A high-speed particle intelligent recognition and fault diagnosis method based on electrostatic-image feature fusion

Abstract

The application discloses a kind of high-speed particle intelligent identification and fault diagnosis method based on electrostatic-image feature fusion, belong to mechanical fault diagnosis and condition monitoring technical field.The present application aims at solving the problem of low recognition accuracy and poor reliability caused by image blur and weak signal in single modal monitoring in high-speed fluid environment.The method comprises the following steps: firstly, synchronously collect the electrostatic induction signal and visual image of the wear particles in the high-speed fluid;Secondly, EMD-ICA joint noise reduction is performed on the electrostatic signal to enhance the particle pulse feature, and the time series feature is extracted by inputting the Transform model, and the electrostatic recognition probability is output;The visual image is preprocessed by template matching and adaptive region scoring, background interference is eliminated, and the CNN model with fusion ECA attention mechanism is input, and the visual recognition probability is output;Finally, an adaptive decision fusion model based on meta-learning is constructed, and the fusion weight is dynamically calculated by the class-specific weight generation network after splicing the dual-modal probability vector.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis and condition monitoring technology, specifically relating to a wear particle identification and fault diagnosis method based on dual-modal information of electrostatic monitoring and machine vision, combined with meta-learning for adaptive fusion. Background Technology

[0002] Rotating machinery (such as wind turbine gearboxes and aircraft engines) is core equipment in modern industry, and its operating status directly affects the safety and economy of the entire system. Wear is a major form of mechanical failure. During the wear process, particles generated circulate in the system with lubricating oil or airflow, and their material, size, and morphology directly reflect the wear type, severity, and even the failure evolution trend. Therefore, inferring the internal condition of equipment through online monitoring of wear particles is a key technical path to achieving predictive maintenance.

[0003] However, real-time and accurate monitoring of wear particles in high-speed fluid environments faces significant challenges. On the one hand, the significant displacement of particles during high-speed motion leads to severe blurring of visual images, making it difficult for traditional image processing methods to extract effective morphological features. On the other hand, the electrostatic signals generated by particle collisions are weak in amplitude and exhibit non-stationary and transient characteristics, making them easily submerged by background noise and power frequency interference. Current monitoring methods mostly rely on single-modal data, visual methods are limited by imaging conditions, and electrostatic methods struggle to distinguish particles with similar geometric shapes, resulting in low recognition accuracy and poor robustness in high-speed dynamic environments. Furthermore, existing multi-source information fusion strategies often employ fixed weights or simple voting mechanisms, making it difficult to dynamically adjust the fusion strategy based on real-time operating conditions such as image blurring and signal-to-noise ratio, thus failing to meet the needs of complex industrial environments. Summary of the Invention

[0004] The purpose of this invention is to provide a high-speed intelligent particle identification and fault diagnosis method based on electrostatic-image feature fusion, to overcome the problems of incomplete single-modal monitoring information and the lack of adaptability of traditional fusion strategies in existing technologies. To achieve the above objective, this invention provides the following technical solution: S1: Synchronously acquire the raw electrostatic signals and raw visual images of wear particles in high-speed fluid; specifically, this is achieved as follows: a synchronization pulse signal is generated using a field-programmable gate array (FPGA) or a microcontroller, which simultaneously triggers frame capture by both the electrostatic acquisition system and the high-speed camera. The electrostatic acquisition system records the precise timestamp of each sampling point relative to the system startup time based on the rising edge of the synchronization pulse; the high-speed camera records the exposure start timestamp of each frame image based on the same synchronization pulse. During data post-processing, using the synchronization pulse as a reference, a timestamp interpolation matching algorithm is used to achieve sub-millisecond alignment between each pulse event in the electrostatic signal and the corresponding image frame. S2: Denoise the original electrostatic signal, extract the temporal features of the denoised signal, and output the first recognition probability vector based on the temporal features; preprocess the original visual image, extract the morphological features of the preprocessed image, and output the second recognition probability vector based on the morphological features. S3: Construct an adaptive decision fusion model based on meta-learning, adaptively fuse the first recognition probability vector and the second recognition probability vector, and output the final particle category.

[0005] In step S2, the original electrostatic signal is subjected to noise reduction processing, specifically including: S21: The original electrostatic signal is decomposed into multiple intrinsic mode function (IMF) components using empirical mode decomposition (EMD); S22: Perform independent component analysis (ICA) on all IMF components to obtain multiple independent source components; S23: Calculate the negative entropy value of each independent source component, select the top N components with the highest negative entropy values ​​for linear reconstruction, and obtain the denoised electrostatic signal.

[0006] This EMD-ICA joint noise reduction method can effectively remove background noise, enhance the non-Gaussian impulse characteristics generated by particle collisions, and improve signal quality. In a specific embodiment, for stainless steel fatigue ball particles, after EMD-ICA processing, the negative entropy value of each independent component is calculated and a line graph is plotted. It is observed that the negative entropy value shows a clear inflection point after the second component and tends to flatten out. Based on the elbow rule, the first two components with the highest negative entropy values ​​are selected for reconstruction, effectively preserving the particle collision impulse characteristics and suppressing background noise.

[0007] In step S2, the original visual image is preprocessed, specifically including: S24: Use template matching to identify and remove the interference region at the lower boundary of the pipe in the original visual image; S25: Perform adaptive threshold segmentation and morphological operations on the image after removing the boundary, extract candidate particle contours, and filter out effective particles according to preset geometric feature (such as area, roundness, aspect ratio, solidity) thresholds. S26: A sliding window method based on multi-index scoring is adopted, which comprehensively considers factors such as the number of particles, spatial density, regional centrality, and contrast within a region to adaptively select regions in the image and crop out the region containing the richest particle information. For example, the multi-index scoring function is: Score = W1·Nparticles + W2·Density - W3·CenterDist + W4·Contrast. Optimization is performed on the validation set through grid search, and an exemplary set of weights is determined as: W1=0.4, W2=0.2, W3=0.1, W4=0.3. This set of weights ensures that the region containing the most particles, with the most uniform distribution, good centrality, and high contrast receives the highest score. S27: Scale the cropped area to a uniform size and perform image enhancement processing to obtain a standardized particle image.

[0008] This preprocessing process can accurately locate particle targets, eliminate background interference, and standardize image specifications, providing high-quality input for subsequent recognition models.

[0009] In step S2, the second recognition probability vector is output based on morphological features. Specifically, this includes: inputting a standardized particle image into a convolutional neural network (CNN) that incorporates an efficient channel attention (ECA) mechanism; adaptively recalibrating the feature channel weights through the ECA module; strengthening the key morphological features of the particles; suppressing the interference caused by motion blur; and outputting the second recognition probability vector after passing through a fully connected layer.

[0010] In S3, constructing an adaptive decision fusion model based on meta-learning specifically includes: S31: Concatenate the first recognition probability vector with the second recognition probability vector to obtain the joint input vector; S32: Input the joint input vector into a shared feature extraction layer (e.g., a multilayer perceptron) to map and obtain a joint feature representation, which encodes the bimodal output characteristics of the current sample and the implicit confidence information; S33: Construct an independent weight generation network (e.g., a single-layer or multi-layer perceptron) for each particle category. Each network takes the joint feature representation as input and outputs a set of normalized electrostatic modal weights and visual modal weights for that category through the Softmax function. S34: Use the weights to perform a weighted summation of the first and second recognition probability vectors to obtain the preliminary fusion probability for each category; S35: Introduce a global adjustment factor generated from joint features to fine-tune the initial fusion probability, and output the final particle category after Softmax normalization.

[0011] In S3, the training loss function of the model is a comprehensive loss function, including: Cross-entropy classification loss is used to guide the model to classify correctly; The weighted objective guides the loss, based on the physical prior knowledge of different particles (e.g., metal particles rely more on electrostatic modes, while non-metal particles rely more on visual modes), guiding the weight allocation to conform to the preset direction and enhancing the interpretability of the model. Weight difference loss is used to limit the weight difference between two modes to be too large, preventing the model from going to extremes during training and completely ignoring the information of a certain mode.

[0012] The wear particles include, but are not limited to, metallic particles (such as stainless steel fatigue balls and brass cutting chips) and non-metallic particles (such as alumina ceramic balls, invasive corundum abrasives, exfoliated chromite particles, and graphite flake particles), covering different wear mechanisms.

[0013] Compared with the prior art, the present invention has the following beneficial effects: Information complementarity and improved accuracy: By fusing electrostatic monitoring, which is sensitive to material properties, and visual monitoring, which is sensitive to morphology, multi-source information complementarity is achieved, solving the problem of incomplete information from a single modality. Experimental results show that the fusion model achieves an average recognition accuracy of over 96% on six typical particle types, which is a significant improvement over the single-modality model.

[0014] Adaptive fusion with strong robustness: By introducing meta-learning concepts, a category-specific and sample-adaptive decision fusion framework is constructed. This framework can dynamically and intelligently adjust the fusion weights of the two modes based on contextual information such as the image blur level and signal-to-noise ratio of the current sample, effectively addressing the challenges of drastic changes in operating conditions in high-speed fluid environments.

[0015] In-depth signal processing and effective feature extraction: For electrostatic signals, an EMD-ICA joint denoising method is proposed, which effectively improves the signal-to-noise ratio; for blurred images, a preprocessing workflow based on template matching and adaptive scoring is designed, and an ECA attention mechanism is introduced to effectively enhance the extraction of key morphological features. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This invention provides a high-speed intelligent particle identification and fault diagnosis method based on electrostatic-image feature fusion. Figure 2 This is a flowchart of the single-modal data processing and recognition in this invention; Figure 3 This is a flowchart of the adaptive decision fusion based on meta-learning in this invention; Figure 4 This is a diagram of the Transformer model architecture in this invention. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0019] This embodiment provides a high-speed intelligent particle identification and fault diagnosis method based on electrostatic-image feature fusion, the overall process of which is as follows: Figures 1-4 As shown.

[0020] S1: Simultaneously acquire the raw electrostatic signals and raw visual images of wear particles in high-speed fluid: A simulated high-speed gas-solid two-phase flow wear particle experimental system was constructed. This system integrates a variable frequency gas supply module, a particle injection module, a quartz glass observation pipe, a high-speed imaging module, and an electrostatic acquisition module. Six typical wear particles were selected as research objects: stainless steel fatigue balls, alumina ceramic spheres, intrusive corundum abrasives, exfoliated chromite particles, graphite flakes, and brass cutting chips. Under a stable wind speed of 10 m / s, high-speed cameras (such as MemrECAm Hx-7S, 5000 fps) were used to synchronously acquire particle motion images, and a copper ball induction electrode combined with a high-precision charge meter (1000 Hz sampling rate) was used to synchronously acquire electrostatic signals, constructing a paired dataset including motion blur and background noise.

[0021] S2: Single-modal data processing and recognition: S21: The original electrostatic signal is decomposed into multiple intrinsic mode function (IMF) components using empirical mode decomposition (EMD).

[0022] EMD-ICA combined noise reduction: This process transforms the acquired raw electrostatic signal... Empirical mode decomposition (EMD) is performed to obtain a series of intrinsic mode function (IMF) components arranged from high frequency to low frequency. and a residual term : ; In the formula It consists of a series of IMF components arranged from high frequency to low frequency, each IMF representing an oscillation mode at a specific time scale in the original electrostatic signal; The residual term reflects the average trend or DC component of the signal.

[0023] S22: Perform independent component analysis (ICA) on all IMF components to obtain multiple independent source components.

[0024] All IMF components Constructing the observation signal matrix Assume the observed signal is a linear mixture of several statistically independent source signals, i.e., a mixing matrix exists. Source signal matrix satisfy The goal of Independent Component Analysis (ICA) is to find an unmixing matrix. This makes the output components To achieve statistical independence as much as possible, thereby approximating the true source signal. .

[0025] S23: Calculate the negative entropy value of each independent source component, select the top N components with the highest negative entropy values ​​for linear reconstruction, and obtain the denoised electrostatic signal.

[0026] To measure the independence of each component, each independent source component is calculated. negative entropy The larger the negative entropy value, the stronger the non-Gaussianity of the signal and the richer the effective granular information it contains. ; In the formula: Represents the mathematical expectation; It is a Gaussian variable with zero mean and unit variance; It is a nonlinear function.

[0027] Transformer model recognition: The denoised electrostatic signal sequence is input into a classification model based on a Transformer encoder. First, a feature embedding layer maps the input to a high-dimensional space, and then positional encoding is superimposed to incorporate temporal order information. For the input matrix First, through a learnable weight matrix , , They are linearly mapped to query matrices respectively. Key matrix Sum matrix : ; After that, Split into Each head independently computes its scaled dot product attention: ; In the formula: The dimension of the key vector, and the scaling factor. Used to maintain gradient stability.

[0028] Finally, the outputs of all heads are concatenated, and the final output of the multi-head self-attention layer is obtained through a linear transformation. : ; In the formula: This is the learnable output weight matrix.

[0029] Feedforward networks and residual connections are used to perform nonlinear transformations and stability optimizations on the attention output. (Output from the attention layer) The input will be a feedforward network consisting of fully connected layers, which performs the same independent processing on the representations of each position in the sequence: ; In the formula: These represent the weights and bias parameters of the two network layers, respectively. The output of each sub-layer (self-attention mechanism layer and feedforward layer) undergoes residual connection and layer normalization processing, and its calculation can be uniformly expressed as: ; In the formula: Input for sub-layer; For sub-layer functions; Presentation layer normalization operation.

[0030] The introduction of residual connections enables the training of deep networks, which is particularly crucial for tasks requiring multiple layers, such as electrostatic signal processing. Shallow networks may only extract pulse amplitude information, while deep networks can progressively abstract more complex patterns, such as pulse shape symmetry and multi-pulse resonance modes. Layer normalization ensures a relatively stable distribution of outputs at each layer, avoiding gradient fluctuations during training. Through this structure, the Transformer encoder can perform deep feature extraction and global dependency modeling of input sequences, providing a robust feature representation foundation for subsequent tasks.

[0031] S24: Use template matching to identify and remove the interference region at the lower boundary of the pipe in the original visual image.

[0032] Using a pre-cut pipe lower boundary template In the original image Template matching is performed, and the normalized correlation coefficient is calculated. ; In the formula: This represents the average grayscale value of the template. This represents the average grayscale value of a local region of the image. Select the location corresponding to the maximum response value. As the boundary location, remove the following area at that location: ; In the formula: For safety margins, ensure that the boundary area is completely removed.

[0033] S25: Perform adaptive thresholding and morphological operations on the image after removing the boundaries, extract candidate particle contours, and select effective particles according to the preset geometric feature thresholds.

[0034] For the image after removing the boundaries Adaptive threshold segmentation combined with morphological operations is used to extract candidate particle contours: ; In the formula: It is a local mean; This is a constant offset. Then, the binary image is processed... Perform closing and opening operations to connect adjacent regions and eliminate noise: ; right Extracting the contour set And calculate the geometric features of each contour for filtering: area: ; Circularity: ,in The perimeter of the outline; Aspect Ratio: ,in , The length and width of the circumscribed rectangle; Solidity: ,in This represents the area of ​​the convex hull.

[0035] Retain contours that meet the following conditions as valid particles: .

[0036] S26: A sliding window method based on multi-index scoring is used to adaptively select regions in the image and crop out the region containing the richest particle information.

[0037] To select the most representative region from the image, a multi-index scoring function is designed: ; In the formula: Score the number of particles in the region; Score the spatial density of the particles; The region centrality score (distance from the image center); Assign a regional contrast score; These are the weighting coefficients. Iterate through all possible regions and select the region with the highest score as the cropping candidate: ; S27: Scale the cropped area to a uniform size and perform image enhancement processing to obtain a standardized particle image.

[0038] ; Where: the contrast gain coefficient is taken. Brightness shift This makes the outline clearer. Finally, an unsharpened mask (USM) is used to enhance edge sharpness: first, a Gaussian blur is applied to the image to obtain the low-frequency components. Then extract the edge details and overlay them back onto the original image: ; Where: Gaussian kernel standard deviation Sharpening intensity This process naturally sharpens grain edges without producing white edge artifacts. The final image is obtained after three enhancement steps. , as input to the visual model.

[0039] CNN-ECA model recognition: The standard residual block is calculated as follows: ; In the formula: For input features; Represents the residual mapping to be learned; These are the weights for the convolutional layers. This shortcut connection structure helps alleviate the vanishing gradient problem in deep networks, enabling the model to be trained effectively.

[0040] An ECA attention module is introduced after residual mapping. Given the input feature map... The ECA module first aggregates spatial information through global average pooling: ; Subsequently, one-dimensional convolution is used to capture local channel interactions, and the kernel size is... Based on the number of channels Adaptive determination: ; Channel attention weights are normalized using the Sigmoid function: ; Finally, the original feature map is multiplied channel by channel with the attention weights to achieve feature recalibration: ; In the formula: This indicates element-wise multiplication. This mechanism can adaptively enhance the response of important channels and suppress redundant information, thus enabling the focus on key morphological features of particles even under motion blur conditions.

[0041] After four stages of residual learning and attention enhancement, the network ends with adaptive average pooling to unify the feature map size to 1×1, resulting in a 512-dimensional feature vector. .

[0042] To prevent overfitting, a Dropout layer is introduced to randomly drop off some neurons (dropout probability). Finally, the recognition probabilities of six types of wear particles are output through a fully connected layer and a softmax function: ; In the formula: This is the probability distribution vector output by the model, with each component corresponding to the predicted probability of six types of wear particles; This is the weight matrix of the fully connected layer; The feature vector after Dropout processing; This is the bias term; the Softmax function normalizes the linear transformation result into a probabilistic form, ensuring that the sum of all output components is 1.

[0043] S3: Meta-learning-based adaptive decision fusion: S31: Concatenate the first recognition probability vector with the second recognition probability vector to obtain the joint input vector.

[0044] Given a sample, its electrostatic probability vector and visual probability vector Concatenate them into a single joint input vector: .

[0045] S32: Input the joint input vector into the shared feature extraction layer and map it to obtain the joint feature representation.

[0046] To extract deeper features to reveal complex relationships between modalities, a multilayer perceptron (MLP) is used. Mapped to a hidden feature space : ; In the formula: Represents the ReLU activation function; , , This is the weight matrix; , , This is the bias vector.

[0047] S33: Construct an independent weight generation network for each particle category. Each network takes the joint feature representation as input and outputs a set of normalized electrostatic modal weights and visual modal weights for that category through the Softmax function.

[0048] For each category Design an independent weight generation network. Each With shared features As input, output a two-dimensional vector, then... Function normalization yields the fusion weight vector for that category. : ; In the formula: and These are the weighting coefficients; and This is the bias vector. The Softmax function guarantees... .

[0049] S34: Use weights to sum the first and second recognition probability vectors to obtain the preliminary fusion probability for each category.

[0050] Calculate the initial fusion probability for each category: .

[0051] S35: Introduce a global adjustment factor to fine-tune the initial fusion probability, and output the final particle category after Softmax normalization.

[0052] Introducing a global adjustment factor The initial fusion probability was fine-tuned, with the adjustment range controlled within... Between, to avoid over-correction: ; Finally, the adjusted probabilities of all categories are normalized using the Softmax function to obtain the final fused probability vector. : ;

[0053] Pick The class with the largest component is used as the final decision for that sample.

[0054] Model training: The overall loss function of the model is as follows: ; Classification loss Using the standard cross-entropy loss, its expression is: ; In the formula: One-Hot encoding for real labels; for The Each component has a uniform number of samples in each class of the training set, so there is no need to introduce additional class weights; the standard cross-entropy loss can effectively guide the model's learning.

[0055] Weighted target guided loss Based on the physical priors of different particles, a reasonable weight allocation direction is preset for each category. For metal particles that are mainly distinguished by electrostatic features, the loss function will incentivize the model to learn... The pattern; for non-metallic particles where morphological characteristics are more critical, then the excitation... For graphite particles with unique negative charge, the same excitation is also possible. Its mathematical form can be uniformly represented as: ; In the formula: The ideal weight vector is defined based on physical priors. These are the constraint strength coefficients for different categories. Through this loss, the model is guided to adjust the weights in a direction consistent with physical priors, but not strictly required to be equal to the target value, thus enhancing interpretability while preserving adaptability.

[0056] Weighted difference loss Used to prevent the model from going to extremes and completely ignoring a certain mode, penalizing excessively large differences in the weights between two modes: ; In the formula: The number of samples; This is the maximum allowable difference threshold. This loss ensures that the weight difference between the two modes is not too large, thus preserving the potential contribution of the other mode.

[0057] For training strategies, the model uses the Adam optimizer with an initial learning rate of 1×10⁻³ and weight decay of 1×10⁻⁴. Cosine annealing is used to adjust the learning rate, with a maximum epoch of 50 and a minimum learning rate of 1×10⁻⁵. The training epochs are 150, with an early stopping mechanism (Patience=20) and a batch size of 16. During training, The weight decreases linearly from 0.2 to 0.05, allowing the model to focus on learning a reasonable weight allocation in the early stages and on classification accuracy in the later stages. The value is fixed at 0.05 to maintain constraints on extreme weights. All hyperparameters are determined through grid search on the validation set.

[0058] To verify the effectiveness of the proposed comprehensive loss function, ablation experiments were conducted. Experimental results show that when only cross-entropy loss is used, the model's average recognition accuracy decreases by approximately 2%, and the weight allocation for some categories deviates from the physical prior (e.g., the visual weight of metal particles is abnormally increased). After removing the weighted objective-guided loss, the model's recognition accuracy for non-metallic particles (e.g., alumina ceramic spheres) decreases by 3.5%, indicating that this loss effectively guides the model to learn a fusion strategy consistent with physical mechanisms. After removing the weight difference loss, the model exhibits extreme weight allocations on some samples (e.g., the weight of a certain modality exceeds 0.95), leading to a decrease in robustness. This demonstrates that all components in the comprehensive loss function of this invention significantly contribute to the model's performance.

[0059] For training hyperparameter settings, the Adam optimizer was used with an initial learning rate of 1e-3 and weight decay of 1e-4. Cosine annealing was employed to adjust the learning rate. The weight-target guided loss coefficient λ was used. warm The weighted difference loss coefficient λ decreases linearly from 0.2 to 0.05. diff The value was fixed at 0.05. All hyperparameters were determined through grid search on the validation set.

[0060] Experimental results: The model was trained and tested on a dataset containing 1200 paired samples. Experimental results show that the fusion model of this invention achieves an average recognition accuracy of 96.0% on six types of particles, significantly higher than the single electrostatic model (91.0%) and the single visual model (88.2%). Weight allocation analysis shows that the model automatically assigns a higher weight (approximately 0.67) to the electrostatic mode for metallic particles (such as stainless steel fatigue balls), while assigning a higher weight (approximately 0.66) to the visual mode for non-metallic particles (such as alumina ceramic spheres). This is highly consistent with the physical mechanism, verifying the effectiveness and interpretability of the model.

[0061] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0062] This invention is intended to cover all such substitutions, modifications, and variations falling within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A high-speed intelligent particle identification and fault diagnosis method based on electrostatic-image feature fusion, characterized in that, Includes the following steps: S1: Simultaneously acquire the raw electrostatic signals and raw visual images of wear particles in high-speed fluid; S2: Denoise the original electrostatic signal, extract the temporal features of the denoised signal, and output the first recognition probability vector based on the temporal features; preprocess the original visual image, extract the morphological features of the preprocessed image, and output the second recognition probability vector based on the morphological features. S3: Construct an adaptive decision fusion model based on meta-learning, adaptively fuse the first recognition probability vector and the second recognition probability vector, and output the final particle category.

2. The method according to claim 1, characterized in that, In step S2, the original electrostatic signal is subjected to noise reduction processing, specifically including: S21: The original electrostatic signal is decomposed into multiple intrinsic mode function (IMF) components using empirical mode decomposition (EMD); S22: Perform independent component analysis (ICA) on all IMF components to obtain multiple independent source components; S23: Calculate the negative entropy value of each independent source component. Based on the cumulative contribution rate of the negative entropy value or the elbow rule after sorting the negative entropy values, determine the number of components N to be selected. Select the top N components with the highest negative entropy values ​​for linear reconstruction to obtain the denoised electrostatic signal.

3. The method according to claim 2, characterized in that, In step S2, the first recognition probability vector is output based on temporal features. Specifically, this includes: inputting the denoised electrostatic signal into a classification model based on a Transformer encoder, using its multi-head self-attention mechanism to mine the local pulse details and global temporal dependencies of the signal in parallel, and outputting the first recognition probability vector after passing through the classification layer.

4. The method according to claim 1, characterized in that, In step S2, the original visual image is preprocessed, specifically including: S24: Use template matching to identify and remove the interference region at the lower boundary of the pipe in the original visual image; S25: Perform adaptive threshold segmentation and morphological operations on the image after removing the boundary, extract candidate particle contours, and select effective particles according to the preset geometric feature threshold. S26: A sliding window method based on multi-index scoring is used to adaptively select regions in the image and crop out the regions containing the richest particle information; wherein, the weight coefficients of each item in the multi-index scoring function are predetermined by the analytic hierarchy process or by grid search based on the validation set. S27: Scale the cropped area to a uniform size and perform image enhancement processing to obtain a standardized particle image.

5. The method according to claim 4, characterized in that, In step S2, the second recognition probability vector is output based on morphological features. Specifically, this includes: inputting a standardized particle image into a convolutional neural network (CNN) that incorporates an efficient channel attention (ECA) mechanism; adaptively recalibrating the feature channel weights through the ECA module to enhance the key morphological features of the particles; and outputting the second recognition probability vector after passing through a fully connected layer.

6. The method according to claim 1, characterized in that, In S3, constructing an adaptive decision fusion model based on meta-learning specifically includes: S31: Concatenate the first recognition probability vector with the second recognition probability vector to obtain the joint input vector; S32: Input the joint input vector into the shared feature extraction layer to obtain the joint feature representation; S33: Construct an independent weight generation network for each particle category, with no parameter sharing. Each network takes the joint feature representation as input and outputs a set of normalized electrostatic modal weights and visual modal weights for that category through the Softmax function. S34: Use the weights to perform a weighted summation of the first and second recognition probability vectors to obtain the preliminary fusion probability for each category; S35: Introduce a global adjustment factor to fine-tune the initial fusion probability, and output the final particle category after Softmax normalization.

7. The method according to claim 6, characterized in that, In S3, the model's training loss function is a comprehensive loss function, including: cross-entropy classification loss; and weighted objective guidance loss; and, Weight difference loss; wherein the coefficients of the weight target guidance loss and the weight difference loss are dynamically adjusted during the training phase or determined through grid search optimization.

8. The method according to any one of claims 1 to 7, characterized in that, The wear particles include metallic particles and non-metallic particles. The metallic particles include stainless steel fatigue balls and brass cutting chips. The non-metallic particles include alumina ceramic balls, invasive corundum abrasives, exfoliated chromite particles, and graphite flake particles.

Images