A method and system for identifying gearbox health status stages based on oil characteristics

Abstract

This invention belongs to the field of mechanical equipment wear condition monitoring technology. It proposes a method and system for identifying gearbox health status stages based on oil characteristics. The method involves preprocessing acquired oil wear particle images, extracting oil features from the preprocessed images, and then obtaining gearbox health status classification results based on these oil features and a pre-trained long short-term memory (LSTM) neural network model. The training of the LSM model includes: fusing oil features from each oil wear particle image sample to obtain a one-dimensional health factor; obtaining a health index sequence based on the one-dimensional health factor of each oil wear particle image sample; performing breakpoint detection on the health index sequence to obtain three breakpoints; and adding health status labels to all oil wear particle image samples based on the obtained breakpoints. This invention can autonomously and accurately classify wear stages and capture evolution trends, ultimately forming a decision-making judgment that surpasses human experience.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment wear condition monitoring technology, specifically to a method and system for identifying gearbox health condition stages based on oil characteristics. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Oil characteristic analysis is an indispensable technical means for condition monitoring and health management of wind turbine gearboxes. Under the harsh operating environment of wind turbine generators, which are subjected to variable loads and strong impacts for extended periods, the gearbox undergoes a progressive evolution from a healthy state to early wear and finally severe failure. The ability to accurately identify the entire lifecycle health stage through oil characteristics, especially to provide timely warnings of early wear, directly impacts the timeliness of fault prevention and control and maintenance costs. Failure to identify early conditions can lead to rapid wear escalation, triggering catastrophic failures, causing unplanned unit shutdowns and significant economic losses.

[0004] Currently, wear condition identification methods based on oil analysis still have significant technical shortcomings. At the feature level, most methods still rely on single macroscopic indicators such as abrasive particle concentration or fixed thresholds for judgment. While these indicators can reflect the "quantitative change" in the degree of wear, they cannot effectively identify the "qualitative change" in the wear mechanism. For example, they are difficult to distinguish between different failure modes such as cutting wear, fatigue spalling, and corrosion oxidation, resulting in insufficient sensitivity in identifying complex early failures. At the model level, existing methods mostly rely on fixed rules set by expert experience, failing to fully consider the dynamic coupling relationship between multiple oil features and their temporal evolution. This makes the models have poor generalization ability under actual complex operating conditions, blurry identification boundaries for the "health-early wear" critical state, significant lag in early warning, and diagnostic results that are greatly affected by subjectivity, making it difficult to effectively transfer between different wind farms or units. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for identifying gearbox health status stages based on oil characteristics. This avoids situations where equipment maintenance timing is not accurately determined due to single characteristics or human error, improves the automation level and accuracy of equipment health status identification, and makes the division of health stages more consistent with the actual wear evolution process of the equipment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for identifying the health status stage of a gearbox based on oil characteristics.

[0008] A method for identifying gearbox health status stages based on oil characteristics includes the following process:

[0009] The acquired oil abrasive images are preprocessed, and oil features are extracted based on the preprocessed oil abrasive images. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0010] Based on the characteristics of the oil and a pre-trained long short-term memory neural network model, the gearbox health status classification results are obtained; the training of the long short-term memory neural network model includes:

[0011] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0012] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0013] Secondly, the present invention provides a gearbox health status stage identification system based on oil characteristics.

[0014] A gearbox health status stage identification system based on oil characteristics, comprising:

[0015] The oil feature extraction unit is configured to: preprocess the acquired oil abrasive image, and extract oil features based on the preprocessed oil abrasive image. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0016] The health status classification unit is configured to: obtain the gearbox health status classification result based on oil characteristics and a pre-trained long short-term memory neural network model; wherein, the training of the long short-term memory neural network model includes:

[0017] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0018] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0019] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium;

[0020] A processor, adapted to execute computer programs;

[0021] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the gearbox health status stage identification method based on oil characteristics according to the first aspect of the present invention.

[0022] Fourthly, the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the method for identifying gearbox health status stages based on oil characteristics according to the first aspect of the present invention.

[0023] Fifthly, the present invention provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the gearbox health status stage identification method based on oil characteristics according to the first aspect of the present invention.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This invention employs a comprehensive approach encompassing oil abrasive image preprocessing, multi-dimensional oil feature extraction, pre-trained long short-term memory neural network model, oil feature fusion to generate a one-dimensional health factor, and health index sequence breakpoint detection with added health status labels. This method combines oil feature analysis with a temporal deep learning model, achieving automated labeling of sample health status through breakpoint detection. This solution addresses the problem of traditional oil monitoring relying solely on single abrasive features, making it difficult to accurately correspond to different equipment health stages. It fills the gap in the correlation between multiple oil features and the entire health cycle of equipment, overcoming the inefficiencies of manual oil abrasive identification, the significant influence of subjective experience on judgment results, and the inability of ordinary machine learning models to capture the temporal changes in oil features. This allows the model to better adapt to the dynamic process of equipment wear, improving the automation and accuracy of equipment health status identification. The division of health stages more closely aligns with the actual wear evolution process of the equipment, while also improving the utilization efficiency of oil monitoring data. Furthermore, it avoids inaccurate timing of equipment maintenance due to single features or human error, reducing resource waste from over-maintenance and the risk of equipment failure caused by untimely maintenance.

[0026] This invention employs a weighted fusion method based on multiple oil features. By calculating the weights of each feature in each oil wear particle sample through index normalization, it integrates multi-dimensional oil features such as wear rate and the proportion of various wear particles into a single-dimensional health factor that can uniformly characterize the equipment's health status. This solves the problems of fragmented and dispersed multi-dimensional oil features, making it difficult to comprehensively reflect the equipment's health status. It transforms complex multi-feature data into a single quantitative indicator that can be directly used for subsequent time-series analysis. This overcomes the shortcomings of information redundancy and unreasonable weight allocation between features when multiple features are input into the model in parallel, as well as the limitation that a single oil feature cannot fully cover different dimensions of equipment wear. It ensures that the contribution of each feature is reasonably reflected. This method improves the representation efficiency of oil features, making the subsequently constructed health index sequence more concise and containing complete equipment health information. It also enhances the correlation between features and the actual health status of the equipment, making the feature's reflection of health status more accurate. Furthermore, it avoids the problems of information fragmentation and inability to comprehensively assess equipment health when multiple features are analyzed independently, ensuring the reliability and comprehensiveness of subsequent health status analysis.

[0027] This invention employs a binary segmentation method based on L2 loss. It treats the health index sequence as a complete segment, pre-segmenting it by traversing all possible segmentation positions. By calculating the sum of the L2 losses of the two segments after segmentation, the position with the minimum loss is selected as the optimal breakpoint. This process is repeated until three breakpoints that can delineate the health stages are obtained. This solves the problem of accurately locating breakpoints for different health stages in the health index sequence, achieving clear boundaries between stages within the equipment's health cycle. It overcomes the shortcomings of traditional breakpoint detection methods, such as sensitivity to local fluctuations in the sequence and unstable segmentation results, as well as the lack of quantitative basis and strong subjectivity in manual breakpoint division. It provides a clear objective standard for breakpoint selection, improving the objectivity and accuracy of breakpoint detection. This makes the boundaries of health stages more closely reflect the actual changes in oil characteristics with equipment wear, while also increasing the automation level of health state classification and reducing the cost of manual intervention. Furthermore, this invention avoids inaccurate sample health state labels caused by ambiguous breakpoint division, providing reliable label data for subsequent training of long short-term memory neural network models and ensuring the model's recognition performance.

[0028] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1A flowchart illustrating a gearbox health status stage identification method based on oil characteristics, provided as an exemplary embodiment of the present invention;

[0031] Figure 2 A schematic diagram of image preprocessing provided as an exemplary embodiment of the present invention;

[0032] Figure 3 A schematic diagram of an LSTM health status assessment model provided as an exemplary embodiment of the present invention; wherein, , , ··· They represent the 1st, 2nd, 3rd...th respectively The input consists of oil features at each time step, where Layer1 represents the first layer of the LSTM. , , ··· Don't represent Layer 1 in the 1st, 2nd, 3rd... The hidden state at each time step; Layer2 represents the second layer of the LSTM; , , ··· , , The 16 components represent the expanded feature dimensions; Represents Layer 2 in the The final hidden state at each time step;

[0033] Figure 4 A schematic diagram of concentration evolution trend provided for an exemplary embodiment of the present invention;

[0034] Figure 5 A schematic diagram illustrating the evolution trend of the proportion of large abrasive particles provided as an exemplary embodiment of the present invention;

[0035] Figure 6 A schematic diagram illustrating the evolution trend of the proportion of cutting abrasive particles provided as an exemplary embodiment of the present invention;

[0036] Figure 7 A schematic diagram illustrating the evolution trend of fatigue wear particle proportion provided for an exemplary embodiment of the present invention;

[0037] Figure 8 A schematic diagram illustrating the evolution trend of the proportion of oxide abrasive particles provided as an exemplary embodiment of the present invention;

[0038] Figure 9 A schematic diagram of the training data health status stages provided in an exemplary embodiment of the present invention;

[0039] Figure 10 An oil feature fusion confusion matrix provided as an exemplary embodiment of the present invention;

[0040] Figure 11 A single concentration feature confusion matrix is ​​provided as an exemplary embodiment of the present invention;

[0041] Figure 12 The confidence distribution of prediction results provided as an exemplary embodiment of the present invention;

[0042] Figure 13 The predicted probabilities at each stage are provided as an exemplary embodiment of the present invention;

[0043] Figure 14 A schematic diagram of a gearbox health status stage identification system based on oil characteristics, provided as an exemplary embodiment of the present invention;

[0044] Figure 15 A schematic diagram of a computer device provided for an exemplary embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0047] As a core component of wind turbines, the gearbox of a wind turbine is prone to wear failure due to long-term high-load operation, leading to huge economic losses. Therefore, effective monitoring of its wear condition is crucial. Among various monitoring methods, oil wear condition monitoring, by analyzing abrasive particles in the lubricating oil, is like performing a "blood test" on the gearbox. It can directly and proactively diagnose early abnormal wear, making it a key technology for achieving predictive maintenance and ensuring the economic benefits of wind farms. Currently, oil wear condition monitoring of wind turbine gearboxes faces two major technical bottlenecks: First, traditional methods rely on single macroscopic indicators such as the abrasive coverage area index (IPCA) and the number of large abrasive particles (LPQ), which only reflect the "quantitative change" of wear but cannot identify differences in wear mechanisms, thus limiting the accuracy of condition assessment. Second, diagnostic decisions still heavily rely on human experience, resulting in strong subjectivity in assessment and delayed early warning, making it difficult to meet the needs of intelligent operation and maintenance. To address the aforementioned issues, this invention proposes a data-driven intelligent assessment model based on multi-dimensional features. This model simultaneously extracts five types of features from abrasive grain images: wear rate, wear degree, cutting index, fatigue index, and oxidation index, achieving multi-dimensional feature fusion assessment. Based on this, an intelligent assessment model is constructed to automatically fuse and diagnose multi-dimensional features, replacing the traditional judgment mode that relies on human experience. This forms an objective, accurate, and automatically operable health status assessment system, providing reliable technical support for predictive maintenance of wind turbine gearboxes. Specifically, for example... Figure 1 As shown, the process includes the following:

[0048] S101: Data Acquisition.

[0049] Through a full-life-cycle wear test of a wind turbine gearbox, real-time monitoring of different wear states of the gearbox was achieved using a high-throughput online oil abrasive image monitoring sensor. This sensor collects real-time oil abrasive information of the gearbox under various wear states, covering key aspects such as abrasive-related characteristics and oil auxiliary properties, ultimately forming a complete full-lifecycle oil dataset.

[0050] S102: Image preprocessing.

[0051] To extract the features of abrasive grains from the image, it is necessary to eliminate interference from other factors in the image and preprocess the abrasive grain image. The processing flow is as follows: Figure 2 As shown, the preprocessing includes the preprocessing of the abrasive annular image and the preprocessing of the dispersed abrasive image.

[0052] The original abrasive grain annular image and the abrasive grain annular image both contain information redundancy in the RGB three channels, and the color information has a weak correlation with key features such as abrasive grain morphology and concentration. Therefore, a weighted average method is first used to convert the image to grayscale, and the calculation formula is as follows:

[0053] (1);

[0054] in, , , These represent the red, green, and blue channel intensities, respectively, with weights reflecting the human eye's sensitivity to different colors. After grayscale processing, each pixel in the image is represented by only one grayscale value (0-255), making brighter areas easier to distinguish.

[0055] Abrasive grain images are often affected by background interference such as uneven lighting, equipment reflection, and oil flow, which easily introduce noise and affect the abrasive grain recognition effect. Therefore, a difference operation method is used to subtract the background image from the original image, thereby suppressing background interference and enhancing the saliency of the abrasive grain region. Since the abrasive grains themselves are dark gray, the gray-level difference with the background region is small, and their contours are still not clear enough in the gray-level image, which is not conducive to the extraction of morphological features. Therefore, the Otsu global threshold segmentation algorithm is further employed to automatically determine the optimal segmentation threshold, achieving separation of the foreground (abrasive grains) from the background.

[0056] (2);

[0057] In the formula, Pixel Value is the pixel grayscale value. The threshold is calculated by the algorithm. This step effectively removes the background, improves the recognizability of the abrasive grain contour, and lays the foundation for subsequent statistical extraction of features such as grain size and shape.

[0058] To improve the accuracy of feature extraction, further morphological processing was performed on the dispersed abrasive grain images based on the initial preprocessing. To address potential issues such as holes or discontinuities within the abrasive grains after binarization, dilation and erosion operations were sequentially applied to restore the integrity of the abrasive grain regions. Subsequently, based on the grayscale difference between the abrasive grain edges and the background, the contour of each abrasive grain was extracted and numbered, providing basic data for subsequent shape parameter analysis.

[0059] S103: Multidimensional feature construction.

[0060] To overcome the limitations of traditional wear monitoring methods that rely solely on macroscopic indicators to reflect quantitative changes and accurately identify wear mechanisms, this paper selects five key features from multiple dimensions and mechanisms to achieve a more comprehensive and accurate assessment of equipment wear conditions. These five features are: wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles. Through the synergistic analysis of these features, the changes in wear rate, severity, and the differentiated manifestations of different wear mechanisms can be effectively captured, thus providing a more mechanism-related basis for wear condition identification.

[0061] S103-1: Wear rate.

[0062] Wear rate is characterized by the actual area of ​​the abrasive grains. The abrasive coverage area index (IPCA) converts the material loss rate into a visible spatial coverage index by statistically analyzing the area of ​​all foreground pixels in the abrasive grain ring image. The foreground pixel coverage area can be obtained by counting all foreground pixels.

[0063] (3);

[0064] Wherein, FPCA represents the foreground pixel coverage area. The number of foreground pixels in the abrasive ring image, when the first... When the grayscale value of each pixel is 255, When the grayscale value is 1; when the grayscale value is 0, It is 0.

[0065] Further combining image pixel resolution This can be converted into actual physical area:

[0066] (4);

[0067] in, For height, For width.

[0068] S103-2: Proportion of large abrasive particles.

[0069] Abrasive grains with an equivalent circular diameter exceeding 30 μm are defined as large abrasive grains, and a significant increase in their number usually indicates an accelerated wear stage. The formula for calculating the proportion of large abrasive grains is:

[0070] (5);

[0071] in, The proportion of large abrasive particles For a large number of abrasive grains, This represents the total number of abrasive grains.

[0072] S103-3: Proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidation abrasive particles.

[0073] Besides normal sliding wear, the main wear mechanisms include cutting wear, fatigue wear, and oxidative wear, which are quantified by the proportion of each type of abrasive grain. First, the following morphological features need to be extracted as classification criteria:

[0074] Abrasive grain area A: The actual area of ​​a single abrasive grain based on pixel statistics.

[0075] (6);

[0076] in, The area of ​​a single abrasive grain. To disperse the number of individual abrasive grain pixels in the abrasive grain image, when the first... When the grayscale value of each pixel is 255, When the grayscale value is 1; when the grayscale value is 0, It is 0.

[0077] Calculate the equivalent circle diameter That is, the diameter of the circle with the same area as the abrasive grain:

[0078] (7);

[0079] Calculate the aspect ratio AR, which is the ratio of the long side to the short side of the smallest bounding rectangle of the abrasive grain. When AR≈1, the abrasive grain is approximately square; when AR>1, the abrasive grain is slender or flat.

[0080] (8);

[0081] Calculate roundness Describes the degree to which a shape approximates a circle; the roundness of an ideal circle. A value of 1 indicates that the abrasive grains are nearly perfectly round, indicating good roundness. A value less than 1 indicates that the abrasive grain shape deviates from a circle, and the closer it is to 0, the more irregular the shape.

[0082] (9);

[0083] The perimeter of the abrasive grain can be obtained from its profile.

[0084] Calculate the standard deviation of curvature It reflects the degree of fluctuation in the profile curvature and is used to identify spherical oxide abrasive particles.

[0085] Based on the above characteristics, a three-level discrimination tree model is established to classify all abrasive grains. The specific criteria are as follows: when the equivalent circle diameter is greater than or equal to the first set threshold and the roundness is greater than or equal to the second set threshold, it is a normal abrasive grain; otherwise, proceed to the next step of judgment. When the standard deviation of curvature is less than or equal to the third set threshold and the roundness is greater than or equal to the fourth set threshold, it is an oxidized abrasive grain; otherwise, proceed to the next step of judgment. When the aspect ratio is greater than or equal to the fifth set threshold and the roundness is less than or equal to the second set threshold, it is a cutting abrasive grain; otherwise, it is a fatigue sliding abrasive grain.

[0086] S104: Model building.

[0087] This model employs an LSTM (Long Short-Term Memory) architecture for wind turbine gearbox health status assessment. The health status assessment model consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. It effectively extracts temporal features and long-term dependencies from time-series data. The overall model architecture is as follows: Figure 3 As shown.

[0088] S104-1: Input layer.

[0089] Input data: Time series feature data after sliding window processing;

[0090] Data preprocessing: Feature normalization was performed using MinMaxScaler.

[0091] (10);

[0092] in, Represents the characteristics after normalization. Represents the characteristics before normalization. Represents the minimum value of the characteristic. Represents the maximum value of the characteristic.

[0093] S104-2: LSTM coding layer.

[0094] LSTM controls the flow of information through three gating units. The flow of information about oil characteristics in LSTM is as follows:

[0095] Forgetting gate: Feature importance screening determines which historical oil features are forgotten or retained from the cellular state.

[0096] (11)

[0097] in, Represents the forget gate weight matrix. The bias term representing the forget gate; represent The hidden state at any given moment; Represents the current moment The input feature vector; The representative will and Concatenate them into a single vector; Represents the sigmoid activation function; Represents the current moment The output of the forget gate;

[0098] Input gate: new feature information is taken in, determining which new information is stored in the cell state.

[0099] (12)

[0100] (13)

[0101] in, Represents the input gate weight matrix; A weight matrix representing the state of candidate cells; The bias term representing the input gate; Bias terms representing the candidate cell state; Represents the tanh activation function; Represents matrix multiplication; Represents the current moment Input gate output; Represents the current moment Candidate cell state;

[0102] Cell state update: Integrating historical oil characteristics with current characteristics to form long-term memory.

[0103] (14);

[0104] in, Represents element-wise multiplication; Represents the current moment cellular state; represent The state of a cell at any given moment.

[0105] Output gate: determines which information to output, and outputs the hidden state based on all current oil feature information.

[0106] (15);

[0107] in, The weight matrix representing the output gate; The bias term represents the output gate; Represents the current moment The output gate outputs;

[0108] S104-3: Feature mapping classification layer.

[0109] Temporal feature aggregation :

[0110] (16);

[0111] in, This represents the last time step of the LSTM model (i.e., the 1st time step). The hidden state of the last time step is taken as the compressed representation of the entire oil feature sequence.

[0112] Health status classification:

[0113] (17);

[0114] (18);

[0115] in, The weight matrix representing the fully connected classification layer; The bias term represents the fully connected classification layer; represent Activation function; The predicted probability distribution represents the health status, and finally the health status corresponding to the input oil feature data is classified and output (four categories in total).

[0116] S105: Model Training and Optimization.

[0117] S105-1: Processing of training data health status labels.

[0118] This invention proposes a method based on attention mechanism and time series segmentation, which is used to perform stage segmentation and labeling on the feature set of training data throughout the entire life cycle.

[0119] First, we need to use the attention mechanism fusion method to fuse oil features and construct a one-dimensional health factor. The core of the attention mechanism is to learn the "importance weight" for each feature dimension to ensure that the contribution of key features is higher.

[0120] Weight vector The generation is achieved through two steps ( (Representing the set of real numbers)

[0121] Step 1: Calculate the original attention score.

[0122] By performing a nonlinear transformation on the input features, the original attention score for each feature is obtained. The formula is:

[0123] (19);

[0124] in: This is the weight matrix; It is the bias vector; The input feature matrix is ​​denoted as .

[0125] The purpose of this step is to capture the correlation between features (such as the synergistic effect of different vibration features) and provide a basis for subsequent weight allocation.

[0126] Step 2: Weight normalization (Softmax activation).

[0127] To ensure that the weights meet the normalization condition of "summing up to 1" (to facilitate the interpretation of feature contribution ratios), the original attention score is... Applying the Softmax function, we obtain the final weights:

[0128] (20);

[0129] in, Representing the The first oil abrasive image sample The original score of each oil feature, Representing the The first oil abrasive image sample The original score of each oil feature, The total number representing oil characteristics, This represents the total number of oil abrasive particle image samples. Representing the The first oil abrasive image sample Weights of individual oil characteristics Representing the The first oil abrasive image sample Individual oil characteristics, Representing the A one-dimensional health factor for each oil abrasive particle image sample; this formula guarantees... and Weight It can be directly interpreted as the first The feature is related to the first The contribution percentage of each sample's health status.

[0130] One-dimensional health factor calculation compresses multidimensional features into a one-dimensional health factor through "feature-weight weighted multiplication + dimension summation". The formula is:

[0131] (twenty one);

[0132] in, For the first n The physical meaning of this step, which considers the health factors of a sample, is to amplify the information of high-weight features (such as fault-sensitive features) and suppress the interference of low-weight features (such as environmental noise), ultimately obtaining a quantitative indicator that can accurately reflect the health status of the bearing.

[0133] Three key breakpoints (corresponding to four state boundaries) were detected from the health index HI sequence (i.e., a sequence composed of one-dimensional health factors over time) using a time series segmentation algorithm. The segmented intervals were then mapped to discrete states in chronological order. The evolution of bearing health exhibits phased characteristics: healthy stage (HI stable), early wear (HI slowly rising), severe wear (HI rapidly increasing), and failure stage (HI approaching its limit). This characteristic leads to abrupt change points between different stages, which serve as the basis for segmentation.

[0134] Breakpoint detection relies on the Binseg algorithm (binary segmentation based on L2 loss) of the FSW system, which locates breakpoints by "recursive segmentation + loss minimization": the HI sequence is initialized as a complete segment, all possible segmentation positions are traversed for each segment, the sum of the L2 losses of the two segments after segmentation is calculated, and the point with the minimum loss is selected as the optimal breakpoint, until 3 breakpoints (corresponding to 4 segments) are obtained.

[0135] The L2 loss formula is:

[0136] (twenty two);

[0137] in, Represents the dividing point The loss value at that point, The dividing point represents the health index sequence, which is then divided into [...]. , ]as well as[ , Two sections, Represents the starting point of the health index sequence. Represents the end point of the health index sequence. Representing the A one-dimensional health factor for an oil abrasive particle image sample represent[ , The mean of the one-dimensional health factor of all oil abrasive particle image samples within the [database]. represent[ , The mean of one-dimensional health factors for all oil abrasive particle image samples within the [database].

[0138] After obtaining the breakpoints, the `classify_stages` function completes the state mapping: based on the unidirectional irreversible nature of bearing failure, the time interval is mapped to four states "from early to late": [:breakpoints[0]] represents the healthy state (HI is the lowest and stable), [breakpoints[0]:breakpoints[1]] represents early wear (HI rises slowly), [breakpoints[1]:breakpoints[2]] represents severe wear (HI increases rapidly), and [breakpoints[2]:] represents the failure state (HI reaches a high level or remains stable). Based on the identified breakpoints, the training data is labeled with health status for model training.

[0139] S105-2: Loss Function Design.

[0140] The model training process constructed in this invention uses the cross-entropy loss function to learn the feature distribution of the gearbox's break-in, mild, abnormal, and severe wear states;

[0141] (twenty three);

[0142] In the formula, This represents the cross-entropy loss value; The total number of categories representing the classification task; Representing the One category; The actual label encoding representing the k-th category; The representative model predicted the first The probability of each category.

[0143] S105-3: Adam optimizer update rules.

[0144] To enable the network to find the optimal solution for wear state identification more quickly, this invention uses Adam for update optimization, which can be expressed as:

[0145] (twenty four);

[0146] In the formula, For the first The weight of each step; Indicates the learning rate; , These are the first-order momentum (mean) and second-order momentum (square mean) of the gradient, respectively. To represent a very small constant, preventing division by zero; For the first The weight of each step.

[0147] S106: Stage identification.

[0148] After model training, its prediction accuracy needs to be evaluated using validation set data of wind turbine gearboxes. Core evaluation metrics include wear feature classification accuracy and confusion matrix. These metrics ensure that the model can accurately determine the state of different wear stages of the gearbox in practical applications, thereby optimizing model performance and providing a reliable reference for wear state classification.

[0149] In summary, this invention achieves a full-spectrum characterization of wear states by introducing multi-dimensional indicators such as cutting, fatigue, and oxidation, overcoming the limitations of single-concentration indicators. Through deep integration with equipment health status, it provides accurate and reliable decision-making basis for predictive maintenance. The core of this invention lies in realizing intelligent decision-making for wear diagnosis. The model dynamically focuses on key indicators through an attention mechanism, simulating expert decision-making thinking. Combining FSW and LSTM algorithms, it autonomously completes the accurate division of wear stages and captures evolution trends, ultimately forming a decision judgment that surpasses human experience. The modeling parameters used in this invention (such as attention weights, segmentation thresholds, and LSTM network parameters) are adaptively optimized through data-driven processes, making them easier to implement in engineering and deploy in batches compared to traditional methods that rely on manually setting thresholds. This ease of use not only simplifies the model debugging process but also supports adaptive adjustments under different operating conditions, making this invention more efficient and reliable in the health management of complex equipment such as wind turbine gearboxes.

[0150] To further verify the technical effects of the present invention, a specific example is provided below:

[0151] Gearbox accelerated life tests were conducted using a laboratory power transmission test bench, with oil wear monitoring performed via a high-throughput online oil wear image sensor. This sensor employs a CMOS1 imaging system to directly generate oil wear ring images. The experiment lasted 5183 minutes, with online oil monitoring data automatically and synchronously acquired at 10-minute intervals, accumulating 518 data sets. These data comprehensively record the entire wear process of the gear system from the break-in period, normal wear, abnormal wear to severe wear, providing crucial data support for wear state assessment and fault diagnosis based on oil wear images. The acquired oil wear images were preprocessed using the image preprocessing method proposed in this invention, performing grayscale conversion, background reduction, threshold segmentation, and morphological operations sequentially.

[0152] Feature extraction is performed on the processed oil abrasive images. Based on the aforementioned feature extraction method, five major features are analyzed in the preprocessed abrasive images: concentration, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles. A full lifecycle oil feature dataset is constructed, and the evolution trends of these five major oil abrasive features are as follows: Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 As shown.

[0153] The FSW (Time Series Segmentation) algorithm is used to divide the training oil feature data throughout its entire life cycle into four stages: healthy state, early wear, severe wear, and failure state. A schematic diagram of the stage division is shown below. Figure 9As shown, the state labeling process is performed on the oil feature dataset.

[0154] The oil feature dataset covering the entire reference lifecycle was randomly partitioned into training, testing, and validation sets in a ratio of 7:1.5:1.5. The oil feature data was then transformed using the Z-score normalization method.

[0155] (25);

[0156] in, This is the original data. The mean of the data. Let be the standard deviation of the data. After the above transformation, each feature is transformed into a distribution with a mean of 0 and a standard deviation of 1.

[0157] This paper describes the construction of an LSTM model based on Python, forming a complete health status recognition model. The main structure of the constructed LSTM model is as follows: The model employs an LSTM network structure. The input is time-series feature data processed by a sliding window, specifically an oil feature sequence containing 10 time steps. Each time step includes all oil features extracted from the original data, i.e., the input shape is (4, 10, 5). The data transformation process in the model is as follows:

[0158] Temporal feature extraction: The input sequence is first fed into an LSTM layer (containing 2 hidden layers, each with 16 hidden units). The temporal dependencies are modeled by a recurrent neural network, and the hidden state of each time step is output. Finally, the hidden state of the last time step (with a shape of (4, 16)) is taken as the core feature representation of the sequence.

[0159] Classification mapping: The 16-dimensional hidden features output by the LSTM are fed into the fully connected layer. The features are mapped to a dimension that matches the number of wear state categories through linear transformation, and the output probability distribution vector has a shape of (4, 4).

[0160] The final output is the classification result of the wear state. By taking the maximum value index of the output vector, the predicted label of the corresponding wear state (healthy state, early wear, wear failure, etc.) is obtained. The built model is trained using the processed feature dataset.

[0161] To highlight the advantages of the multi-dimensional oil feature fusion method proposed in this patent, the performance of a model using only a single abrasive particle concentration feature was simultaneously compared and tested, such as... Figure 10 and Figure 11As shown, using the multi-feature fusion method for oil, 150 out of 153 samples were successfully identified. Using the single abrasive particle concentration feature method, only 145 out of 153 samples were successfully identified, and the identification of severe wear and failure states was inconsistent. Experimental results demonstrate that the LSTM-based multi-dimensional oil feature health state identification method of this invention significantly improves the identification accuracy compared to the single abrasive particle concentration feature method, validating the effectiveness of the data-level fusion strategy.

[0162] Using oil characteristic data during equipment operation, a health status stage identification test was conducted on the wind turbine gearbox based on a trained health status stage identification model. The test results are as follows: Figure 12 and Figure 13 As shown, the model identifies the current state as "early wear." This indicates that the intelligent assessment model has sensitively captured the initial signal of gearbox wear transitioning from a normal stage to an abnormal one, revealing the early "qualitative change" in wear. This result successfully verifies the superiority of the intelligent model over traditional methods, enabling earlier detection of potential faults and providing crucial and timely scientific evidence for predictive maintenance and proactive intervention.

[0163] Figure 14 A gearbox health status stage identification system based on oil characteristics is shown, including:

[0164] The oil feature extraction unit 1401 is configured to: preprocess the acquired oil abrasive image, and extract oil features based on the preprocessed oil abrasive image. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0165] Health status classification unit 1402 is configured to: obtain gearbox health status classification results based on oil characteristics and a pre-trained long short-term memory neural network model; wherein, the training of the long short-term memory neural network model includes:

[0166] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0167] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0168] It is understood that the aforementioned units can be individually or entirely merged into one or more other units, or some of the units can be further divided into multiple functionally smaller units. This achieves the same operation without affecting the technical effects of the embodiments of the present invention. The aforementioned units are based on logical functional division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of the present invention, the system may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by multiple units.

[0169] According to another embodiment of the present invention, the system of this embodiment can be constructed by running a computer program (including program code) capable of performing the steps involved in the corresponding method of the present invention on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, loaded into the aforementioned computing device through the computer-readable recording medium, and run therein.

[0170] Figure 15 A computer device is shown, which includes a processor 1501, a communication interface 1502, and a computer-readable storage medium 1503. The processor 1501, communication interface 1502, and computer-readable storage medium 1503 can be connected via a bus or other means.

[0171] The communication interface 1502 is used to receive and send data. The computer-readable storage medium 1503 can be stored in the memory of the electronic device. The computer-readable storage medium 1503 is used to store computer programs, which include program instructions. The processor 1501 is used to execute the program instructions stored in the computer-readable storage medium 1503.

[0172] Processor 1501 is the computing and control core of electronic devices. It is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions.

[0173] Processor 1501 is configured to perform the following procedure:

[0174] The acquired oil abrasive images are preprocessed, and oil features are extracted based on the preprocessed oil abrasive images. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0175] Based on the characteristics of the oil and a pre-trained long short-term memory neural network model, the gearbox health status classification results are obtained; the training of the long short-term memory neural network model includes:

[0176] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0177] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0178] This invention also provides a computer-readable storage medium, which is a memory device in an electronic device for storing programs and data. It is understood that the computer-readable storage medium here may include both built-in storage media in the electronic device and extended storage media supported by the electronic device. The computer-readable storage medium provides storage space for storing the processing system of the electronic device.

[0179] Furthermore, this storage space also contains one or more instructions suitable for loading and execution by the processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory; alternatively, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0180] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes the one or more instructions stored in the computer-readable storage medium to perform the following process:

[0181] The acquired oil abrasive images are preprocessed, and oil features are extracted based on the preprocessed oil abrasive images. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0182] Based on the characteristics of the oil and a pre-trained long short-term memory neural network model, the gearbox health status classification results are obtained; the training of the long short-term memory neural network model includes:

[0183] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0184] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0185] The present invention also provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the following process:

[0186] The acquired oil abrasive images are preprocessed, and oil features are extracted based on the preprocessed oil abrasive images. The oil features include wear rate, proportion of large abrasive particles, proportion of cutting abrasive particles, proportion of fatigue abrasive particles, and proportion of oxidized abrasive particles.

[0187] Based on the characteristics of the oil and a pre-trained long short-term memory neural network model, the gearbox health status classification results are obtained; the training of the long short-term memory neural network model includes:

[0188] The oil features of each oil abrasive image sample are fused to obtain a one-dimensional health factor, and a health index sequence is obtained based on the one-dimensional health factor of each oil abrasive image sample.

[0189] Breakpoint detection was performed on the health index sequence to obtain three breakpoints. Based on the obtained breakpoints, health status labels were added to all oil abrasive image samples. The health status labels include healthy state, early wear, severe wear, and failure state.

[0190] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described functions using different methods for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0191] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic cable, digital cable) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0192] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for gear box health condition phase identification based on oil signature, characterized in that, The method comprises the following steps: The acquired oil abrasive particle image is preprocessed, and oil features are extracted according to the preprocessed oil abrasive particle image, wherein the oil features include wear rate, large abrasive particle proportion, cutting abrasive particle proportion, fatigue abrasive particle proportion and oxidation abrasive particle proportion; According to the oil features and the pre-trained long short-term memory neural network model, a gearbox health state classification result is obtained; wherein the training of the long short-term memory neural network model comprises: The oil features of each oil abrasive particle image sample are fused to obtain a one-dimensional health factor, including: ； wherein, , represents a raw score of a first oil feature of a first oil debris image sample, represents a raw score of a first oil feature of a first oil debris image sample, represents a raw score of a first oil feature of a first oil debris image sample, represents a total number of oil features, represents a weight of a first oil feature of a first oil debris image sample, represents a first oil feature of a first oil debris image sample, represents a first oil feature of a first oil debris image sample, represents a one-dimensional health factor of a first oil debris image sample;​​​​​​​ According to the one-dimensional health factors of each oil abrasive particle image sample, a health index sequence is obtained according to time; A binary segmentation based on L2 loss is adopted to detect breakpoints of the health index sequence to obtain three breakpoints, including: The health index sequence is regarded as a complete segment, all possible segmentation positions are pre-segmented, and the sum of L2 losses of two segments after segmentation is calculated, including: wherein, a loss value at a split point, a split point, dividing the health index sequence into two segments, a start point of the health index sequence, an end point of the health index sequence, a one-dimensional health factor of the i-th oil debris image sample, a mean value of the one-dimensional health factors of all oil debris image samples in the interval [i, j], selecting the minimum loss as the optimal split point until three split points are obtained; and adding health status labels to all oil debris image samples according to the obtained split points, the health status labels including a healthy state, an early wear state, a severe wear state and a failure state.​​​​​​​​​​ 2. The gearbox health state phase identification method based on oil features according to claim 1, wherein The acquired oil abrasive particle image is preprocessed, including sequentially performing grayscale processing, background reduction processing, binaryzation processing and morphological processing on the oil abrasive particle image.

3. The gearbox health state phase identification method based on oil features according to claim 1, wherein The loss function of the long short-term memory neural network model is: ； wherein, represents a cross-entropy loss value; represents the total number of classes for a classification task; represents the th class; represents the true label encoding for the th class; represents the probability of the th class predicted by the model.

4. A gearbox health condition phase recognition system based on oil signature, characterized in that, including: The oil feature extraction unit is configured to preprocess the acquired oil abrasive particle image, and extract oil features according to the preprocessed oil abrasive particle image, wherein the oil features include wear rate, large abrasive particle proportion, cutting abrasive particle proportion, fatigue abrasive particle proportion and oxidation abrasive particle proportion; The health state classification unit is configured to obtain a gearbox health state classification result according to the oil features and the pre-trained long short-term memory neural network model; wherein the training of the long short-term memory neural network model comprises: The oil features of each oil abrasive particle image sample are fused to obtain a one-dimensional health factor, including: ； in, , Representing the The first oil abrasive image sample The original score of each oil feature, Representing the The first oil abrasive image sample The original score of each oil feature, The total number representing oil characteristics, Representing the The first oil abrasive image sample Weights of individual oil characteristics Representing the The first oil abrasive image sample Individual oil characteristics, Representing the One-dimensional health factors for each oil abrasive image sample; a health index sequence is obtained over time based on the one-dimensional health factors of each oil abrasive image sample. A binary segmentation based on L2 loss is adopted to detect breakpoints of the health index sequence to obtain three breakpoints, including: The health index sequence is regarded as a complete segment, all possible segmentation positions are pre-segmented, and the sum of L2 losses of two segments after segmentation is calculated, including: in, Represents the dividing point The loss value at that point, The dividing point represents the health index sequence, which is then divided into [...]. , ]as well as[ , Two sections, Represents the starting point of the health index sequence. Represents the end point of the health index sequence. Representing the A one-dimensional health factor for an oil abrasive particle image sample represent[ , The mean of the one-dimensional health factor of all oil abrasive particle image samples within the [database]. represent[ , The mean of the one-dimensional health factor of all oil wear particle image samples is obtained; the point with the minimum loss is selected as the optimal breakpoint, until three breakpoints are obtained; based on the obtained breakpoints, health status labels are added to all oil wear particle image samples, and the health status labels include healthy status, early wear, severe wear and failure status.

5. A computer device, comprising: including: A processor and a computer readable storage medium; The processor is adapted to execute a computer program; The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the gearbox health state phase identification method based on oil features according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is adapted to be loaded and executed by the processor to implement the gearbox health state phase identification method based on oil features according to any one of claims 1 to 3.

7. A computer program product, characterised in that, The computer program product comprises a computer program which, when executed by a processor, implements the oil characteristic-based gear box health state stage identification method according to any one of claims 1 to 3.

Images