap using a sliding window with a step size and window length of 1024 data points. Then, the generalized S-transform is used to perform a time-frequency transformation on the segmented denoised signal, converting the one-dimensional vibration signal into a two-dimensional time-frequency graph. The time-frequency features of different signals are extracted. The two-dimensional time-frequency graphs are shown in Figures 4(a)-4(l). The formula for the generalized S-transform is: ;

[0022] In the formula, For frequency; It is a Gaussian window function; The position of the window function in the time domain; i is the imaginary unit; This is the window length adjustment coefficient; f ( t () represents the reconstructed signal after denoising.

[0023] Step 3: Feature Fusion and Sample Division. The two-dimensional time-frequency maps of centrifugal fire pump vibration signals and drive motor vibration signals with the same fault labels obtained in Step 2 are fused using a data-level fusion strategy. The fused samples are then randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0024] Step 4: Building and improving the fault diagnosis model. Load the pre-trained EfficientNet-b0 model, fine-tune the pre-trained EfficientNet-b0 model, replace the global average pooling layer, fully connected layer and output layer with pre-trained weights and parameters in the original pre-trained EfficientNet-b0 model with brand new global average pooling layer, fully connected layer and output layer without pre-trained weights and parameters, and adjust the number of output neurons in the output layer after the replacement to meet the fault diagnosis task, thereby building the transfer learning model EfficientNet-b0-TL; At this point, the transfer learning model EfficientNet-b0-TL contains seven MBConv modules (MBConv_1~MBConv_7) with pre-trained weights, one global average pooling layer without pre-trained weights and parameters, one fully connected layer without pre-trained weights and parameters, and one output layer without pre-trained weights and parameters. To address the problem of traditional neural networks using static features for weight allocation during fault diagnosis, which leads to a lack of network adaptability and performance degradation, the CBAM attention module is introduced to improve the EfficientNet-b0-TL model. An improved EfficientNet-b0-TL model is created by inserting a CBAM attention module after the 7th MBConv (MBConv_7) module and before the global average pooling layer. The improved EfficientNet-b0-TL model contains a total of 7 MBConv modules (MBConv_1~MBConv_7) with pre-trained weights, 1 CBAM attention module with weights and parameters but no pre-training, 1 global average pooling layer without pre-trained weights and parameters, 1 fully connected layer without pre-trained weights and parameters, and 1 output layer without pre-trained weights and parameters.

[0025] Step 5: Model Training. The improved EfficientNet-b0-TL model is trained using the training sample set partitioned in Step 3. To fully utilize the advantages of transfer learning models and achieve the transfer of pre-trained parameters, a strategy of freezing the weights and parameters of shallow networks is adopted during training. The parameters of the seven MBConv modules (MBConv_1~MBConv_7) with pre-trained weights are frozen, ensuring that the weights and parameters of these seven MBConv modules (MBConv_1~MBConv_7) are not updated during model training. During model training, only the weights and parameters of the inserted CBAM attention module, as well as the weights and parameters of the non-pre-trained global average pooling layer, fully connected layer, and output layer with weights and parameters, are updated. The weights are dynamically adjusted using the backpropagation algorithm to accelerate model convergence and improve model fault diagnosis capabilities. A validation set is used to verify and monitor model performance, avoiding the risk of overfitting.

[0026] Step 6: Model Testing and Evaluation. The model trained in Step 5 is evaluated using the test set defined in Step 3 to test its ability to identify unknown faults. One or more evaluation metrics, such as accuracy, recall, precision, and F1 score, are used to assess the model's capabilities. If the evaluation results do not meet expectations, the model is adjusted (those skilled in the art can adjust it based on experience) or the process is repeated in Step 1 to collect more data, which is then processed in Steps 2 and 3. The model is then retrained in Step 5 until the evaluation results meet expectations. The model is then used to diagnose the centrifugal fire pump to be tested.

[0027] This invention is verified using specially constructed centrifugal fire pump fault simulation experimental data. This dataset contains fault-free state data of centrifugal fire pumps and common fault data (bearing inner ring fault, bearing outer ring fault, foundation loosening, shaft imbalance fault, bearing compound fault). All faults are artificially generated to simulate the vibration of centrifugal fire pumps and drive motors caused by different faults.

[0028] The experiment used the XY-U9004 signal acquisition card produced by Yangzhou Xiyuan Electronic Technology Co., Ltd. as the data acquisition device. This acquisition card supports 4-channel synchronous acquisition with a maximum sampling frequency of 128KHz / channel. The accelerometer used is the AD26D100 / T08IEPE single-axis accelerometer. During the data acquisition process, the drive motor speed was set to 2000r / min and the signal sampling frequency was set to 16KHz through the frequency converter. The vibration signals of the centrifugal fire pump and the drive motor under different fault conditions were acquired synchronously. Then, the signal processing and sample division were performed according to the steps described in steps 1 to 3. In this case, the number of samples for a single fault was 600. After being randomly divided into training set, validation set and test set in a 6:2:2 ratio, the number of samples in the training set, validation set and test set were 360, 120 and 120 respectively.

[0029] The obtained data is used to train and evaluate the improved EfficientNet-b0-TL model according to steps 5 and 6. During training, the initial learning rate is 10. -4 With a batch size of 32 and 20 iterations, the final model achieved 100% fault classification accuracy in the experimental data, as shown in Figures 5(a)-5(c).

[0030] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Many modifications and variations can be made based on the content of this specification. The selection and detailed description of these embodiments are intended to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. They should not be considered as limitations on the invention.

Claims

1. A fault diagnosis method for centrifugal fire pumps based on multi-source information fusion and an improved EfficientNet-b0-TL model, comprising the following steps: Step 1: Data Acquisition and Preprocessing A first accelerometer is installed on the casing of the centrifugal fire pump, and a second accelerometer is installed on the casing of the drive motor of the centrifugal fire pump. The vibration signals of the centrifugal fire pump and the drive motor are collected simultaneously under the fault-free working state and under various common fault states. The collected vibration signals are cleaned, aligned and denoised. Step 2: Time-Frequency Feature Extraction of Multi-Source Signals The denoised signal obtained in step 1 is segmented into a non-overlapping sliding window. Then, the generalized S-transform is used to perform time-frequency transformation on the segmented denoised signal, transforming the one-dimensional vibration signal into a two-dimensional time-frequency graph, and extracting the time-frequency features of different signals. Step 3: Feature Fusion and Sample Splitting The two-dimensional time-frequency graphs obtained in step 2 are subjected to feature fusion using a data-level fusion strategy. The fused samples are then randomly divided into training set, validation set, and test set according to a certain ratio. Step 4: Building and improving the fault diagnosis model Load the pre-trained EfficientNet-b0 model, and replace the global average pooling layer, fully connected layer, and output layer with pre-trained weights and parameters in the original pre-trained EfficientNet-b0 model with completely new global average pooling layer, fully connected layer, and output layer without pre-trained weights and parameters. Adjust the number of output neurons in the output layer after the replacement to meet the fault diagnosis task, thus building the transfer learning model EfficientNet-b0-TL. Then, insert a CBAM attention module after the 7th MBConv module and before the global average pooling layer in the EfficientNet-b0-TL model to improve the EfficientNet-b0-TL model. The improved EfficientNet-b0-TL model contains a total of 7 MBConv modules with pre-trained weights, 1 CBAM attention module with weights and parameters but no pre-training, 1 global average pooling layer without pre-trained weights and parameters, 1 fully connected layer without pre-trained weights and parameters, and 1 output layer without pre-trained weights and parameters. Step 5: Model Training The improved EfficientNet-b0-TL model was trained using the training sample set partitioned in step 3. During training, a strategy of freezing the weights and parameters of shallow networks was adopted. The parameters of the seven MBConv modules with pre-trained weights were frozen, so that the weights and parameters of the seven MBConv modules with pre-trained weights were not updated during model training. During model training, only the weights and parameters of the inserted CBAM attention module, as well as the weights and parameters of the global average pooling layer, fully connected layer, and output layer without pre-trained weights and parameters, were updated. The weights were dynamically adjusted through the backpropagation algorithm to accelerate the model convergence speed and improve the model's fault diagnosis capability. At the same time, a validation set was used to verify and monitor the model performance to avoid the risk of overfitting. Step 6: Model Testing and Evaluation The model trained in step 5 is evaluated using the test set partitioned in step 3 to test its ability to identify unknown faults. One or more evaluation metrics, such as accuracy, recall, precision, and F1 score, are used to evaluate the model's capabilities. If the evaluation results do not meet the expectations, the model is adjusted or the process is repeated in step 1 to collect more data. After processing in steps 2 and 3, the model is retrained in step 5 until the evaluation results meet the expectations. Then, the model is used to diagnose the centrifugal fire pump to be tested.

2. The centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved EfficientNet-b0-TL as described in claim 1, characterized in that: In step 1, the first acceleration sensor is vertically arranged on the centrifugal fire pump housing, and the second acceleration sensor is vertically arranged on the drive motor housing. The multi-source signal synchronous acquisition refers to the synchronous acquisition of the drive motor vibration signal and the centrifugal fire pump vibration signal under the same time reference.

3. The centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved EfficientNet-b0-TL as described in claim 1, characterized in that: The operating status of the centrifugal fire pump includes no fault, bearing inner ring fault, bearing outer ring fault, foundation loosening, shaft imbalance fault, and bearing combined fault.

4. The centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved EfficientNet-b0-TL as described in claim 1, characterized in that: The specific process of denoising the acquired vibration signal in step 1 is as follows: By selecting appropriate wavelet basis functions and decomposition levels, the cleaned and aligned signal is decomposed using the Maximum Overlap Discrete Wavelet Transform (MODWT) to obtain approximation coefficients and detail coefficients. A suitable threshold and threshold function are then selected to perform threshold filtering on the signal detail coefficients. Finally, the filtered detail coefficients and approximation coefficients are reconstructed using the inverse MODWT transform to obtain the denoised signal.

5. The centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved Efficient-b0-TL as described in claim 1, characterized in that, In step 2, the one-dimensional vibration signal is first segmented without overlap using a sliding window with a step size and a window length of 1024 data points. Then, the segmented one-dimensional vibration signal is transformed into a two-dimensional time-frequency graph using a generalized S-transform to simultaneously characterize the fault features in the time and frequency domains, thereby enhancing the expressive power of the sample features. The one-dimensional vibration signal includes centrifugal fire pump vibration signals and drive motor vibration signals under different fault labels. The two-dimensional time-frequency graph includes two-dimensional time-frequency graphs of centrifugal fire pump vibration signals and drive motor vibration signals under different fault labels.

6. The centrifugal fire pump fault diagnosis method based on multi-source information fusion and improved EfficientNet-b0-TL as described in claim 1, characterized in that: In step 3, the feature fusion adopts a data-level feature fusion strategy, which splices the two-dimensional time-frequency diagrams of the centrifugal fire pump vibration signal and the drive motor vibration signal obtained in step 2 with the same fault label to complete the time-frequency feature fusion.

Images