Metallic abrasive grain shape recognition method and system based on dynamic polar coordinate transformation and attention mechanism

Abstract

This invention discloses a method and system for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and an attention mechanism. The method includes the following steps: acquiring and preprocessing a one-dimensional induced voltage time-series signal of metal abrasive particles; constructing a sliding window; traversing the one-dimensional induced voltage time-series signal point by point; selecting the point with the minimum amplitude within the window as the pole; calculating the polar radius and polar angle; nonlinearly reconstructing the preprocessed one-dimensional induced voltage time-series signal into a polar radius feature sequence and a polar angle feature sequence; constructing an attention-enhanced abrasive particle recognition model; inputting the polar radius feature sequence and the polar angle feature sequence, along with the preprocessed one-dimensional induced voltage time-series signal, into the attention-enhanced abrasive particle recognition model; performing adaptive weighting to obtain an attention-enhanced feature map; and inputting this map into a classifier to output the specific morphology category of the abrasive particle. This technical solution utilizes nonlinear polar coordinate reconstruction to capture local dynamic fluctuations, improving the accuracy and robustness of metal abrasive particle morphology recognition.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensing technology for mechanical equipment, and relates to a method and system for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism. Background Technology

[0002] During the entire lifecycle of mechanical equipment, the relative motion and contact friction between components inevitably generate metal abrasive particles. Under different operating stages (such as the break-in period, stable operation period, and period of frequent failures) and different wear patterns, the generated abrasive particles exhibit significant differences in morphology, size, and structure. For example, normal wear conditions primarily produce regularly shaped spherical abrasive particles, while fatigue wear and adhesive wear often produce large, plate-like and elongated particles. These morphological differences not only affect the voltage peak generated when abrasive particles pass through an inductive sensor but also exhibit characteristic differences in timing waveforms, polarity changes, and peak distribution.

[0003] Therefore, accurate identification of abrasive particle morphology can not only effectively determine the specific wear pattern of the current mechanical equipment, but also quantitatively assess the severity of wear, providing a reliable basis for fault warning and operation and maintenance of mechanical equipment.

[0004] In current online abrasive monitoring, inductive metal abrasive sensors are commonly used to collect signals. However, due to the small size of the abrasive particles themselves, the waveform of the induced voltage signal output by different types of metal abrasive particles passing through the sensor is almost indistinguishable. The peak output voltage of the sensor is mainly related to the size of the abrasive particles and cannot effectively reflect the morphological differences of the abrasive particles. Furthermore, the subtle feature changes caused by the differences in abrasive particle morphology are only hidden in the dynamic evolution of the voltage signal. Traditional identification methods that rely on intuitive features or directly input raw time-series data into the model not only have low classification accuracy but also significantly reduce the model training efficiency.

[0005] To enhance the differences between the minute features of different abrasive grains, static polar coordinate transformation and wavelet transform feature extraction are commonly used in the field of time series data preprocessing.

[0006] However, existing static polar coordinate transformation methods typically use the extreme points of the entire time series as fixed poles, which cannot adapt to the dynamic fluctuation characteristics of local data. This easily leads to key local features being masked by global features, making it difficult to accurately capture subtle morphological differences in abrasive grain signals. On the other hand, methods such as wavelet transform cannot fully preserve the temporal correlation of the original abrasive grain time series data during feature extraction, and their computational complexity is high, making them unsuitable for the practical application requirements of rapid preprocessing of abrasive grain detection signals.

[0007] Furthermore, existing static polar coordinate mapping or wavelet feature extraction methods cannot adapt to local fluctuations and are prone to losing temporal correlation. In particular, their adaptability and accuracy are limited when facing multi-peak, alternating polarity signals output by four-coil sensors.

[0008] Therefore, there is an urgent need for a high-precision identification method that can combine abrasive particle morphology characteristics, sensor physical properties, and deep learning. Summary of the Invention

[0009] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a method and system for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism.

[0010] To achieve the above objectives, the basic solution of this invention is: a method for metal abrasive grain morphology recognition based on dynamic polar coordinate transformation and attention mechanism, comprising the following steps:

[0011] S1, using a fully excited four-coil inductive sensor, acquires the one-dimensional induced voltage timing signal of metal abrasive particles and performs preprocessing;

[0012] S2, a sliding window is constructed by combining the abrasive flow velocity and the effective magnetic field length of the coil. The pre-processed one-dimensional induced voltage timing signal is traversed point by point. The point with the minimum amplitude within the window is dynamically selected as the pole, and the weighting coefficient is calculated by combining the local and global amplitude ranges.

[0013] S3, based on the poles and weighting coefficients, the pole radius and pole angle are obtained, and the preprocessed one-dimensional induced voltage time sequence signal is nonlinearly reconstructed into a pole radius feature sequence and a pole angle feature sequence.

[0014] S4. The polar radius feature sequence and polar angle feature sequence, together with the preprocessed one-dimensional induced voltage time series signal, are input into the attention-enhanced abrasive particle recognition model. First, basic feature extraction and dimensionality increase are performed through the initial convolutional layer, and then the model is split into local and global branches for feature extraction. After the outputs of the two branches are aligned in size, they are spliced ​​and fused, and adaptive weighting of channel and spatial dimensions is performed to obtain the attention-enhanced feature map with dynamic polar coordinate fusion.

[0015] S5 inputs the attention-enhanced feature map fused with dynamic polar coordinates into the classifier, and outputs the specific morphological category of the abrasive grains.

[0016] The working principle and beneficial effects of this basic scheme are as follows: This technical scheme achieves high-dimensional feature reconstruction and enhancement of the one-dimensional induced voltage timing signal of metal abrasive particles by combining dynamic polar coordinate transformation with an attention mechanism. This method can effectively capture subtle differences in abrasive particle morphology in both local and global ranges, improving the discriminative ability of feature representation.

[0017] By dynamically selecting poles through a sliding window and combining local and global amplitude ranges for nonlinear polar coordinate reconstruction, the local fluctuation characteristics in the original time series signal can be preserved, and the sensitivity to wear particle signals of different morphologies can be enhanced.

[0018] By employing an adaptive weighting mechanism with dual channel and spatial attention, the model's focus on key morphological features is significantly enhanced, and redundant information is suppressed. This improves the accuracy and robustness of metal abrasive particle morphology recognition, making it suitable for online monitoring of wear conditions in mechanical equipment under complex working conditions.

[0019] Furthermore, the one-dimensional induced voltage timing signal of the collected metal abrasive particles is preprocessed, specifically as follows:

[0020] The min-max method is used to process the original time series data as follows:

[0021] ,

[0022] in, This is the original timeline data. This is the original time series data; The timeline data is normalized. The normalized induced voltage amplitude data, , These are the minimum and maximum values ​​of the original timeline data, respectively. , The minimum and maximum amplitude values ​​are the original induced voltage time series data, and n is the index number of the current discrete data point.

[0023] The min-max normalization preprocessing method can eliminate the dimensional effects of sensor differences or changes in the acquisition environment in the original signal, making the subsequent dynamic polar coordinate transformation and model training have better numerical stability and consistency, which is beneficial to improving the model convergence speed and recognition performance.

[0024] Furthermore, the method for constructing a sliding window by combining the abrasive flow velocity and the effective magnetic field length of the coil is as follows:

[0025] Let the total length of the single-layer solenoid coil of the fully excited four-coil inductive sensor be... The controlled flow rate of metal abrasive particles passing through the sensor in the lubricating oil circuit is: The sampling frequency of data acquisition is The time period during which a single metal abrasive grain passes through a single-layer solenoid coil, causing an effective local magnetic induction intensity disturbance. for:

[0026] ,

[0027] Sliding window length (Total number of data points included) and window half-length ,for:

[0028] ,

[0029] ,

[0030] in, This indicates a floor operation, ensuring that the window length is the number of discrete data points that are valid integers;

[0031] Let the preprocessed one-dimensional induced voltage time series data sequence be... The corresponding timeline sequence is ,in The total length of the sequence is represented by data points. Create a sliding window centered on the element and add a boundary judgment mechanism:

[0032] ,

[0033] ,

[0034] in, This is the starting index of the sliding window. This is the ending index of the sliding window.

[0035] By combining the abrasive grain flow rate, coil length, and sampling frequency to construct the sliding window length, the window has a clear physical meaning and can match the effective signal period when a single abrasive grain passes through the coil.

[0036] A boundary judgment mechanism is introduced to ensure the integrity of the sliding window in the beginning and end regions of the signal, thereby improving the rationality and stability of pole selection.

[0037] Furthermore, the method for dynamically selecting the point with the minimum amplitude within the window as the pole by iterating through the preprocessed one-dimensional induced voltage time-series signal point by point is as follows:

[0038] Within the effective sliding window corresponding to the current data point, each point is traversed and the data point with the smallest normalized amplitude is selected as the dynamic pole of the current window and used as the reference point for the current local interval to participate in subsequent feature calculations. (Pole index...) for:

[0039] ,

[0040] in, This indicates the location of the data point with the smallest amplitude within the sliding window; j is the sequence number of the preprocessed one-dimensional induced voltage time series data.

[0041] Normalized time axis coordinates of dynamic poles With normalized amplitude coordinates Calculate the current data point respectively The difference between the coordinates of the dynamic pole and the time axis and the coordinate difference in amplitude The time axis coordinate difference and the amplitude coordinate difference together characterize the time-amplitude deviation of the current data point relative to the local peak and valley reference point, providing basic variables for polar coordinate mapping.

[0042] It adaptively captures changes in local signal structure, avoiding feature masking problems caused by globally fixed poles, and enhances the ability of features to express local morphological differences.

[0043] Furthermore, weighting coefficients are calculated by combining local and global amplitude ranges to achieve adaptive adjustment of the feature representation intensity for different local intervals, as follows:

[0044] ,

[0045] in, This represents the range of original data amplitudes within the current sliding window, that is, the difference between the maximum and minimum values ​​of the data within the window. It represents the range of the entire original data amplitude.

[0046] By introducing weighting coefficients for local and global amplitude ranges, the calculation of polar radius and polar angle takes into account both local signal fluctuations and overall signal distribution, improving the sensitivity of polar coordinate mapping to subtle morphological differences and enhancing the discriminative ability of feature sequences. The weighting coefficients vary with the relative changes in the local and global amplitude ranges, and are used to strengthen the regions with obvious local fluctuation characteristics and suppress the regions with weak feature responses in the calculation of polar radius and polar angle.

[0047] Furthermore, based on the time axis coordinate difference between the current data point and the dynamic pole... Amplitude coordinate difference with weighting coefficients One-dimensional time-domain features are uniformly compressed and mapped to two-dimensional polar coordinate space to construct a joint representation of local fluctuation features, peak-valley distribution features, and multi-peak time-series response features, generating polar radii respectively. With polar angle ,for:

[0048] ,

[0049] in, It is the arctangent function in the fourth quadrant, and its range is [0, 2π).

[0050] Generate the corresponding polar radius sequences based on the polar radius formula and the polar angle formula, respectively. and polar sequence ,for:

[0051] ,

[0052] ,

[0053] in, and These represent the polar radius and polar angle generated by the nth data point in the preprocessed one-dimensional time-domain feature, respectively. .

[0054] By mapping a one-dimensional time-domain signal to a two-dimensional polar coordinate space through weighted polar radius and four-quadrant polar angles, nonlinear compression and reconstruction of the signal morphology are achieved. This effectively preserves the temporal structure and amplitude variation trend of the original signal, providing more discriminative feature inputs for subsequent attention models.

[0055] Furthermore, the backbone network of the attention-enhanced abrasive particle recognition model in step S4 adopts a dual-branch residual architecture, specifically as follows:

[0056] After the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time series signal are subjected to convolution, batch normalization, ReLU activation, and max pooling to complete the initial extraction of basic features and dimensionality increase, the signal is then split into local and global branches:

[0057] The local branch is composed of two CS-ResNeXt blocks embedded with Convolutional Block Attention (CBAM) stacked in series. It is used to perform continuous convolution extraction while maintaining the spatial resolution of the feature map. This allows the local details of the input signal to continue to be transmitted during the feature extraction process. The key responses are highlighted by CBAM weighting, and the input features are then fused through residual connections to enhance the local feature representation.

[0058] The global branch consists of two CS-ResNeXt blocks for downsampling, a one-dimensional transposed convolutional layer, and a CS-ResNeXt block for feature refinement. Specifically, the first two CS-ResNeXt blocks downsample the feature map twice consecutively using a convolution extraction method with a stride of 2, thereby compressing the feature map length step by step, thus expanding the network's receptive field and capturing global contextual features.

[0059] The global features obtained after downsampling are then passed through the one-dimensional transposed convolutional layer to restore the spatial resolution of the feature map, so that it is consistent with the size of the local branch output;

[0060] The feature map after resolution restoration is input into the last CS-ResNeXt block for feature refinement to obtain the global branch output features;

[0061] The CS-ResNeXt block is a residual bottleneck structure with an attention enhancement mechanism, based on the ResNeXt architecture. (The main path sequentially includes the first convolutional layer, the first batch of normalization layers, the first ReLU activation layer, the second convolutional layer, the second batch of normalization layers, the second ReLU activation layer, a 1×1 convolutional layer, the third batch of normalization layers, the convolutional block attention module (CBAM), and the third ReLU activation layer. Basic features are extracted through two group convolutions, channels are normalized by a 1×1 convolution, and adaptive weighting is performed by embedding the convolutional block attention module (CBAM). Finally, the input features are fused through residual connections.

[0062] After the feature maps output by the local branch and the global branch are aligned in size, they are spliced ​​and fused in the channel dimension by the Concat layer to achieve feature unification of the micro-contour of the abrasive particles and the macro-motion state.

[0063] The fused features are transformed into a one-dimensional global feature vector through adaptive global average pooling, mapped to the classification space through a fully connected layer, and finally output as a classification result of abrasive grain type through the Softmax activation function.

[0064] The attention-enhanced abrasive particle recognition model of the present invention extracts local fine features, retains the subtle fluctuations, peak-valley transitions and edge abrupt changes of the abrasive particle induced voltage signal in the local range, and enhances the ability to characterize the weak responses corresponding to the local morphological differences of abrasive particles.

[0065] Further, in step S4, the polar radius feature sequence and polar angle feature sequence, along with the preprocessed one-dimensional induced voltage time-series signal, are input into the attention-enhanced abrasive particle recognition model for adaptive weighting based on channel and spatial dimensions. The specific steps are as follows:

[0066] Feature enhancement is achieved using the model's embedded Convolutional Block Attention (CBAM) module, which involves the collaborative computation of channel attention and spatial attention sub-modules. Specifically:

[0067] Let the intermediate feature map obtained by feature extraction and dimensionality upscaling fusion of the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time-series signal through the convolutional layer inside the attention-enhanced abrasive particle recognition model be the CBAM input feature map, and let it be denoted as... C, H, and W represent the number of channels, height, and width of the feature map, respectively. Global average pooling (GAP) and global max pooling (GMP) are performed on the input feature map to obtain two dimensions. The channel feature vectors are denoted as follows: and :

[0068] ,

[0069] in, This represents the feature value of the c-th channel, h-th row, and w-th column of the input feature map; This represents the average response intensity of the c-th channel across the entire feature map, characterizing the overall response level of this channel to abrasive morphology-related modes. This represents the maximum response intensity in the c-th channel, describing the channel's ability to capture locally significant responses;

[0070] Will and The data is concatenated and fed into a shared network consisting of two fully connected layers (FC) and an activation function to achieve feature compression and reconstruction. It comprehensively utilizes global statistical responses and local salient responses to learn the dependencies between different channels. Finally, it outputs channel attention weight vectors through a sigmoid activation function. ,for:

[0071] ,

[0072] in, ReLU represents the Sigmoid activation function, and ReLU represents the rectified linear activation function. and These represent two fully connected layers, used to compress and restore the feature dimension. This indicates channel concatenation of two pooled feature vectors;

[0073] Channel attention weight vector The input feature map X is multiplied channel by channel to enhance the feature channel responses related to abrasive grain morphology recognition, suppress redundant channel responses that contribute little to recognition, and highlight the separability of local fluctuation features, peak-valley distribution features, and multi-peak temporal response features in the channel dimension, resulting in a weighted feature map enhanced by the channel attention mechanism. ;

[0074] right Global average pooling and global max pooling are performed separately along the channel dimension. Average pooling is used to calculate the average response intensity of each time position across all feature channels to characterize the overall activation level of wear-grain morphology-related features at that position. Max pooling is used to extract the strongest response value of each time position across all feature channels to characterize the prominence of local significant wear-grain morphology features at that position, resulting in two dimensions. Spatial feature maps, respectively denoted as and ,for:

[0075] ,

[0076] in, This represents the feature value of the c-th channel, h-th row, and w-th column of the channel-weighted feature map.

[0077] Will and By concatenating the channels, we obtain the dimension as follows: The feature map is then compressed to 1 channel number through a convolutional layer (Conv), and the spatial attention weight map is output after passing through a sigmoid activation function. ,for:

[0078] ,

[0079] Where Conv represents the convolution operation, the spatial attention weight map can characterize the contribution of different temporal positions to the recognition of abrasive grain morphology, and the positions with larger weights correspond to key positions where local fluctuation features, peak and valley distribution features and multi-peak temporal response features are more significant.

[0080] Spatial attention weight map Feature map weighted by channel Element-wise multiplication is performed to obtain the final attention-enhanced feature map. .

[0081] By using the CBAM module to perform dual weighting of the feature map in terms of both channel and spatial dimensions, the key feature channels and regions related to abrasive grain morphology can be adaptively enhanced, while suppressing irrelevant or interfering information, significantly improving the model's ability to model complex signal morphologies and its recognition accuracy.

[0082] Furthermore, in step S5, the attention-enhanced feature map fused by dynamic polar coordinates is input into the classifier, and the specific morphological category of the abrasive grains is output. The specific steps are as follows:

[0083] The attention-enhanced feature map is passed sequentially through a global average pooling layer and a fully connected layer to reduce the dimensionality of the high-dimensional features and integrate them to obtain a one-dimensional feature vector for abrasive grain morphology discrimination.

[0084] The integrated one-dimensional feature vector is input into a classifier with a softmax activation function to calculate the probability distribution of the current metal abrasive particles belonging to different morphological categories, as follows:

[0085] ,

[0086] in, The current metal abrasive particles belong to the first... The probability of each morphological category. The first output of the fully connected layer The feature logarithmic value of each morphological category For category indexing, The total number of morphological categories, It is a natural constant;

[0087] The category label with the highest confidence in the probability distribution is extracted as the final morphology recognition result of the metal abrasive particles.

[0088] By using global average pooling and fully connected layers to reduce the dimensionality of the enhanced feature maps and combining them with the morphological probability distribution output by the Softmax classifier, efficient and interpretable abrasive morphology classification is achieved.

[0089] The present invention also provides a metal abrasive particle morphology recognition system based on dynamic polar coordinate transformation and attention mechanism, including a front-end acquisition device, a metal abrasive particle morphology recognition device and a result display device;

[0090] The pre-acquisition device is used to acquire the one-dimensional induced voltage timing signal of metal abrasive particles. The pre-acquisition device includes a fully excited four-coil inductive sensor, which includes four multilayer solenoid coils with completely identical parameters. The four coils are coaxially and closely arranged outside the insulating oil pipe skeleton, forming a hollow fluid detection channel. The four coils are sequentially divided into a first bridge arm coil, a second bridge arm coil, a third bridge arm coil, and a fourth bridge arm coil along the axial direction, forming a symmetrical Wheatstone bridge circuit. The first bridge arm coil and the second bridge arm coil form the first group, and the third bridge arm coil and the fourth bridge arm coil form the second group. Two adjacent coils in the same group are wound in the same direction to form a positive reinforcement of mutual inductance. The coils in the first group and the coils in the second group are wound in opposite directions. The metal abrasive particles pass through the fluid detection channel along the axial direction with the lubricating oil, causing alternating changes in the effective inductance of each bridge arm coil, and differentially outputting the induced voltage signal through the Wheatstone bridge circuit.

[0091] The input end of the metal abrasive particle morphology recognition device is connected to the output end of the pre-acquisition device. The metal abrasive particle morphology recognition device executes the method described in this invention and outputs the specific morphology category of the abrasive particles to the result display device for display.

[0092] This system acquires high signal-to-noise ratio one-dimensional induced voltage signals through a front-end fully excited four-coil inductive sensor, and combines it with a morphology recognition device to achieve fully automated processing from signal acquisition to morphology output, so as to map the wear pattern and operating status of the equipment based on the classification results.

[0093] Furthermore, the metal abrasive grain morphology recognition device includes:

[0094] The signal acquisition and preprocessing module is used to receive the one-dimensional induced voltage timing signal of metal abrasive particles and to perform normalization processing to eliminate the influence of dimensions.

[0095] The dynamic polar coordinate transformation module is used to construct a sliding window to traverse the normalized signal point by point, dynamically select the point with the minimum amplitude within the window as the pole, calculate the weighting coefficients by combining the local and global amplitude ranges, and obtain the polar radius and polar angle to realize the dynamic mapping of one-dimensional time domain signals to polar coordinate feature sequences.

[0096] The feature enhancement module is used to receive the polar coordinate feature sequence and the preprocessed one-dimensional induced voltage time-series signal, and perform adaptive weighting of the channel and spatial dimensions through the convolutional block attention module (CBAM) to enhance key morphological features and suppress redundant information, and output attention-enhanced feature map.

[0097] The morphology classification and state assessment module is used to receive attention-enhanced feature maps and output the specific morphology category of metal abrasive grains.

[0098] Through modular design, the functions of signal acquisition, dynamic polar coordinate transformation, feature enhancement and morphological classification are decoupled, which facilitates system integration and functional expansion. The introduction of the CBAM module further enhances the adaptive capability of feature enhancement, and the overall system has good real-time performance, deployability and recognition reliability. Attached Figure Description

[0099] Figure 1 This is a flowchart illustrating the metal abrasive particle morphology recognition method based on dynamic polar coordinate transformation and attention mechanism of the present invention.

[0100] Figure 2 This is a schematic diagram of a four-coil sensor for the metal abrasive particle morphology recognition system based on dynamic polar coordinate transformation and attention mechanism of the present invention.

[0101] Figure 3 This is an architecture diagram of the CS-ResNeXt block of the metal abrasive particle morphology recognition method based on dynamic polar coordinate transformation and attention mechanism of the present invention. Detailed Implementation

[0102] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0103] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0104] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0105] This invention discloses a method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism. The technical problem to be solved is that the weak features of abrasive particle morphology in the output signal of inductive sensor are easily masked by the global amplitude trend, and existing preprocessing methods are difficult to adaptively capture local dynamic fluctuations and are prone to losing temporal correlation, resulting in insufficient morphological feature extraction and low classification accuracy.

[0106] like Figure 1 As shown, the metal abrasive grain morphology recognition method based on dynamic polar coordinate transformation and attention mechanism includes the following steps:

[0107] S1 uses a fully excited four-coil inductive sensor to acquire the one-dimensional induced voltage timing signal of metal abrasive particles and performs preprocessing (normalization) to eliminate the influence of dimensions.

[0108] S2, a sliding window is constructed by combining the abrasive flow velocity and the effective magnetic field length of the coil. The pre-processed one-dimensional induced voltage timing signal is traversed point by point. The point with the minimum amplitude within the window is dynamically selected as the pole, and the weighting coefficient is calculated by combining the local and global amplitude ranges.

[0109] S3, based on the poles and weighting coefficients, the pole radius and pole angle are obtained, and the preprocessed one-dimensional induced voltage time sequence signal is nonlinearly reconstructed into a pole radius feature sequence and a pole angle feature sequence.

[0110] S4. The polar radius feature sequence and polar angle feature sequence, along with the preprocessed one-dimensional induced voltage time-series signal, are input into the attention-enhanced abrasive particle recognition model. The model first passes through an initial convolutional layer for basic feature extraction and dimensionality upscaling, and then splits into local and global branches for feature extraction. After the outputs of the two branches are aligned in size, they are spliced ​​and fused, and adaptive weighting of the channel and spatial dimensions is performed to obtain a dynamically polar coordinate fused attention-enhanced feature map (including local fluctuation features, peak-valley distribution features, and multi-peak time-series response features).

[0111] S5 inputs the attention-enhanced feature map fused with dynamic polar coordinates into the classifier, outputting the specific morphological category of the wear particles (such as spherical or sheet-like), and maps the wear pattern and operating status of the equipment based on the classification result.

[0112] In a preferred embodiment of the present invention, the one-dimensional induced voltage timing signal of the collected metal abrasive particles is preprocessed, specifically as follows:

[0113] The min-max method is used to process the original time series data as follows:

[0114] ,

[0115] in, This is the original timeline data. This is the original time series data; The timeline data is normalized. The normalized induced voltage amplitude data, , These are the minimum and maximum values ​​of the original timeline data, respectively. , The minimum and maximum amplitude values ​​of the original induced voltage time-series data are given by , and n is the index number of the current discrete data point. This process eliminates the influence of the dimensions of the original signal, avoids the interference of the absolute magnitude difference of amplitude caused by different sized abrasive grains on subsequent morphological feature extraction, and in particular overcomes the feature masking and calculation error problems caused by the significantly enhanced eddy current effect induced by the abrasive grains themselves as the grain size increases.

[0116] In a preferred embodiment of the present invention, the method for constructing a sliding window by combining the abrasive flow velocity and the effective magnetic field length of the coil is as follows:

[0117] To perform polar coordinate transformation on a normalized signal, the length of the dynamic sliding window must first be determined. Let the total length of the single-layer solenoid coil of the fully excited four-coil inductive sensor be... The controlled flow velocity of the metal abrasive particles passing through the sensor in the lubricating oil circuit is v, and the sampling frequency of the data acquisition is... The time period during which a single metal abrasive grain passes through a single-layer solenoid coil, causing an effective local magnetic induction intensity disturbance. for:

[0118] ,

[0119] To ensure the sliding window can completely cover the transient magnetic flux disturbance period caused by metal abrasive particles and avoid truncating local morphological distortion features, the length of the sliding window... (Total number of data points included) and window half-length ,for:

[0120] ,

[0121] ,

[0122] in, This indicates a floor operation, ensuring that the window length is the number of discrete data points that are valid integers;

[0123] Let the preprocessed one-dimensional induced voltage time series data sequence be... The corresponding timeline sequence is ,in The total length of the sequence is represented by data points. Create a sliding window centered on the element and add a boundary judgment mechanism:

[0124] ,

[0125] ,

[0126] in, This is the starting index of the sliding window. This serves as the termination index for the sliding window. A boundary check mechanism prevents the window from exceeding the data range, ensuring the validity of the window selection.

[0127] In a preferred embodiment of the present invention, the method for dynamically selecting the point with the minimum amplitude within the window as the pole by traversing the preprocessed one-dimensional induced voltage time-series signal point by point is as follows:

[0128] Within the effective sliding window corresponding to the current data point, each point is traversed and the data point with the smallest normalized amplitude is selected as the dynamic pole of the current window and used as the reference point for the current local interval to participate in subsequent feature calculations. (Pole index...) for:

[0129] ,

[0130] in, This indicates the location of the data point with the smallest amplitude within the sliding window; j is the sequence number of the preprocessed one-dimensional induced voltage time series data.

[0131] Normalized time axis coordinates of dynamic poles With normalized amplitude coordinates Calculate the current data point respectively The difference between the coordinates of the dynamic pole and the time axis and the coordinate difference in amplitude The time axis coordinate difference and the amplitude coordinate difference together characterize the time-amplitude deviation of the current data point relative to the local peak and valley reference point, providing basic variables for polar coordinate mapping.

[0132] The dynamic poles are updated as the sliding window position changes, so that the coordinate difference of the current data point relative to the dynamic poles reflects the local temporal fluctuation characteristics.

[0133] In a preferred embodiment of the present invention, to enhance local data features and avoid the masking of local features by global features, a weighting coefficient is introduced. The weighting coefficient is calculated based on the ratio of the amplitude range of the data within the window to the amplitude range of the global data. Both the amplitudes within the window and the global data are the original data amplitudes before normalization, ensuring consistency of data sources. The weighting coefficient is calculated by combining the local and global amplitude ranges to achieve adaptive adjustment of the feature representation intensity for different local intervals, as follows:

[0134] ,

[0135] in, This represents the range of original data amplitudes within the current sliding window, that is, the difference between the maximum and minimum values ​​of the data within the window. This represents the range of the entire original data amplitude. Since the output signal of the four coils forms two characteristic peaks in each of the positive and negative half-cycles, and the amplitudes of the four characteristic peaks are not symmetrical and equal, the introduction of this weighting coefficient essentially achieves adaptive amplification of the local induced voltage fluctuations caused by the sawtooth edges of irregular abrasive particles such as sheet-like particles, so as to fully integrate the non-equal amplitude multi-peak characteristics.

[0136] The weighting coefficient is calculated by combining the data amplitude range within the current sliding window with the data amplitude range of the entire one-dimensional induced voltage time-series signal. This weighting coefficient characterizes the relative change of the local fluctuation intensity within the current sliding window relative to the overall amplitude distribution and is used to adjust the calculation of the polar radius and polar angle.

[0137] In a preferred embodiment of the present invention, the polar radius feature sequence and the polar angle feature sequence are used to characterize the local fluctuation characteristics, peak-valley distribution characteristics, and multi-peak time-series response characteristics in a one-dimensional induced voltage time-series signal. This is based on the time axis coordinate difference between the current data point and the dynamic pole. Amplitude coordinate difference with weighting coefficients One-dimensional time-domain features are uniformly compressed and mapped to two-dimensional polar coordinate space to construct a joint representation of local fluctuation features, peak-valley distribution features, and multi-peak time-series response features, generating polar radii respectively. With polar angle ,for:

[0138] ,

[0139] in, This is the four-quadrant arctangent function, with a range of [0, 2π). Compared to a single arctangent function, this function does not require additional quadrant determination and can be directly determined based on... and The positive or negative value determines the quadrant in which the polar angle is located, effectively avoiding quadrant ambiguity in the polar angle calculation process.

[0140] Generate the corresponding polar radius sequences based on the polar radius formula and the polar angle formula, respectively. and polar sequence ,for:

[0141]

[0142]

[0143] in, and These represent the polar radius and polar angle generated by the nth data point in the preprocessed one-dimensional time-domain feature, respectively. .

[0144] Polar diameter sequence The change characterizes the joint deviation strength of data points relative to dynamic poles, including signal fluctuation characteristics within local intervals; polar angle sequence The changes in amplitude represent the relative directional relationship between data points and dynamic poles in the time-amplitude plane, including peak and valley distribution information. Through the above point-by-point dynamic pole finding and weighted mapping process, the original one-dimensional abrasive induced voltage signal is finally nonlinearly reconstructed into a polar diameter feature sequence and a polar angle feature sequence that accurately characterize the dynamic evolution of abrasive morphology.

[0145] In a preferred embodiment of the present invention, the backbone network of the attention-enhanced abrasive particle recognition model in step S4 adopts a dual-branch residual architecture, specifically as follows:

[0146] After the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time series signal are subjected to convolution, batch normalization, ReLU activation, and max pooling to complete the initial extraction of basic features and dimensionality increase, the signal is then split into local and global branches:

[0147] like Figure 3As shown, the local branch is composed of two CS-ResNeXt blocks embedded with Convolutional Block Attention Modules (CBAM) stacked in series. It is used to perform continuous convolution extraction while maintaining the spatial resolution of the feature map, so that the local details of the input signal can continue to be transmitted during the feature extraction process. The key responses are highlighted by CBAM weighting, and the input features are then fused through residual connections to enhance the local feature representation.

[0148] The global branch consists of two CS-ResNeXt blocks for downsampling, a one-dimensional transposed convolutional layer, and a CS-ResNeXt block for feature refinement. Specifically, the first two CS-ResNeXt blocks downsample the feature map twice consecutively using a convolution extraction method with a stride of 2, thereby compressing the feature map length step by step, expanding the network's receptive field, and capturing global contextual features. The global features obtained after downsampling are then used by the one-dimensional transposed convolutional layer to restore the spatial resolution of the feature map, ensuring that it maintains the same size as the output of the local branch.

[0149] The feature maps after resolution restoration are further input into the last CS-ResNeXt block for feature refinement to obtain the global branch output features. The CS-ResNeXt block is a residual bottleneck structure with attention enhancement mechanism, based on the ResNeXt architecture. (The main path sequentially includes the first convolutional layer, the first batch of normalization layers, the first ReLU activation layer, the second convolutional layer, the second batch of normalization layers, the second ReLU activation layer, a 1×1 convolutional layer, the third batch of normalization layers, the convolutional block attention module (CBAM), and the third ReLU activation layer. Basic features are extracted through two group convolutions, channels are normalized by a 1×1 convolution, and adaptive weighting is performed by embedding the convolutional block attention module (CBAM). Finally, the input features are fused through residual connections.

[0150] After the feature maps output by the local and global branches are aligned in size, they are spliced ​​and fused in the channel dimension by the Concat layer to achieve feature unification of the micro-contour of the abrasive grains and the macro-motion state. The fused features are transformed into a one-dimensional global feature vector by adaptive global average pooling, mapped to the classification space through a fully connected layer, and finally output as the classification result of the abrasive grain type by the Softmax activation function.

[0151] Instead of downsampling, the local branches are based on high-resolution feature maps. The local features of the abrasive induced voltage signal are refined by two grouped convolutions of the CS-ResNeXt block. This captures the subtle fluctuations, peak-valley transitions, and abrupt edge changes in the signal within the local area, avoiding the loss of local detail information due to downsampling.

[0152] Local branches maintain high spatial resolution of the feature map, preserve local temporal details of the wear-induced voltage signal, ensure accurate capture of weak signal features corresponding to local morphological differences, and provide basic data support for characterizing weak responses.

[0153] The global branch provides global perspective support for local features and provides global context for local feature extraction. Its first two CS-ResNeXt blocks expand the network's receptive field by progressively compressing the feature map length through continuous downsampling, capturing global context features of abrasive particle sensing signals, providing global trend background for local fine features, and enhancing the collaborative representation ability of local and global features.

[0154] One-dimensional transposed convolutional layers restore the spatial resolution of downsampled global features, aligning the global branch output with the local branch feature map size, thus ensuring the integrity of subsequent feature fusion.

[0155] The last CS-ResNeXt block with CBAM attention mechanism first extracts global basic features through two group convolutions, then regularizes the channels through 1×1 convolutions, and embeds a convolutional block attention module (CBAM) for adaptive weighting. This highlights the weights of weak response features corresponding to the local morphological differences of abrasive particles, suppresses irrelevant noise interference, and enhances the ability to represent weak responses.

[0156] In a preferred embodiment of the present invention, step S4 inputs the polar radius feature sequence and polar angle feature sequence, along with the preprocessed one-dimensional induced voltage timing signal, into the attention-enhanced abrasive particle recognition model, and performs adaptive weighting based on channel and spatial dimensions. The specific steps are as follows:

[0157] Feature enhancement is achieved using the model's embedded Convolutional Block Attention (CBAM) module, which involves the collaborative computation of channel attention and spatial attention sub-modules. Specifically:

[0158] Let the intermediate feature map obtained by feature extraction and dimensionality upscaling fusion of the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time-series signal through the convolutional layer inside the attention-enhanced abrasive particle recognition model be the CBAM input feature map, and let it be denoted as... The different channels of the feature map contain different response patterns to local fluctuation features, peak-valley distribution features, and multi-peak time-series response features. C, H, and W represent the number of channels, height, and width of the feature map, respectively. Global average pooling (GAP) and global max pooling (GMP) are performed on the input feature map. GAP calculates the average response intensity of each feature channel within the overall feature map range to characterize the overall distribution of the wear particle morphology-related features corresponding to that channel in the entire time-series feature map. GMP extracts the strongest response value in each feature channel to characterize the sensitivity of that channel to local significant wear particle morphology features, resulting in two dimensions. The channel feature vectors are denoted as follows: and :

[0159] ,

[0160] in, This represents the feature value of the c-th channel, h-th row, and w-th column of the input feature map; This represents the average response intensity of the c-th channel across the entire feature map, characterizing the overall response level of this channel to abrasive morphology-related modes. This represents the maximum response intensity in the c-th channel, describing the channel's ability to capture locally significant responses;

[0161] Will and The data are concatenated and input into a shared network consisting of two fully connected layers (FC) and an activation function to achieve feature compression and reconstruction. The statistical responses of each channel are converted into channel importance representations to characterize the contribution of different feature channels to the abrasive grain morphology recognition task. By comprehensively utilizing global statistical responses and local salient responses, the dependencies between different channels are learned. Finally, the channel attention weight vector is output through a sigmoid activation function. To achieve adaptive quantization of the contribution levels of different feature channels, the following is provided:

[0162] ,

[0163] in, ReLU represents the Sigmoid activation function, and ReLU represents the rectified linear activation function. and These represent two fully connected layers, used to compress and restore the feature dimension. This indicates channel concatenation of two pooled feature vectors; Adaptive recalibration is performed on the abrasive morphology discrimination information carried by different channels in the input feature map, which enhances the response of channels that are more relevant to abrasive morphology recognition, while suppressing channels with strong noise, redundant information and weak correlation response, thereby improving the model's ability to represent the differences in abrasive signals of different morphologies.

[0164] Channel attention weight vector The input feature map X is multiplied channel by channel to enhance the feature channel responses related to abrasive grain morphology recognition, suppress redundant channel responses that contribute little to recognition, and highlight the separability of local fluctuation features, peak-valley distribution features, and multi-peak temporal response features in the channel dimension, resulting in a weighted feature map enhanced by the channel attention mechanism. ;

[0165] The channel attention mechanism can automatically learn and extract effective features from the deep local features of the data, dynamically suppress redundant channels that are greatly affected by oil flow rate, environmental fluctuations, etc., and give higher weights to the mapping channels that are sensitive to changes in abrasive particle morphology and voltage peak size.

[0166] Spatial attention submodule: Feature map weighted by channel attention As input, Global average pooling and global max pooling are performed separately along the channel dimension. Average pooling is used to calculate the average response intensity of each time position across all feature channels to characterize the overall activation level of wear-grain morphology-related features at that position. Max pooling is used to extract the strongest response value of each time position across all feature channels to characterize the prominence of local significant wear-grain morphology features at that position, resulting in two dimensions. Spatial feature maps, respectively denoted as and ,for:

[0167] ,

[0168] in, This represents the feature value of the c-th channel, h-th row, and w-th column of the channel-weighted feature map.

[0169] Will and By concatenating the channels, we obtain the dimension as follows: The feature map is then compressed to 1 channel number through a convolutional layer (Conv), and the spatial attention weight map is output after passing through a sigmoid activation function. ,for:

[0170] ,

[0171] Where Conv represents the convolution operation, the spatial attention weight map can characterize the contribution of different temporal positions to the recognition of abrasive grain morphology, and the positions with larger weights correspond to key positions where local fluctuation features, peak and valley distribution features and multi-peak temporal response features are more significant.

[0172] Spatial attention weight map Feature map weighted by channel Element-wise multiplication is performed to enhance key positional feature responses in the multi-peak time-series response and suppress positional responses that contribute less to abrasive grain morphology recognition, resulting in the final attention-enhanced feature map. .

[0173] The attention-enhanced abrasive grain recognition model adaptively weights the input features in both channel and spatial dimensions through a convolutional block attention module. Specifically, the channel attention submodule generates a channel attention weight vector based on the input feature map and weights different feature channels to enhance the response of feature channels related to abrasive grain morphology recognition. The spatial attention submodule generates a spatial attention weight map based on the channel-weighted feature map and weights features at different locations to enhance the response of key location features in the time-series signal, thereby improving the model's ability to represent different abrasive grain morphology features.

[0174] The spatial attention mechanism is mainly used to compress the dimensionality of the temporal feature map of the convolution output, which strengthens the salient features in the signal and weakens the noise, while also enhancing the robustness of the model to small temporal shifts and deformations. Its adaptive weighted guidance model focuses attention on the boundary region of the four temporal and polarity alternating electromagnetic disturbances caused when particles enter the same group of mutually inductive coupled coils and the opposite magnetic field region, suppressing the information of the zero magnetic field equilibrium region when there is no metal particle interference.

[0175] In a preferred embodiment of the present invention, step S5 inputs the attention-enhanced feature map fused by dynamic polar coordinates into the classifier and outputs the specific morphological category of the abrasive particles. The specific steps are as follows:

[0176] The attention-enhanced feature map is passed sequentially through a global average pooling layer and a fully connected layer to reduce and integrate the high-dimensional features, thereby obtaining a one-dimensional feature vector for abrasive grain morphology discrimination.

[0177] The integrated one-dimensional feature vector is input into a classifier with a softmax activation function to calculate the probability distribution of the current metal abrasive particles belonging to different morphological categories, as follows:

[0178] ,

[0179] in, The current metal abrasive particles belong to the first... The probability of each morphological category. The first output of the fully connected layer The feature logarithmic value of each morphological category For category indexing, The total number of morphological categories, It is a natural constant;

[0180] The category label with the highest confidence in the probability distribution is extracted as the final morphology recognition result of the metal abrasive particles.

[0181] The present invention also provides a metal abrasive particle morphology recognition system based on dynamic polar coordinate transformation and attention mechanism, including a front-end acquisition device, a metal abrasive particle morphology recognition device, and a result display device.

[0182] The pre-acquisition device is used to acquire the one-dimensional induced voltage timing signal of metal abrasive particles, such as... Figure 2 As shown, the front-end acquisition device includes a fully excited four-coil inductive sensor. The fully excited four-coil inductive sensor includes four multilayer solenoid coils with completely identical parameters. The four coils are arranged coaxially and closely, and are divided into four bridge arm coils (a), (b), (c), and (d) along the axial direction. Coil (a) and coil (b) form one group, and coil (c) and coil (d) form another group. The two groups of coils are wound in opposite directions, and the spacing between adjacent coils is equal, forming a symmetrical Wheatstone bridge arm.

[0183] The number of turns, axial length, inner and outer radii, wire material and winding density of the four coils are completely consistent, ensuring that when there is no interference from metal particles, the self-inductance, resistance and mutual inductance parameters of each coil are strictly symmetrical, forming a zero magnetic field balance region.

[0184] The fully excited four-coil inductive sensor includes four multilayer solenoid coils with identical parameters. These four coils are coaxially and closely arranged outside an insulating oil pipe frame, forming a hollow internal fluid detection channel. The four coils are sequentially divided into a first bridge arm coil, a second bridge arm coil, a third bridge arm coil, and a fourth bridge arm coil along the axial direction, forming a symmetrical Wheatstone bridge circuit. The first and second bridge arm coils form the first group, and the third and fourth bridge arm coils form the second group. Two adjacent coils within the same group are wound in the same direction to create positive mutual inductance enhancement. The coils in the first group and the second group are wound in opposite directions. Metal abrasive particles pass sequentially through the fluid detection channel along the axial direction with the lubricating oil, causing alternating changes in the effective inductance of each bridge arm coil, and differentially outputting induced voltage signals through the Wheatstone bridge circuit.

[0185] By employing a symmetrical coil layout and a differential signal detection mechanism, common-mode noise can be effectively offset and the effects of environmental fluctuations can be compensated in real time, significantly improving detection sensitivity and stability. At the same time, its bridge topology further optimizes energy transmission efficiency and enhances signal response capability.

[0186] The input end of the metal abrasive particle morphology recognition device is connected to the output end of the pre-acquisition device. The metal abrasive particle morphology recognition device executes the method described in this invention and outputs the specific morphology category of the abrasive particles to the result display device for display.

[0187] In a preferred embodiment of the present invention, the metal abrasive particle morphology identification device includes:

[0188] The signal acquisition and preprocessing module is used to receive the one-dimensional induced voltage timing signal of metal abrasive particles and to perform normalization processing to eliminate the influence of dimensions.

[0189] The dynamic polar coordinate transformation module is used to construct a sliding window to traverse the normalized signal point by point, dynamically select the point with the minimum amplitude within the window as the pole, calculate the weighting coefficients by combining the local and global amplitude ranges, and obtain the polar radius and polar angle to realize the dynamic mapping of the one-dimensional time domain signal to the polar coordinate feature sequence.

[0190] The feature enhancement module receives the polar coordinate feature sequence and the preprocessed one-dimensional induced voltage time-series signal, performs adaptive weighting of the channel and spatial dimensions through the convolutional block attention module (CBAM), enhances key morphological features and suppresses redundant information, and outputs an attention-enhanced feature map.

[0191] The morphology classification and state assessment module is used to receive attention-enhanced feature maps and output the specific morphology category of metal abrasive grains.

[0192] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.

[0193] In the embodiments of this application, terms such as "fixed," "fixed connection," and "fixed connection" refer to common fixing methods in the prior art, such as welding, riveting, and screws. "Rotary connection" refers to common rotary connection methods in the prior art, such as hinges and bearing rotation. If electrical components are provided, the functions, control, and power supply methods of all electrical components are common technical means in the prior art. This application has not improved them and they are not within the protection scope of this application. Therefore, this application will not elaborate on them.

[0194] Furthermore, the selection of materials and strength limitations for all components in this application can be made and arranged by those skilled in the art based on the site environment and the requirements of relevant national or industry standards, and are not within the scope of protection of this application. Therefore, this application will not elaborate on these points.

Claims

1. A method for recognizing the morphology of metal abrasive particles based on dynamic polar coordinate transformation and attention mechanism, characterized in that, Includes the following steps: S1, using a fully excited four-coil inductive sensor, acquires the one-dimensional induced voltage timing signal of metal abrasive particles and performs preprocessing; S2, a sliding window is constructed by combining the abrasive flow velocity and the effective magnetic field length of the coil. The pre-processed one-dimensional induced voltage timing signal is traversed point by point. The point with the minimum amplitude within the window is dynamically selected as the pole, and the weighting coefficient is calculated by combining the local and global amplitude ranges. S3, based on the poles and weighting coefficients, the pole radius and pole angle are obtained, and the preprocessed one-dimensional induced voltage time sequence signal is nonlinearly reconstructed into a pole radius feature sequence and a pole angle feature sequence. S4. The polar radius feature sequence and polar angle feature sequence, together with the preprocessed one-dimensional induced voltage time series signal, are input into the attention-enhanced abrasive particle recognition model. The model first passes through the initial convolutional layer for basic feature extraction and dimension upscaling, and then splits into local and global branches for feature extraction. After the outputs of the two branches are aligned in size, they are spliced ​​and fused. Adaptive weighting of channel and spatial dimensions is performed to obtain the attention-enhanced feature map with dynamic polar coordinate fusion. S5 inputs the attention-enhanced feature map fused with dynamic polar coordinates into the classifier, and outputs the specific morphological category of the abrasive grains.

2. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 1, characterized in that, The one-dimensional induced voltage timing signal of the collected metal abrasive particles is preprocessed as follows: The min-max method is used to process the original time series data as follows: ， in, This is the original timeline data. This is the original time series data; The timeline data is normalized. The normalized induced voltage amplitude data, , These are the minimum and maximum values ​​of the original timeline data, respectively. , The minimum and maximum amplitude values ​​are the original induced voltage time series data, and n is the index number of the current discrete data point.

3. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 1, characterized in that, The method for constructing a sliding window by combining the abrasive flow velocity and the effective magnetic field length of the coil is as follows: Let the total length of the single-layer solenoid coil of the fully excited four-coil inductive sensor be... The controlled flow rate of metal abrasive particles passing through the sensor in the lubricating oil circuit is: The sampling frequency of data acquisition is The time period during which a single metal abrasive grain passes through a single-layer solenoid coil, causing an effective local magnetic induction intensity disturbance. for: ， Sliding window length (Total number of data points included) and window half-length ,for: ， ， in, This indicates a floor operation, ensuring that the window length is the number of discrete data points that are valid integers; Let the preprocessed one-dimensional induced voltage time series data sequence be... The corresponding timeline sequence is ,in The total length of the sequence is represented by data points. Create a sliding window centered on the element and add a boundary judgment mechanism: ， ， in, This is the starting index of the sliding window. This is the ending index of the sliding window.

4. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 3, characterized in that, The method for dynamically selecting the point with the minimum amplitude within the window as the pole by iterating through the preprocessed one-dimensional induced voltage time-series signal point by point is as follows: Within the effective sliding window corresponding to the current data point, each point is traversed and the data point with the smallest normalized amplitude is selected as the dynamic pole of the current window and used as the reference point for the current local interval to participate in subsequent feature calculations. (Pole index...) for: ， in, This indicates the location of the data point with the smallest amplitude within the sliding window; j is the sequence number of the preprocessed one-dimensional induced voltage time series data. Normalized time axis coordinates of dynamic poles With normalized amplitude coordinates Calculate the current data point respectively The difference between the coordinates of the dynamic pole and the time axis and the coordinate difference in amplitude The time axis coordinate difference and the amplitude coordinate difference together characterize the time-amplitude deviation of the current data point relative to the local peak and valley reference point.

5. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 4, characterized in that, The weighting coefficients are calculated by combining the local and global amplitude ranges, as follows: ， in, This represents the range of original data amplitudes within the current sliding window, that is, the difference between the maximum and minimum values ​​of the data within the window. It represents the range of the entire original data amplitude.

6. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 5, characterized in that, Based on the time axis coordinate difference of the current data point relative to the dynamic pole Amplitude coordinate difference with weighting coefficients One-dimensional time-domain features are uniformly compressed and mapped to two-dimensional polar coordinate space to generate polar radii. With polar angle ,for: ， in, It is the arctangent function in the fourth quadrant, and its range is [0, 2π). Generate the corresponding polar radius sequences based on the polar radius formula and the polar angle formula, respectively. and polar sequence ,for: ， ， in, and These represent the polar radius and polar angle generated by the nth data point in the preprocessed one-dimensional time-domain feature, respectively. .

7. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 1, characterized in that, The backbone network of the attention-enhanced abrasive particle recognition model in step S4 adopts a dual-branch residual architecture, specifically as follows: After the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time series signal are subjected to convolution, batch normalization, ReLU activation, and max pooling to complete the initial extraction of basic features and dimensionality increase, the signal is then split into local and global branches: The local branch is composed of two CS-ResNeXt blocks embedded with Convolutional Block Attention (CBAM) stacked in series. It is used to perform continuous convolution extraction while maintaining the spatial resolution of the feature map. This allows the local details of the input signal to continue to be transmitted during the feature extraction process. The key responses are highlighted by CBAM weighting, and the input features are then fused through residual connections to enhance the local feature representation. The global branch includes two CS-ResNeXt blocks for downsampling, a one-dimensional transposed convolutional layer, and a CS-ResNeXt block for feature refinement. The first two CS-ResNeXt blocks perform two consecutive downsampling operations on the feature map using a convolution extraction method with a stride of 2 to capture global context features. The global features obtained after downsampling are then passed through a one-dimensional transposed convolutional layer to restore the spatial resolution of the feature map, so that it is consistent with the size of the local branch output; The feature map after resolution restoration is input into the last CS-ResNeXt block for feature refinement, resulting in the global branch output features. The CS-ResNeXt block is a residual bottleneck structure with an attention enhancement mechanism. Based on the ResNeXt architecture, the main path sequentially includes the first convolutional layer, the first batch of normalization layers, the first ReLU activation layer, the second convolutional layer, the second batch of normalization layers, the second ReLU activation layer, a 1×1 convolutional layer, the third batch of normalization layers, the convolutional block attention module (CBAM), and the third ReLU activation layer. Basic features are extracted through two group convolutions, channels are normalized by 1×1 convolutions, and adaptive weighting is performed by embedding the convolutional block attention module (CBAM). Finally, the input features are fused through residual connections. After the feature maps output by the local branch and the global branch are aligned in size, they are spliced ​​and fused in the channel dimension by the Concat layer to achieve feature unification of the micro-contour of the abrasive particles and the macro-motion state. The fused features are transformed into a one-dimensional global feature vector through adaptive global average pooling, mapped to the classification space through a fully connected layer, and finally output as a classification result of abrasive grain type through the Softmax activation function.

8. The method for metal abrasive particle morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 7, characterized in that, Step S4 inputs the polar radius feature sequence and polar angle feature sequence, along with the preprocessed one-dimensional induced voltage time-series signal, into the attention-enhanced abrasive particle recognition model, and performs adaptive weighting based on channel and spatial dimensions. The specific steps are as follows: Feature enhancement is achieved using the model's embedded Convolutional Block Attention (CBAM) module, which involves the collaborative computation of channel attention and spatial attention sub-modules. Specifically: Let the intermediate feature map obtained by feature extraction and dimensionality upscaling fusion of the polar radius feature sequence, polar angle feature sequence, and preprocessed one-dimensional induced voltage time-series signal through the convolutional layer inside the attention-enhanced abrasive particle recognition model be the CBAM input feature map, and let it be denoted as... C, H, and W represent the number of channels, height, and width of the feature map, respectively. Global average pooling (GAP) and global max pooling (GMP) are performed on the input feature map to obtain two dimensions. The channel feature vectors are denoted as follows: and : ， in, This represents the feature value of the c-th channel, h-th row, and w-th column of the input feature map; This represents the average response intensity of the c-th channel across the entire feature map, characterizing the overall response level of this channel to abrasive morphology-related modes. This represents the maximum response intensity in the c-th channel, describing the channel's ability to capture locally significant responses; Will and The data is concatenated and fed into a shared network consisting of two fully connected layers (FC) and an activation function to achieve feature compression and reconstruction. It comprehensively utilizes global statistical responses and local salient responses to learn the dependencies between different channels. Finally, it outputs channel attention weight vectors through a sigmoid activation function. ,for: ， in, ReLU represents the Sigmoid activation function, and ReLU represents the rectified linear activation function. and These represent two fully connected layers, used to compress and restore the feature dimension. This indicates channel concatenation of two pooled feature vectors; Channel attention weight vector The input feature map X is multiplied channel by channel to obtain the weighted feature map enhanced by the channel attention mechanism. ; right Perform global average pooling and global max pooling along the channel dimension respectively to obtain two dimensions. Spatial feature maps, respectively denoted as and ,for: ， in, This represents the feature value of the c-th channel, h-th row, and w-th column of the channel-weighted feature map. Will and By concatenating the channels, we obtain the dimension as follows: The feature map is then compressed to 1 channel number through a convolutional layer (Conv), and the spatial attention weight map is output after passing through a sigmoid activation function. ,for: ， Where Conv represents the convolution operation; Spatial attention weight map Feature map weighted by channel Element-wise multiplication is performed to obtain the final attention-enhanced feature map. .

9. The method for metal abrasive grain morphology recognition based on dynamic polar coordinate transformation and attention mechanism according to claim 1, characterized in that, Step S5 inputs the attention-enhanced feature map fused by dynamic polar coordinates into the classifier, and outputs the specific morphological category of the abrasive grains. The specific steps are as follows: The attention-enhanced feature map is passed sequentially through a global average pooling layer and a fully connected layer to reduce the dimensionality of the high-dimensional features and integrate them to obtain a one-dimensional feature vector for abrasive grain morphology discrimination. The integrated one-dimensional feature vector is input into a classification layer with a softmax activation function to calculate the probability distribution of the current metal abrasive particles belonging to different morphological categories, as follows: ， in, The current metal abrasive particles belong to the first... The probability of each morphological category. The first output of the fully connected layer The feature logarithmic value of each morphological category For category indexing, The total number of morphological categories, It is a natural constant; The category label with the highest confidence in the probability distribution is extracted as the final morphology recognition result of the metal abrasive particles.

10. A metal abrasive grain morphology recognition system based on dynamic polar coordinate transformation and attention mechanism, characterized in that, It includes a pre-acquisition device, a metal abrasive particle morphology recognition device, and a result display device; The pre-acquisition device is used to acquire the one-dimensional induced voltage timing signal of metal abrasive particles. The pre-acquisition device includes a fully excited four-coil inductive sensor, which includes four multilayer solenoid coils with completely identical parameters. The four coils are coaxially and closely arranged outside the insulating oil pipe skeleton, forming a hollow fluid detection channel. The four coils are sequentially divided into a first bridge arm coil, a second bridge arm coil, a third bridge arm coil, and a fourth bridge arm coil along the axial direction, forming a symmetrical Wheatstone bridge circuit. The first bridge arm coil and the second bridge arm coil form the first group, and the third bridge arm coil and the fourth bridge arm coil form the second group. Two adjacent coils in the same group are wound in the same direction to form a positive reinforcement of mutual inductance. The coils in the first group and the coils in the second group are wound in opposite directions. The metal abrasive particles pass through the fluid detection channel along the axial direction with the lubricating oil, causing alternating changes in the effective inductance of each bridge arm coil, and differentially outputting the induced voltage signal through the Wheatstone bridge circuit. The input end of the metal abrasive particle morphology recognition device is connected to the output end of the pre-acquisition device. The metal abrasive particle morphology recognition device executes the method described in any one of claims 1-9 and outputs the specific morphology category of the abrasive particles to the result display device for display.

11. The metal abrasive grain morphology recognition system based on dynamic polar coordinate transformation and attention mechanism according to claim 10, characterized in that, The metal abrasive grain morphology recognition device includes: The signal acquisition and preprocessing module is used to receive the one-dimensional induced voltage timing signal of metal abrasive particles and to perform normalization processing to eliminate the influence of dimensions. The dynamic polar coordinate transformation module is used to construct a sliding window to traverse the normalized signal point by point, dynamically select the point with the minimum amplitude within the window as the pole, calculate the weighting coefficients by combining the local and global amplitude ranges, and obtain the polar radius and polar angle to realize the dynamic mapping of one-dimensional time domain signals to polar coordinate feature sequences. The feature enhancement module is used to receive the polar coordinate feature sequence and the preprocessed one-dimensional induced voltage time-series signal, and perform adaptive weighting of the channel and spatial dimensions through the convolutional block attention module (CBAM) to enhance key morphological features and suppress redundant information, and output attention-enhanced feature map. The morphology classification and state assessment module is used to receive attention-enhanced feature maps and output the specific morphology category of metal abrasive grains.

Images