Centrifugal pump fault diagnosis method based on YOLOv8

Abstract

Aiming at the problems of feature extraction bottleneck and limited identification precision in the field of centrifugal pump fault diagnosis, the invention provides an innovative solution. The intelligent diagnosis model based on the YOLO architecture is constructed, and the core innovation point of the intelligent diagnosis model lies in the innovation of a signal processing method. Through the introduction of a continuous wavelet transform (CWT) algorithm, a one-dimensional vibration time sequence signal is mapped into a two-dimensional time-frequency domain representation, and the method effectively overcomes the dependence of a traditional diagnosis normal form on expert experience and artificial feature engineering. In the model verification link, the standard centrifugal pump fault data set is used for systematic experiments. The data set covers typical scenes such as normal working conditions, single faults and composite faults. Mechanism research shows that the provided model shows excellent classification performance, experimental data shows that the average recognition accuracy can reach 86%, and the index is remarkably superior to that of a traditional threshold value setting method. Notably, the end-to-end feature learning ability of the model has important application value in the field of rotary machine diagnosis.

Description

Technical Field

[0001] The invention belongs to the field of rotating machinery fault diagnosis and aims to improve the accuracy of centrifugal pump fault diagnosis. Background Art

[0002] A revolutionary breakthrough in object detection, the YOLO (You Only Look Once) algorithm is widely recognized for its ability to locate and classify objects in a single forward pass. This architectural design significantly improves detection efficiency. In its latest iteration, YOLOv8, the superiority of its deep learning model is demonstrated through an improved convolutional neural network architecture, particularly its ability to efficiently extract key features. Notably, the algorithm demonstrates unique advantages in multi-label processing, primarily due to its parameter-optimized non-maximum suppression strategy, which demonstrates excellent discrimination capabilities in scenarios with overlapping objects. However, it is important to note that while YOLOv8 has achieved significant results in general object detection tasks, its application in rotating machinery fault diagnosis remains a research gap. Specifically, for typical rotating machinery fault diagnosis tasks such as centrifugal pumps, there is no standardized solution based on the YOLOv8 framework, nor is there a unified architecture recognized by both academia and industry. This situation urgently needs to be addressed through systematic experimental research and engineering validation. Summary of the Invention

[0003] To address the challenges of fault feature extraction and insufficient fault identification accuracy in centrifugal pump fault diagnosis, this paper proposes a fault diagnosis model based on YOLO. It introduces continuous wavelet transform to convert one-dimensional vibration signals into two-dimensional time-frequency images, breaking through the traditional diagnostic method's reliance on manual feature extraction. The model is then verified on a centrifugal pump fault dataset.

[0004] The technology of the present invention is a centrifugal pump fault diagnosis method based on YOLOv8, which includes:

[0005] Step 1: Data Preprocessing and Noise Reduction: Preprocess the vibration signal. This step includes data cleaning, normalization, and noise reduction to obtain an optimized vibration signal and ensure data quality and consistency.

[0006] Step 2: Feature extraction and transformation: The continuous wavelet transform (CWT) feature extraction method is used on the preprocessed vibration signal to generate two-dimensional image features suitable for deep learning models.

[0007] The basic principle of CWT is: for function If satisfied but It can be written as mother wavelet. A series of scale translation transformations are performed to obtain a continuous wavelet function, which is called an analysis wavelet. The transformation formula is:

[0008]

[0009] In the formula, u represents the translation factor; v represents the scale factor. When v>1, it stretches in the horizontal direction, and when v<1, it compresses in the horizontal direction. u represents the translation parameter and is an arbitrary real number. Both u and v are continuous variables, so it is called a continuous wavelet. In order to keep the energy unchanged after the scaling transformation, it is necessary to multiply the scale factor beforehand. Right now For signal x(t)∈L 2 (R) Perform continuous wavelet transform:

[0010]

[0011] Where, Representative function A complex function whose inverse transform satisfies Specifically:

[0012]

[0013] Only when the values ​​of the mother wavelet function outside the window function are all 0 and satisfy Only in this way can we ensure that the wavelet transform can accurately construct the original signal and have the corresponding inverse transform.

[0014] Step 3: Sample set preparation: The two-dimensional image obtained by the CWT method is resized to a suitable size and saved as a sample set.

[0015] Step 4: Dataset division: The dataset needs to be randomly divided into three parts according to a predetermined ratio: training set, test set, and validation set.

[0016] Step 5: Model establishment and initialization: Establish a YOLOv8 neural network fault diagnosis model and set the initial values ​​of its network parameters.

[0017] Step 6: Model training. The training set is batched and fed into the model for training. During training, the model adjusts weights through backpropagation to minimize the loss function and improve classification performance. A validation set is used to monitor model performance and prevent overfitting.

[0018] Step 7: Model Testing and Evaluation. The trained model is evaluated using the test set to verify its performance in distinguishing between various fault types. Classification performance is evaluated using accuracy, precision, recall, and F1 score.

[0019] The beneficial effects of the present invention are:

[0020] 1) A fault diagnosis model based on the YOLO architecture demonstrates significant advantages in centrifugal pump condition monitoring. Compared to traditional diagnostic methods that rely on manual feature engineering, the model's accuracy is significantly improved, primarily due to the deep convolutional neural network's ability to automatically extract fault features. Through a single-stage target detection mechanism, the model's rapid processing capabilities during training and inference have been demonstrated, and a single forward propagation mechanism has been shown to reduce response latency, thereby meeting the stringent requirements of real-time fault diagnosis in industrial scenarios.

[0021] 2) In the signal preprocessing link, the introduction of continuous wavelet transform is used as the core preprocessing step. The one-dimensional vibration signal is mapped into a two-dimensional time-frequency domain representation. This nonlinear transformation process breaks through the traditional diagnostic paradigm's reliance on expert experience. The fault modes implicit in the time-frequency image are adaptively extracted, effectively solving the problem of insufficient separation of fault features under complex working conditions. Through the matrix expression of the time-frequency energy distribution, multi-scale fault features can be explicitly presented, which provides reliable input for the robust classification of subsequent deep learning models. It is particularly noteworthy that the feature decoupling capability of this method has been verified in the field of rotating machinery. Its synergy with the YOLO architecture has opened up a new path for building end-to-end intelligent diagnostic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the YOLOv8 network model structure;

[0023] Figure 2 It is a centrifugal pump fault simulation test device;

[0024] Figure 3 is a schematic diagram of bearing failure mode;

[0025] Figure 4 is a schematic diagram of the impeller failure mode;

[0026] Figure 5 This is the result obtained by the present invention using the data set. DETAILED DESCRIPTION

[0027] This paper constructs a specialized centrifugal pump fault simulation test bench, using the IHG25-160 centrifugal pump as the research object. Its rated operating parameters are set to a rotational speed of 1200 r / min to simulate the coupling effect of mechanical loads and fluid dynamics in industrial sites. In the data acquisition process, the sampling frequency is set to 50 kHz, and 512,000 data points are collected per sampling cycle. This configuration not only ensures the time series integrity of the high-frequency vibration signal, but also meets the frequency domain resolution and statistical analysis accuracy requirements. The experimental failure mode is set to a dual operating condition including single faults and combined faults, which provides data support for the verification of the model's generalization ability.

[0028] The experimental process is carried out according to a seven-stage structure:

[0029] Step 1: Data preprocessing and noise reduction. The raw vibration signal undergoes a preprocessing process involving data cleaning, numerical normalization, and noise reduction to eliminate environmental interference and improve the signal-to-noise ratio.

[0030] Step 2: Feature extraction and conversion: The preprocessed vibration signal is processed using the Continuous Wavelet Transform (CWT) algorithm, and the implementation of time-frequency joint analysis converts the one-dimensional time series signal into a two-dimensional grayscale image feature.

[0031] Step 3: Sample set preparation: The time-frequency images converted by CWT need to be size-normalized and then archived as a standardized sample library.

[0032] Step 4: Dataset Partitioning. Following the machine learning paradigm, the dataset is randomly divided into three mutually exclusive subsets: a model training set, a validation set, and a test set. The proportions are determined based on the cross-validation criterion.

[0033] Step 5: Model construction and initialization. A deep learning architecture based on the YOLOv8 framework is constructed, the network parameters are initialized, and the convolution kernel size and activation function type are determined based on prior experiments.

[0034] Step 6: Model training. The training set data is batched into the network for iterative training. Weights are updated using the backpropagation algorithm, and the gradient descent of the loss function is monitored. The validation set is used for hyperparameter tuning, and early stopping is used to prevent overfitting.

[0035] Step 7: Model Testing and Evaluation. The final model performance is systematically evaluated on the test set. Classification performance is measured using quantitative metrics, including classification accuracy, precision, recall, and F1 score. These metrics collectively characterize the model's ability to discriminate complex fault modes.

[0036] The present invention is verified in experimental data using a specially constructed centrifugal pump fault simulation test data set, which contains three types of samples: normal operating conditions, single faults, and compound faults. Figure 2 As shown in the figure, the experimental platform consists of core components such as computers, DC-24 multi-function data collectors, consoles, water tanks, centrifugal pumps, etc., and each subsystem is coordinated and controlled through a distributed architecture.

[0037] In the bearing fault simulation module, deep groove ball bearings are used as research objects, and their fault types are realized through artificial defect implantation technology. Figure 3 As shown in the figure, the wire cutting process is used to create standard defects on the inner and outer rings of the bearing, forming three experimental states: normal, inner ring fault, and outer ring fault. The working condition switching is achieved through physical replacement. The impeller fault simulation uses a closed impeller structure, such as Figure 4 As shown in the figure, the mass imbalance defect is constructed by mechanical drilling process, and the blade deformation is simulated by plastic deformation process, finally forming three experimental conditions: normal impeller, unbalanced impeller and deformed impeller.

[0038] The data acquisition system is equipped with a VALENIAN DC-24 8-channel data acquisition unit. Its hardware architecture features a modular design, integrated 24-bit A / D converters, and 110dB inter-channel isolation. Each independent channel supports a sampling rate capped at 100kHz, enabling real-time data transmission via Gigabit Ethernet. This configuration meets the required standards for rotating machinery vibration monitoring. The data acquisition unit is connected to sensors to acquire vibration signals from the centrifugal pump under various operating conditions. This vibration data is crucial for analyzing pump health, identifying potential faults, and performing predictive maintenance, providing a vital information foundation for fault diagnosis and performance evaluation.

[0039] 500 groups of centrifugal pump vibration signals were extracted under various fault conditions, and the sample length of each group was 2048. After preprocessing using a noise reduction method, the signal was converted into a two-dimensional time-frequency graph and used as the input of the network for training and testing. The sample length was set to 2048×k, where k is the decomposition scale and the scale sequence length is 224. These generated two-dimensional images or time-frequency graphs were randomly divided into three different data sets: training set, test set and validation set, and each category contained 400 training data, 50 test data and 50 validation data. During the training process of the YOLO model, in order to optimize the model performance and ensure that it can achieve higher accuracy and efficiency in target detection tasks, the present invention defines specific hyperparameters during the training process, sets the training data set to 200 rounds, the number of samples to 500, and adjusts the image size of the input model to 224×224 pixels.

[0040] Figure 5 The results obtained by the present invention using the data set are given, and the accuracy of the model is 86%.

Claims

1. A centrifugal pump fault diagnosis method based on YOLOv8, the method comprising the following steps: Step 1: Data Preprocessing and Noise Reduction: Preprocess the vibration signal. This step includes data cleaning, normalization, and noise reduction to obtain an optimized vibration signal and ensure data quality and consistency. Step 2: Feature extraction and transformation: The continuous wavelet transform (CWT) feature extraction method is used on the preprocessed vibration signal to generate two-dimensional image features suitable for deep learning models. Step 3: Sample set preparation: The two-dimensional image obtained by the CWT method is resized to a suitable size and saved as a sample set. Step 4: Dataset division: The dataset needs to be randomly divided into three parts according to a predetermined ratio: training set, test set, and validation set. Step 5: Model establishment and initialization: Establish a YOLOv8 neural network fault diagnosis model and set the initial values ​​of its network parameters. Step 6: Model training. The training set is batched and fed into the model for training. During training, the model adjusts weights through backpropagation to minimize the loss function and improve classification performance. A validation set is used to monitor model performance and prevent overfitting. Step 7: Model Testing and Evaluation. The trained model is evaluated using the test set to verify its performance in distinguishing between various fault types. Classification performance is evaluated using accuracy, precision, recall, and F1 score.

2. The centrifugal pump fault diagnosis method based on YOLOv8 according to claim 1 It is characterized by Innovative design of the signal conversion method in step 2. The continuous wavelet transform (CWT) algorithm is applied to vibration signal processing. It reconstructs one-dimensional time series signals into two-dimensional grayscale images through joint time-frequency analysis. This nonlinear mapping process effectively enhances the completeness of feature representation, enabling the deep learning model to autonomously extract multi-scale fault features from the time-frequency energy distribution, thereby improving pattern recognition capabilities under complex operating conditions.

3. The centrifugal pump fault diagnosis method based on YOLOv8 according to claim 1 It is characterized by The adaptive nature of the model architecture in step 5. The constructed YOLOv8 neural network diagnostic model features reconfigurable parameters, and its backbone network depth and feature pyramid structure can be dynamically adjusted based on the spectral characteristics of the diagnostic object. By optimizing the anchor frame size configuration strategy and loss function weight distribution mechanism, the model significantly improves the accuracy of identifying complex centrifugal pump faults while maintaining real-time detection speed. This flexible architecture design enables it to adapt to industrial detection scenarios with different sensor configurations.

Images