Impact type fault interpretable detection method for ship key equipment based on wavelet scattering and time-frequency feature enhancement

Abstract

The invention discloses a ship key equipment impact type fault interpretable detection method based on wavelet scattering and time-frequency feature enhancement, and belongs to the field of ship equipment fault detection. According to the method, a vibration signal of ship key equipment is acquired through a sensor, and the vibration signal is converted into a time-frequency image with cross-domain consistency by adopting improved wavelet scattering transform; adaptive enhancement processing based on statistical features is carried out on the time-frequency image to realize noise suppression and fault feature enhancement; and inputting the processed time-frequency image into a pre-training model embedded with class activation mapping to complete fault identification and decision process visualization. Aiming at the impact type fault detection requirement of the ship key equipment, the method solves the problems that a traditional method is weak in generalization ability and cannot explain the diagnosis result, has the advantages of being high in detection precision, high in adaptability and traceable in diagnosis result, and can be widely applied to bearing fault detection of key equipment such as a ship main lubricating oil pump, an oil separator and a turbine.

Description

Technical Field

[0001] This invention belongs to the field of ship equipment fault detection, specifically relating to an interpretable detection method for impact-type faults in key ship equipment based on wavelet scattering and time-frequency feature enhancement, applicable to bearing fault detection in key equipment such as main lubricating oil pumps, oil separators, and turbines. Background Technology

[0002] Critical ship equipment is the core of ensuring safe navigation, and its operational status directly affects the ship's navigational stability, operational efficiency, and even the safety of the crew. Under marine conditions, critical ship equipment is subjected to multiple harsh conditions such as variable loads and salt spray corrosion over long periods, making it prone to impact-type failures such as bearing wear, gear tooth breakage, and blade damage. These failures are characterized by their sudden onset; if not detected and diagnosed in a timely manner, they can lead to serious accidents such as equipment downtime and loss of ship control, causing enormous economic losses and safety risks.

[0003] Currently, fault detection in critical ship equipment primarily employs traditional signal processing methods and conventional intelligent diagnostic techniques. Traditional methods, such as Fourier transform and ordinary wavelet analysis, struggle to effectively handle the non-stationarity and strong noise interference of ship vibration signals, and have limited ability to extract features from impact-type faults. Conventional intelligent diagnostic techniques, such as support vector machines and basic convolutional neural networks, while achieving some success in laboratory environments, suffer from insufficient model generalization ability in real-world scenarios due to significant differences in data distribution across different ships and operating conditions, making it difficult to adapt to the detection needs across different equipment and operating conditions. Furthermore, existing intelligent diagnostic methods are mostly black-box models, lacking interpretability in their diagnostic results. Ship maintenance personnel not only need to know if equipment is faulty, but also need to understand the location and characteristic basis of the fault to formulate accurate maintenance plans. Black-box models cannot provide visual support for the decision-making process, leading to low trust in diagnostic results among maintenance personnel and limiting the practical application of intelligent detection technology in the shipbuilding field. In addition, existing methods often neglect the need for collaborative identification of low-frequency steady-state components and high-frequency impact components in the ship environment, making it difficult to comprehensively capture the essential characteristics of impact-type faults.

[0004] Therefore, developing a method for detecting impact-type faults in critical ship equipment that combines high generalization ability, high detection accuracy, and interpretability is of great significance for improving ship navigation safety and reducing operation and maintenance costs. Summary of the Invention

[0005] The purpose of this invention is to address the problems of weak generalization ability, uninterpretable diagnostic results, and poor anti-interference ability of existing methods for detecting impact-type faults in critical ship equipment, and to provide an interpretable detection method based on wavelet scattering and time-frequency feature enhancement.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: An interpretable detection method for impact-type faults in key ship equipment based on wavelet scattering and time-frequency feature enhancement, comprising the following steps:

[0007] Step 1: Collect raw vibration signals during equipment operation by using vibration sensors installed at designated locations on key ship equipment;

[0008] Step 2: Considering strong noise interference and the non-stationary characteristics of the signal, an improved wavelet scattering transform is used to process the original vibration signal. Through multi-scale wavelet decomposition and reconstruction, the one-dimensional vibration signal is converted into a two-dimensional time-frequency image, realizing the visualization of fault characteristics.

[0009] Step 3: Adaptive Time-Frequency Feature Enhancement. The time-frequency image obtained in Step 2 undergoes feature enhancement processing to improve the recognizability of fault features. This includes two operations: regularization and adaptive nonlinear enhancement.

[0010] Step 4: Training the interpretable fault detection model. This includes building a VGG16-based feature extraction network and embedding class activation maps into the fault detection model, and training the model using training samples.

[0011] Step 5: Fault Detection and Result Visualization. The vibration signals of the key equipment on the ship to be detected are converted to time-frequency signals using the method in Step 1, and then processed by feature enhancement in Step 2 before being input into the trained fault detection model. The model outputs the fault category probability and the corresponding class activation map.

[0012] Furthermore, in step 1, the vibration sensor is a piezoelectric accelerometer, and the sampling frequency is determined according to the operating speed of the ship's key equipment, ranging from 12kHz to 48kHz, to ensure complete capture of the high-frequency components of impact-type faults.

[0013] Furthermore, when acquiring the time-frequency image described in step 2, a complex Morlet wavelet is selected as the mother wavelet to avoid phase artifacts caused by real wavelets. Simultaneously, the filter bank structure is optimized, employing a power-law distributed scaling factor in the high-frequency region to ensure the accuracy of impulse feature capture, and an equally spaced scaling factor in the low-frequency region to fully preserve the steady-state component information of the equipment operation. The definition of the Morlet wavelet is as follows:

[0014]

[0015] in, For complex Morlet mother wavelet, For bandwidth parameters, For the center frequency, is the wavelet scaling factor, Q is the number of wavelets per octave, J is the maximum scale parameter, T is the sample length, and k and l are integers;

[0016] The required filter bank is constructed as shown in the following equation:

[0017]

[0018] The original vibration signal is decomposed and reconstructed using multi-scale wavelet decomposition, and the modulus of the wavelet coefficients at each scale is calculated to obtain the time-frequency image.

[0019]

[0020] Where x(t) is the original vibration signal, For the stretched wavelet basis, This indicates the inner product operation.

[0021] Furthermore, in step 3, frequency regularization is used as a preprocessing method, and its definition is as follows:

[0022]

[0023] In the formula Represents the regularization factor. Indicates the bandwidth of the wavelet. This represents the bandwidth of the mother wavelet. Each k corresponds to a scale. .

[0024] Furthermore, in step 3, the adaptive nonlinear time-frequency enhancement is defined as follows:

[0025]

[0026] In the formula, x represents the energy amplitude of the time-frequency representation before mapping. An exponential base that is adaptively determined by statistical characteristics. It is a small positive value close to 0. Bias represents a steepness threshold, which physically means that numbers above the threshold tend to approach 1, and numbers below the threshold tend to approach 0.

[0027] Furthermore, the main body of the interpretable fault detection model constructed in step 4 includes a feature extraction module and a decision visualization module. The former uses the first 5 convolutional blocks of a pre-trained VGG16 model, freezes the parameters of the first 2 convolutional blocks, and allows the remaining parameters to be trained. The input size of the fully connected layer of the model is 256, and the output layer is adaptively designed according to the specific diagnostic task. The specific structural parameters are shown in the table below.

[0028] Table 1: Specific Model Parameters for Feature Extraction Module

[0029]

[0030] The decision visualization module includes a global average pooling layer, a fully connected layer, a softmax layer, and a CAM generation unit. The CAM generation unit obtains the class activation map by weighted summation of the feature maps of the last convolutional layer. The process is defined as follows:

[0031]

[0032] W represents the weights corresponding to the feature map, and c represents the output class of the model. In the formula, k represents the index of the feature map, and CAM... c It directly represents the spatial distribution of the basis for the model's predictions.

[0033] In step 4, the model training process uses the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. The cross-entropy loss function is used to measure the difference between the model's predicted values ​​and the true labels.

[0034] 1.3.1 Beneficial effects of the present invention:

[0035] Strong anti-interference capability: The improved wavelet scattering transform combined with complex Morlet wavelet and optimized filter bank can effectively suppress strong noise interference in the environment, fully preserve the fault impact characteristics and equipment steady-state components, and improve the detection robustness in complex environments.

[0036] Excellent generalization ability: Adaptive enhancement processing based on statistical features reduces the differences in data distribution under different ships and operating conditions, making the fault features obtained by model training consistent across domains, without the need to retrain the model for specific ships or equipment.

[0037] The diagnostic results are interpretable: By using Class Activation Mapping (CAM) technology, the black-box decision-making process of the model is visualized, allowing maintenance personnel to intuitively understand the basis for their decisions, increasing their trust in the diagnostic results, and facilitating the development of accurate maintenance plans. Attached Figure Description

[0038] Figure 1 This is a flowchart of an invention-based method for detecting interpretable impact-type faults.

[0039] Figure 2 This is a schematic diagram of the depth model used in the invention.

[0040] Figure 3It is a visualization of the filter bank constructed by wavelet scattering and a time-domain diagram of the Morlet wavelet.

[0041] Figure 4 This is a diagram showing the principle and effect of a time-frequency enhancement method based on statistical characteristics.

[0042] Figure 5 These are feature enhancement effect diagrams for different data samples.

[0043] Figure 6 It is a heatmap that can explain the fault diagnosis results. Figure 7 It is a time-frequency representation of actual bearing fault samples and a heatmap of diagnostic results. Detailed Implementation

[0044] To enable the present invention to obtain the time-frequency representation from the original vibration signal through a constructed filter bank, and to perform time-frequency enhancement preprocessing based on frequency regularization and statistical characteristics, and then input it into the proposed interpretable fault detection model for training and deployment, the present invention adopts the following example.

[0045] As shown in Figure 1, the specific implementation steps of the detection method of the present invention are as follows:

[0046] Step 1: Vibration signal acquisition and time-frequency conversion. Piezoelectric accelerometers were installed at the input and output bearing housings of the ship's main lubricating oil pump motor. The sampling frequency was set to 48kHz. Vibration signals were acquired under normal motor operation, bearing inner ring fault, and bearing outer ring fault conditions. Each sample had 4096 data points.

[0047] The discrete sequence obtained by sampling the original signal x(t) is as follows:

[0048] x(i): 1≤i≤N

[0049] Where N is the length of the original sequence xi = {x1, x2, ..., xN}, which is 4096;

[0050] Obtain the time-frequency representation of the discrete sequence. A complex Morlet wavelet is selected as the mother wavelet to avoid phase artifacts introduced by real wavelets. Simultaneously, the filter bank structure is optimized, employing a power-law distributed scaling factor in the high-frequency region to ensure the accuracy of impulse feature capture, and an equally spaced scaling factor in the low-frequency region to fully preserve the steady-state components of the equipment operation. The definition of the Morlet wavelet is as follows:

[0051]

[0052] in, For complex Morlet mother wavelet, For bandwidth parameters, For the center frequency, is the wavelet scaling factor, Q is the number of wavelets per octave, J is the maximum scale parameter, T is the sample length, and k and l are integers;

[0053] The required filter bank is constructed as shown in the following equation:

[0054]

[0055] The frequency domain representation of the wavelet waveforms used and the constructed filter banks is as follows: Figure 3 As shown.

[0056] The original vibration signal is decomposed and reconstructed using multi-scale wavelet decomposition, and the modulus of the wavelet coefficients at each scale is calculated to obtain the time-frequency image.

[0057]

[0058] Where x(t) is the original vibration signal, For the stretched wavelet basis, This indicates the inner product operation.

[0059] Step 2: Adaptive enhancement of time-frequency features. First, frequency regularization denoising is performed on the acquired time-frequency image, and the bandwidth of wavelet bases at each scale is calculated. Using its reciprocal as a normalization factor, a weighted average is applied to each pixel of the time-frequency image, as defined below:

[0060]

[0061] In the formula Represents the regularization factor. Indicates the bandwidth of the wavelet. This represents the bandwidth of the mother wavelet. Each k corresponds to a scale. .

[0062] Subsequently, nonlinear feature enhancement is performed. The energy distribution of the time-frequency image is statistically analyzed, and an adaptive mapping function is constructed to process the time-frequency image, resulting in the enhanced fault feature image. The definition of adaptive nonlinear time-frequency enhancement is as follows:

[0063]

[0064] In the formula, x represents the energy amplitude of the time-frequency representation before mapping. An exponential base that is adaptively determined by statistical characteristics. It is a small positive value close to 0. bias represents a steepness threshold, physically meaning that numbers exceeding this threshold tend to approach 1, and numbers below this threshold tend to approach 0. Setting ε = 10−6, bias is 95% of the peak energy of this time-frequency image.

[0065] The constructed adaptive nonlinear mapping function is as follows Figure 4 As shown. Figure 5 This demonstrates the effect before and after adaptive noise reduction. Figure 4 and Figure 5 It is evident that the proposed adaptive time-frequency enhancement method significantly improves the discriminability of fault features and also enhances the consistency of cross-domain fault features for similar faults.

[0066] Step 3: Training the Interpretable Fault Detection Model. An embedded CAM fault detection model is constructed. The feature extraction module uses the first 5 convolutional blocks of a pre-trained VGG16 model, freezing the parameters of the first 2 convolutional blocks, while the remaining parameters are trainable. The fully connected layers of the model are set to 256 neurons, and the output layer is set to 2 neurons (corresponding to 2 device states). Specific parameters are shown in the table below:

[0067] Table 1: Specific Model Parameters for Feature Extraction Module

[0068]

[0069] Figure 2 The overall architecture of the constructed model is shown.

[0070] The decision visualization module includes a global average pooling layer, a fully connected layer, a softmax layer, and a CAM generation unit. The CAM generation unit obtains the class activation map by weighted summation of the feature maps of the last convolutional layer. The process is defined as follows:

[0071]

[0072] W represents the weights corresponding to the feature map, and c represents the output class of the model. In the formula, k represents the index of the feature map, and CAM... c It directly represents the spatial distribution of the basis for the model's predictions.

[0073] Step 4: Model Training. During training, 80% of the samples are used as the training set and 20% as the validation set. The Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 100 training epochs. During training, the model optimizes its parameters using the cross-entropy loss function and generates a class activation map to monitor the rationality of the model's decisions in real time. After training is complete, the model parameters with the highest accuracy on the validation set are saved.

[0074] Step 5: Model Testing. The trained model is tested using data from an actual ship's main lubricating oil pump motor drive under bearing failure conditions. The test samples undergo the preprocessing operations described in Steps 1 and 2, are then input into the trained model to obtain prediction results.

[0075] Table 2. Cross-domain fault detection results. FPR: Precision, Recall: Recall.

[0076]

[0077] Experimental results show that, considering the differences in data distribution between the test and training data, this fault detection task is challenging. The proposed method achieves the best detection performance, demonstrating its strong generalization ability to some extent. To further verify the interpretability of the proposed method, random test samples were obtained, and attention heatmaps of the diagnostic results were plotted. It is evident that the proposed method can accurately capture repetitive transient impact phenomena and make accurate judgments.

Claims

1. A method for interpretable detection of impact-type faults in key ship equipment based on wavelet scattering and time-frequency feature enhancement, characterized in that, Includes the following steps: Step 1: Collect raw vibration signals of key ship equipment using vibration sensors; Step 2: The original vibration signal is converted to time and frequency using an improved wavelet scattering transform to obtain a time-frequency image; Step 3: Perform adaptive enhancement processing based on statistical features on the time-frequency image, including frequency regularization denoising and nonlinear feature enhancement, to obtain the enhanced fault feature image; Step 4: Construct an Embedded Class Activation Map (CAM) fault detection model, using the enhanced fault feature images as training samples, and train the model until convergence; Step 5: After processing the vibration signals of the key equipment of the ship to be tested in Step 1-2, input them into the trained model, output the fault detection results and the corresponding class activation map, and realize the identification and interpretable decision-making of impact faults.

2. The interpretable detection method for impact-type faults in key ship equipment based on wavelet scattering and time-frequency feature enhancement as described in claim 1, characterized in that, The specific implementation of the improved wavelet scattering transform in step 1 is as follows: A complex Morlet wavelet is used as the mother wavelet to construct a filter bank covering the low-to-mid-frequency steady-state components and high-frequency impulse components. The filter bank parameters satisfy the following: in, For complex Morlet mother wavelet, For bandwidth parameters, For the center frequency, is the wavelet scaling factor, Q is the number of wavelets per octave, J is the maximum scale parameter, T is the sample length, and k and l are integers; The required filter bank is constructed as shown in the following equation: Furthermore, the original vibration signal is subjected to multi-scale wavelet decomposition and reconstruction, and the modulus of the wavelet coefficients at each scale is calculated to obtain the time-frequency image: Where x(t) is the original vibration signal, For the stretched wavelet basis, This indicates the inner product operation.

3. The adaptive time-frequency feature enhancement method according to claim 2, characterized in that... Frequency regularization, as a preprocessing method, is defined as follows: In the formula Represents the regularization factor. Indicates the bandwidth of the wavelet. This represents the bandwidth of the mother wavelet. Each k corresponds to a scale. . Furthermore, in step 3, the adaptive nonlinear time-frequency enhancement is defined as follows: In the formula, x represents the energy amplitude of the time-frequency representation before mapping. An exponential base that is adaptively determined by statistical characteristics; It is a small positive value close to 0; bias represents a steepness threshold, which physically means that numbers exceeding the threshold are approached to 1, and numbers below the threshold are approached to 0. According to claim 3, the main body of the interpretable fault detection model constructed in step 4 includes a feature extraction module and a decision visualization module. The former uses the first 5 convolutional blocks of a pre-trained VGG16 model, freezes the parameters of the first 2 convolutional blocks, and the remaining parameters can be trained. The input size of the fully connected layer of the model is 256, and the output layer is adaptively designed according to the specific diagnostic task. The specific structural parameters are shown in the table below. Table 1: Specific model parameters of the feature extraction module 。 4. The decision visualization module according to claim 2 includes a global average pooling layer, a fully connected layer, a softmax layer, and a CAM generation unit. The CAM generation unit obtains the class activation map by weighted summation of the last convolutional feature map, and the process is defined as follows: W represents the weights corresponding to the feature map, and c represents the output class of the model; where k represents the index of the feature map, CAM c It directly represents the spatial distribution of the basis for the model's predictions.

Images