A cross-section profile of a drift buoy state monitoring method

Abstract

The present application relates to a kind of profile drift buoy state monitoring method, belong to state monitoring technical field, including the digital twin model of profile drift buoy is constructed, according to the parameter of digital twin model is updated according to preset fault type, obtain fault simulation data;Fault simulation data is segmented according to preset sliding window, extract multidimensional statistical characteristics;Multidimensional statistical characteristics are input into multilayer perception machine and are calculated to obtain feature score, feature score is converted into probability value by discrete distribution sampling layer, according to preset threshold, probability value is screened by binary coding, and obtain screened feature vector;Screened feature vector is input into gradient boosting machine and is monitored, and state monitoring result is obtained, adaptive feature screening is carried out using multilayer perception machine, avoid the limitation of artificial selection feature, realize state monitoring by gradient boosting machine, combine feature selection and classification advantage, so that monitoring result has high precision and interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of state monitoring, in particular to a profile drift buoy state monitoring method. BACKGROUND

[0002] The profile drift buoy is an autonomous observation equipment for marine environment monitoring, which can periodically move up and down in the water column, collect hydrological data such as temperature, salinity and pressure, and drift with the ocean current to obtain large-scale ocean information. Due to its long-term operation in complex marine environment, the mechanical structure, electrical system and sensor of the buoy are easily affected by faults such as biological attachment, hydraulic leakage and motor wear, resulting in a decrease in data quality or equipment failure. Therefore, real-time monitoring of the working state of the buoy, timely discovery and diagnosis of potential faults, are of great significance to ensure the reliability of observation data, prolong the service life of equipment and reduce maintenance costs.

[0003] At present, the state monitoring of the profile drift buoy mainly relies on black box methods based on artificial intelligence, such as deep learning models such as convolutional neural network (CNN), which automatically identify fault patterns by analyzing sensor data, avoiding the low efficiency of traditional manual detection. In addition, some studies use genetic algorithms combined with ensemble learning techniques to first select key features through genetic algorithms, and then use decision tree models for classification to reduce computational complexity.

[0004] However, the existing black box deep learning model (such as CNN) lacks explainability, making it difficult to clearly show the decision basis for fault diagnosis, reducing the transparency of practical application, and usually relying on high-performance computing devices, making it difficult to deploy in resource-constrained embedded systems such as buoys. Although genetic algorithms can optimize feature selection, they have high computational costs, especially when dealing with high-dimensional features, which significantly increases the time consumption, and still rely on a large amount of training data, while the data scarcity problem in the field of marine equipment further restricts the practicality of such methods. SUMMARY

[0005] In view of the deficiencies in the related art, the present application aims to provide a profile drift buoy state monitoring method to solve the technical problem that the existing deep learning-based black box model lacks explainability, relies on high-performance hardware and has high computational cost, and is limited by data scarcity in the field of marine equipment, making it difficult to achieve efficient and practical fault diagnosis in embedded systems such as buoys.

[0006] The present application provides a profile drift buoy state monitoring method, comprising the following steps:

[0007] Data acquisition step: constructing a digital twin model of the profile drift buoy, updating the parameters of the digital twin model according to the preset fault types of the profile drift buoy, and obtaining fault simulation data;

[0008] a feature extraction step of segmenting the fault simulation data according to a preset sliding window and extracting multi-dimensional statistical features;

[0009] a feature screening step of inputting the multi-dimensional statistical features into a multi-layer perception machine to obtain feature scores, converting the feature scores into probability values through a discrete distribution sampling layer, screening the probability values through binary coding according to a preset threshold, and obtaining a screened feature vector;

[0010] a state monitoring step of inputting the screened feature vector into a gradient boosting machine to obtain a state monitoring result.

[0011] The fault simulation data is generated by constructing a digital twin model, which solves the problem of scarcity of real fault data, so that the digital twin model can be trained in the absence of actual fault samples; the high-dimensional time series data is converted into low-dimensional features that can be processed through sliding window segmentation and statistical feature extraction, improving the calculation efficiency; the self-adaptive feature screening is realized by using the multi-layer perception machine, avoiding the limitations of manual feature selection, and enhancing the generalization ability of the model; finally, the state monitoring is realized by the gradient boosting machine, combining the advantages of feature selection and classification, so that the monitoring result has high precision and interpretability, providing a reliable basis for the maintenance of the profile drift buoy.

[0012] In some embodiments of the present application, the data acquisition step specifically includes:

[0013] constructing a mechanical motion model and an electrical model of the profile drift buoy;

[0014] acquiring historical data of the profile drift buoy, updating parameters of the mechanical motion model and parameters of the electrical model according to the historical data and preset experimental parameters, and obtaining the digital twin model according to the parameter-updated mechanical motion model and the parameter-updated electrical model.

[0015] By respectively constructing the mechanical motion model and the electrical model and updating based on the historical data and the experimental parameters, the physical accuracy of the digital twin model is ensured; the mechanical motion model accurately describes the dynamic behavior of the buoy, and the electrical model reflects the interaction between the motor and the load, so that the fault simulation data is closer to the real scene; through the parameter updating mechanism, the digital twin model can dynamically adapt to different environmental conditions, thereby generating diversified fault data to provide high-quality input for subsequent feature extraction and state monitoring.

[0016] In some embodiments of the present application, the feature screening step specifically includes:

[0017] judging the size of the preset threshold value and the probability value, if the probability value is lower than the preset threshold value, the probability value is simplified to 0;

[0018] Otherwise, the probability value is simplified to 1, and the feature vector corresponding to the probability value of 1 is screened to obtain the screened feature vector.

[0019] The probability value output by the multi-layer perception is binarized by the preset threshold value, the hard selection of the feature is realized, the continuous probability is converted into discrete 0-1 coding, the structure of the feature vector is simplified, the calculation complexity is reduced, the retention and abandonment of the feature are clarified, and the feature that has the greatest influence on the classification result can be focused on; through threshold control, the key features can be retained while the redundant information is removed, so that the efficiency and robustness are improved.

[0020] In some embodiments of the application, the method further comprises:

[0021] The alternating training step: according to a preset alternating round, the multi-layer perception and the gradient boosting machine are alternately trained, the parameters of the multi-layer perception are frozen, and the gradient boosting machine is trained for a first iteration number of times through the screened feature vector;

[0022] After the training of the gradient boosting machine is completed, the parameters of the gradient boosting machine are frozen, the multi-layer perception is trained for a second iteration number of times, and the alternating training of the multi-layer perception and the gradient boosting machine is repeated until the preset round is reached.

[0023] Through the alternating training of the multi-layer perception and the gradient boosting machine, the collaborative optimization of feature selection and the classifier is realized; the way of freezing the parameters of one side to train the other side avoids the parameter conflict caused by simultaneous updating, so that the two can gradually converge to the optimal solution; the alternating training mechanism ensures that the direction of feature selection is always aimed at improving the classification performance, and the feedback of the multi-layer perception and the gradient boosting machine also guides the improvement of feature selection, finally forming a kind of dynamic balance, which significantly improves the overall performance.

[0024] In some embodiments of the application, the alternating training step specifically further comprises:

[0025] The cross-entropy loss of the gradient boosting machine is optimized by the gradient boosting method, and the parameters of the gradient boosting machine are updated;

[0026] The loss function of the multi-layer perception is optimized by the gradient descent method, and the parameters of the multi-layer perception are updated according to the gradient boosting machine, the multi-dimensional statistical feature and the screened feature vector.

[0027] By optimizing the cross-entropy loss of the gradient booster machine using the gradient boosting method, the gradient booster machine can minimize the prediction error and improve the accuracy of fault classification. At the same time, the gradient descent method optimizes the loss function of the multilayer perceptron, ensuring that the feature selection process not only depends on the classification performance, but also combines the causality between features and labels, achieving optimal performance in both feature selection and classification tasks, thus realizing efficient and reliable state monitoring.

[0028] In some embodiments of the present invention, the calculation model for the cross-entropy loss of the gradient booster is as follows:

[0029]

[0030] in, For cross-entropy loss operator; For input Learnable parameters The prediction results of the gradient booster; This is a real label; The number of samples; Number of fault types; The indicator function is used to select the first... Sample Does it belong to the first Check for similar faults; For gradient boosting machines with learnable parameters Below the input feature vector Belongs to the The original prediction score for the class; For gradient boosting machines with learnable parameters Below the input feature vector Belongs to the The original predicted score for the class.

[0031] The classification error of the gradient booster is quantified by the cross-entropy loss function, which enables the gradient booster to clearly distinguish the probability distribution of different fault types. The calculation of cross-entropy loss is based on the logarithmic difference between the true label and the predicted probability, which forces the gradient booster to predict the probability of the true class close to 1, thereby improving the classification accuracy. It is suitable for multi-class problems, can effectively handle class imbalance and data noise, and ensures the stability of the gradient booster in practical applications.

[0032] In some embodiments of the present invention, the calculation model of the loss function of the multilayer perceptron is as follows:

[0033]

[0034] in, Let be the loss function of the multilayer perceptron; It is the first balance factor; It is the second balance factor; This is a causal loss operator.

[0035] By combining classification loss and causal loss to construct the total loss function of the multilayer perceptron, the feature selection process considers both classification performance and emphasizes the causal correlation of features. The classification loss ensures that the selected features can improve classification accuracy, while the causal loss constrains the physical correlation between features and fault labels, avoiding the selection of false features by the multilayer perceptron, enhancing the interpretability of features, and providing a more reliable basis for fault diagnosis.

[0036] In some embodiments of the present invention, the computational model of the causal loss operator is as follows:

[0037]

[0038] in, Multidimensional statistical characteristics; For Gumbel's set of random variables; For the Gumbel random variable set, the first... One element; These are the parameters of the multilayer perceptron; This is an element-wise multiplication operation; The feature vector of a single sample; The probability values ​​generated by the multilayer perceptron; For gradient booster to the first Predicted probability of class-specific faults.

[0039] The causal loss operator quantifies the difference in the impact of retained and discarded features on the classification results, forcing the multilayer perceptron to select features that are crucial to the classification. The calculation of causal loss is based on the comparison of the performance of the multilayer perceptron under retained and discarded features, ensuring that the selected features have a strong causal relationship with the fault labels. This effectively solves the problem that feature selection in traditional methods depends on correlation rather than causation, and improves the generalization ability and reliability of the multilayer perceptron.

[0040] In some embodiments of the present invention, the dynamics of the mechanical motion model are represented as follows:

[0041]

[0042] in, The mass of the profile drifting buoy; For depth The density of the water below; The volume of the glass sphere shell; is the compressibility coefficient of the glass sphere shell; This refers to the volume of the external oil sac. The compressibility coefficient of the hydraulic oil; is an unmodelable item; is a compression coefficient corresponding to the unmodelable item; is a drag coefficient of the profile drift buoy; is a moving speed of the profile drift buoy.

[0043] The mechanical motion behavior of the profile drift buoy is accurately described by the dynamic equation, the motion state of the profile drift buoy under different water depths and environmental conditions can be simulated, a high-fidelity mechanical behavior basis is provided for the digital twin model, and the motion trend of the buoy can be accurately predicted through the analysis of the dynamic equation, thereby providing theoretical support for fault simulation and state monitoring.

[0044] In some embodiments of the application, the representation of the electrical model is:

[0045]

[0046] wherein, is a torque constant of the plunger pump motor; is a current of the plunger pump motor; is a friction coefficient of the plunger pump motor; is a rotational speed of the plunger pump motor; is a first load coefficient of the plunger pump motor; is a second load coefficient of the plunger pump motor.

[0047] The power balance relationship of the plunger pump motor is described by the electrical model, including electromagnetic torque, friction resistance and load effect, which can reflect the current, rotational speed and load characteristics of the plunger pump motor under different working conditions, and provide accurate basis for electrical fault simulation, and through parameter calibration of the electrical model, fault simulation data highly consistent with the actual motor behavior can be generated, thereby supporting training and verification of fault diagnosis algorithms. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, specific embodiments of the present application will be described in detail below with reference to the drawings, and other drawings can also be obtained by those of ordinary skill in the art without creative labor on the premise that the drawings are not attached.

[0049] Figure 1 is a flowchart of a profile drift buoy state monitoring method provided by an embodiment of the application;

[0050] Figure 2 is a flowchart of another profile drift buoy state monitoring method provided by an embodiment of the application. DETAILED DESCRIPTION

[0051] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is described and explained below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0052] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be understood that the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0053] In the field of profile drift buoy state monitoring, existing technical methods mostly rely on artificial intelligence-based black box methods such as convolutional neural networks (CNN) and the like.

[0054] However, the above method has the problem of poor interpretability. Most black box-based deep learning models are usually difficult to explain their decision-making process. This lack of interpretability makes it difficult to understand how the model arrives at its conclusions in fault diagnosis, resulting in low transparency in practical applications.

[0055] In addition, the above deep learning model often needs high-performance computing devices to process large amounts of data and perform complex calculations, limiting its application in resource-limited profile drift buoys.

[0056] Deep learning models are usually computationally complex, expensive, and difficult to run on lightweight devices. Complex models may require specialized knowledge to adjust and maintain.

[0057] According to the problems of the above-mentioned black box-based deep learning model, existing research combines genetic algorithm and ensemble learning for fault diagnosis of profile drift buoys. Genetic algorithm is used for optimization to select the most meaningful, important and useful features from the large statistical feature space extracted from the original sensor data. Subsequently, ensemble learning techniques based on decision trees are used to identify the fault state and fault type of the profile drift buoy.

[0058] Although the above method has simple input, small computational burden and good performance in actual training of decision tree, it cannot be ignored that a large amount of computational resources will be consumed when the genetic algorithm is used for feature subset selection, which is easy to cause the problem of too long calculation time. In addition, the method based on artificial intelligence often depends on a large amount of data, and in the field of marine equipment with scarce data, the application of such method will be seriously restricted.

[0059] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0060] The technical solutions of the application will be described in detail below with specific embodiments and the accompanying drawings.

[0061] As shown in Figure 1 The application provides a cross-section drift buoy state monitoring method, which comprises the following steps:

[0062] The data acquisition step S1: a digital twin model of the cross-section drift buoy is constructed, the parameters of the digital twin model are updated according to the preset fault types of the cross-section drift buoy, and fault simulation data is obtained;

[0063] Optionally, the preset fault types are faults occurring in the development process of the cross-section drift buoy, such as hydraulic oil leakage, biological attachment, plunger pump wear, plunger pump inlet pressure fluctuation, plunger pump motor failure, structural damage and the like.

[0064] In some embodiments, the data acquisition step S1 specifically comprises:

[0065] A mechanical motion model and an electrical model of the cross-section drift buoy are constructed;

[0066] The historical data of the cross-section drift buoy are acquired, the parameters of the mechanical motion model and the electrical model are updated according to the historical data and the experimental preset parameters, and the digital twin model is obtained according to the parameter-updated mechanical motion model and the parameter-updated electrical model; optionally, the method for updating the parameters of the mechanical motion model and the electrical model is the recursive least squares method.

[0067] By respectively constructing the mechanical motion model and the electrical model and updating based on the historical data and the experimental parameters, the physical accuracy of the digital twin model is ensured; the mechanical motion model accurately describes the dynamic behavior of the buoy, and the electrical model reflects the interaction between the motor and the load, and the combination of the two makes the fault simulation data closer to the real scene; through the parameter updating mechanism, the digital twin model can dynamically adapt to different environmental conditions, thereby generating diversified fault data, providing high-quality input for subsequent feature extraction and state monitoring.

[0068] In some embodiments, the dynamics of the mechanical motion model is represented as:

[0069]

[0070] wherein, is the mass of the profiled drift buoy; is the density of water at depth is the volume of the glass sphere shell; is the compressibility of the glass sphere shell; is the volume of the outer oil bladder; is the compressibility of the hydraulic oil; is the unmodelable term; is the compressibility of the unmodelable term; is the drag coefficient of the profiled drift buoy; is the velocity of the profiled drift buoy; In particular,

[0071] is the other volume except the volume of the glass sphere shell and the volume of the outer oil bladder, the other volume is the unmeasurable volume in the process of balancing the mass and buoyancy of the profiled drift buoy in the laboratory, i.e. the unmodelable term. The mechanical motion behavior of the profiled drift buoy is accurately described by the dynamic equation, which can simulate the motion state of the profiled drift buoy under different water depths and environmental conditions, and provides a high-fidelity mechanical behavior basis for the digital twin model; through the analysis of the dynamic equation, the motion trend of the buoy can be accurately predicted, thereby providing theoretical support for fault simulation and state monitoring.

[0072] In some embodiments, the electrical model mainly reflects that the power of the plunger pump motor changes at different depths, and the representation of the electrical model is:

[0073]

[0074] wherein,

[0075] is the torque constant of the plunger pump motor; is the current of the plunger pump motor; is the friction coefficient of the plunger pump motor; is the rotational speed of the plunger pump motor; is the first load coefficient of the plunger pump motor; is the second load coefficient of the plunger pump motor.

[0076] ​The power balance relationship of the plunger pump motor is described by an electrical model, including electromagnetic torque, friction resistance and load effect, which can reflect the current, speed and load characteristics of the plunger pump motor under different working conditions, and provide accurate basis for electrical fault simulation. Through parameter calibration of the electrical model, fault simulation data highly consistent with the actual motor behavior can be generated, thereby supporting training and verification of fault diagnosis algorithms.

[0077] Optionally, in combination with Figure 2 As shown, the process of updating the parameters of the digital twin model and obtaining the fault simulation data can specifically be embodied as follows: if the plunger pump motor inlet pressure fluctuates, the plunger pump motor will have significant current fluctuations at the same speed, so a noise signal can be injected at the current to simulate fluctuations.

[0078] Feature extraction step S2: the fault simulation data is segmented according to a preset sliding window, multi-dimensional statistical features are extracted, the multi-dimensional statistical features are normalized and equally spaced, different degrees of symbols are formed, and the transferability is improved.

[0079] Optionally, the multi-dimensional statistical features include minimum value, maximum value, average value, median value, peak-to-peak value, standard deviation, zero-crossing point number, kurtosis, skewness, root mean square (RMS) value, RMS speed, entropy and wave crest factor;

[0080] Feature selection step S3: the multi-dimensional statistical features are input into a multilayer perceptron to calculate feature scores, the feature scores are converted into probability values through a discrete distribution sampling layer, and the probability values are selected through binary coding according to a preset threshold to obtain a selected feature vector;

[0081] Specifically, the multilayer perceptron (Multilayer Perceptron) is a kind of feedforward artificial neural network, which adopts a full connection layer structure and is composed of multiple full connection layers and batch normalization (BN) layers;

[0082] The discrete distribution sampling layer is a Gumbel-Softmax layer, which adopts a differentiable discrete distribution sampling technology and is used to select or sample categories of discrete features in a neural network while keeping the gradient propagable.

[0083] The multilayer perceptron calculates the influence degree of each feature on the result according to the multi-dimensional statistical features, and outputs a feature score for each feature;

[0084] The output feature score is first converted into a probability by the discrete distribution sampling layer, and the probabilities of the multi-dimensional statistical features sum to 1.

[0085] In some embodiments, the feature selection step S3 specifically includes:

[0086] determining the size of the preset threshold value and the probability value, if the probability value is lower than the preset threshold value, simplifying the probability value to 0;

[0087] Otherwise, the probability value is simplified to 1, and the feature vector corresponding to the probability value of 1 is screened to obtain a screened feature vector, which can sample the most influential feature from the discrete multi-dimensional statistical feature and suppress the remaining feature vectors to achieve the purpose of adaptive selection of features.

[0088] By presetting the threshold value to binarize the probability value output by the multi-layer perception, the hard selection of features is realized, the continuous probability is converted into discrete 0-1 coding, the structure of the feature vector is simplified, the calculation complexity is reduced, and the retention and rejection of the feature are also clear, which can focus on the feature that has the greatest impact on the classification result; through threshold control, redundant information can be removed while retaining key features, thereby improving efficiency and robustness.

[0089] State monitoring step S4: input the screened feature vector into the gradient boosting machine for state monitoring to obtain a state monitoring result.

[0090] Optionally, the gradient boosting machine can use a light gradient boosting machine (LightGBM), which is a high-efficiency gradient boosting decision tree (GBDT) implementation that solves the efficiency and accuracy problems of traditional GBDT in processing large data, and can achieve faster training speed, lower memory consumption and higher accuracy through various optimization techniques.

[0091] The gradient boosting machine can learn the interpretable faults from the screened feature vector, thereby achieving effective monitoring of various states.

[0092] By constructing a digital twin model to generate fault simulation data, the problem of scarcity of real fault data is solved, so that the digital twin model can be trained in the absence of actual fault samples; through sliding window segmentation and statistical feature extraction, high-dimensional time series data is converted into processable low-dimensional features, improving the calculation efficiency; adaptive feature screening is performed using a multi-layer perception to avoid the limitations of manual feature selection, enhancing the generalization ability of the model; finally, state monitoring is realized through a gradient boosting machine, combining the advantages of feature selection and classification, so that the monitoring result has high precision and interpretability, providing a reliable basis for profile drift buoy maintenance.

[0093] In some embodiments, the state monitoring method further comprises:

[0094] Alternating training step: according to a preset alternating round, the multi-layer perception and the gradient boosting machine are alternately trained, the parameters of the multi-layer perception are frozen, and the gradient boosting machine is trained for a first number of iterations through the screened feature vector.

[0095] After the gradient boosting machine is trained, the parameters of the gradient boosting machine are frozen, and the multilayer perceptron is trained for a second number of iterations, and the alternating training of the multilayer perceptron and the gradient boosting machine is repeated until a preset number of rounds is reached.

[0096] Through the alternating training of the multilayer perceptron and the gradient boosting machine, the collaborative optimization of feature selection and the classifier is realized; the way of freezing the parameters of one party to train the other party avoids the parameter conflict caused by simultaneous updating, so that the two can gradually converge to the optimal solution; the alternating training mechanism ensures that the direction of feature selection is always aimed at improving the classification performance, and the feedback of the multilayer perceptron and the gradient boosting machine also guides the improvement of feature selection, eventually forming a dynamic balance, which significantly improves the overall performance.

[0097] In some embodiments, the alternating training step specifically further comprises:

[0098] The cross-entropy loss of the gradient boosting machine is optimized by the gradient boosting method, and the parameters of the gradient boosting machine are updated based on the cross-entropy loss between the predicted label and the real label;

[0099] The loss function of the multilayer perceptron is optimized by the gradient descent method, and the parameters of the multilayer perceptron are updated according to the gradient boosting machine, the multidimensional statistical features and the screened feature vectors;

[0100] Specifically, the parameter update of the multilayer perceptron is jointly determined by the performance of the gradient boosting machine and the interaction information between the screened feature vectors and the remaining feature vectors

[0101] The cross-entropy loss of the gradient boosting machine is optimized by the gradient boosting method, so that the gradient boosting machine can minimize the prediction error and improve the accuracy of fault classification; at the same time, the loss function of the multilayer perceptron is optimized by the gradient descent method, which ensures that the feature selection process not only depends on the classification performance, but also combines the causality of features and labels, and can achieve optimization in both feature selection and classification tasks, thereby realizing efficient and reliable state monitoring.

[0102] In some embodiments, the calculation model of the cross-entropy loss of the gradient boosting machine is:

[0103]

[0104] wherein, is a cross-entropy loss operator; is the prediction result of the gradient boosting machine with input , learnable parameters ; is a real label; is the number of samples; is the number of fault types; ​to indicate whether the first sample belongs to the first fault type; to indicate the original prediction score of the input feature vector belonging to the first fault type by the gradient boosting machine under the learnable parameters; to indicate the original prediction score of the input feature vector belonging to the first fault type by the gradient boosting machine under the learnable parameters. The classification error of the gradient boosting machine is quantified by the cross-entropy loss function, so that the gradient boosting machine can clearly distinguish the probability distribution of different fault types; the calculation of the cross-entropy loss is based on the logarithmic difference between the true label and the prediction probability, which forces the prediction probability of the true class to approach 1, thereby improving the classification accuracy, suitable for multi-classification problems, and can effectively handle class imbalance and data noise, ensuring the stability of the gradient boosting machine in practical applications.

[0105] In some embodiments, the calculation model of the loss function of the multi-layer perception is:

[0106]

[0107]

[0108] wherein, is the loss function of the multi-layer perception; is the first balance factor; is the second balance factor; is the causal loss operator, indicating the causality between the filtered feature vector and the remaining feature vector.

[0109] By combining the classification loss and the causal loss to construct the total loss function of the multi-layer perception, the feature selection process not only considers the classification performance, but also emphasizes the causal correlation of the features; the classification loss ensures that the selected features can improve the classification accuracy, and the causal loss constrains the physical association between the features and the fault label, avoiding the selection of false features by the multi-layer perception and enhancing the explainability of the features, thereby providing a more reliable basis for fault diagnosis.

[0110] In some embodiments, the calculation model of the causal loss operator is:

[0111]

[0112] wherein, is the multi-dimensional statistical feature; is a set of Gumbel random variables; is the first​​ an element; parameters of a multilayer perceptron; an element multiplication operation; a feature vector of a single sample; a probability value generated by the multilayer perceptron; a prediction probability of the multilayer perceptron for the first a prediction probability of the multilayer perceptron for the first

[0113] The causal loss operator quantifies the difference in the influence of the retained features and the discarded features on the classification result, forcing the multilayer perceptron to select features that are crucial to classification; the calculation of the causal loss is based on the performance comparison of the multilayer perceptron under retained and discarded features, ensuring that the selected features have a strong causal relationship with the fault label, effectively solving the problem that feature selection in traditional methods relies on correlation rather than causality, and improving the generalization ability and reliability of the multilayer perceptron.

[0114] It should be noted that the above is a reference way for a profile drift buoy state monitoring method, and the present application is not limited thereto.

[0115] The digital twin model of the embodiment can be trained in the absence of actual fault samples; high-dimensional time series data is converted into processable low-dimensional features through sliding window segmentation and statistical feature extraction, improving the calculation efficiency; adaptive feature screening is performed using a multilayer perceptron, avoiding the limitations of manual feature selection and enhancing the generalization ability of the model; and finally, state monitoring is realized through a gradient boosting machine, combining the advantages of feature selection and classification, so that the monitoring result has high precision and interpretability, providing a reliable basis for profile drift buoy maintenance, solving the technical problems that the existing technology based on deep learning lacks interpretability, it is difficult to clearly show the decision basis of fault diagnosis, reduces the transparency of practical application, and usually relies on high-performance computing devices, which is difficult to deploy in resource-constrained embedded systems such as buoys, and although genetic algorithms can optimize feature selection, the calculation cost is high, especially when processing high-dimensional features, the time consumption is significant, and still relies on a large amount of training data, and the problem of data scarcity in the field of marine equipment further restricts the practicality of such methods.

[0116] Finally, it should be noted that: the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other.

[0117] The above examples are only used to illustrate the technical solutions of the present application but not to limit the present application; although the present application has been described in detail with reference to the preferred embodiments, it is understood by the person of ordinary skill in the art that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the technical solutions of the present application, and all of them should be covered in the technical solution range of the present application claimed.

Claims

1. A method for monitoring the status of a profile drifting buoy, characterized in that, Includes the following steps: Data acquisition steps: Construct a digital twin model of the profile drifting buoy, update the parameters of the digital twin model according to the preset fault type of the profile drifting buoy, and obtain fault simulation data; Feature extraction step: The fault simulation data is segmented according to a preset sliding window, and multidimensional statistical features are extracted; Feature filtering steps: The multidimensional statistical features are input into a multilayer perceptron for calculation to obtain feature scores. The feature scores are converted into probability values ​​through a discrete distribution sampling layer. The probability values ​​are filtered by binary encoding according to a preset threshold to obtain the filtered feature vector. State monitoring step: Input the filtered feature vector into the gradient booster for state monitoring and obtain the state monitoring results; The data acquisition steps specifically include: Construct the mechanical motion model and electrical model of the drifting buoy in the profile; Historical data of the profile drifting buoy is obtained, and the parameters of the mechanical motion model and the electrical model are updated according to the historical data and experimental preset parameters. The digital twin model is obtained based on the updated mechanical motion model and the updated electrical model. The dynamics of the mechanical motion model are expressed as follows: in, The mass of the profile drifting buoy; For depth The density of the water below; The volume of the glass sphere shell; is the compressibility coefficient of the glass sphere shell; This refers to the volume of the external oil sac. The compressibility coefficient of the hydraulic oil; This is a non-modelable term; The compression coefficient corresponding to the unmodelable term; The drag coefficient of the profile drifting buoy; The velocity of the drifting buoy in the profile; The electrical model is represented as follows: in, This represents the torque constant of the piston pump motor. This refers to the current of the plunger pump motor. The friction coefficient of the piston pump motor; This refers to the rotational speed of the plunger pump motor. This is the first load factor for the piston pump motor; This is the second load factor for the piston pump motor.

2. The method for monitoring the status of a drifting buoy in profile according to claim 1, characterized in that, The feature selection step specifically includes: The magnitude of the preset threshold and the probability value is determined. If the probability value is lower than the preset threshold, the probability value is simplified to 0. Otherwise, the probability value is simplified to 1, and the feature vectors corresponding to the probability value of 1 are filtered to obtain the filtered feature vectors.

3. The method for monitoring the status of a drifting buoy in profile according to claim 1, characterized in that, The method further includes: Alternating training steps: The multilayer perceptron and the gradient booster are trained alternately according to a preset alternation round, the parameters of the multilayer perceptron are frozen, and the gradient booster is trained for the first iteration using the selected feature vectors; After the gradient booster is trained, its parameters are frozen, and the multilayer perceptron is trained for a second number of iterations. The alternating training of the multilayer perceptron and the gradient booster is repeated until the preset number of rounds is reached.

4. The method for monitoring the status of a drifting buoy in profile according to claim 3, characterized in that, The alternating training step further includes: The cross-entropy loss of the gradient booster is optimized using the gradient boosting method, and the parameters of the gradient booster are updated. The loss function of the multilayer perceptron is optimized by gradient descent, and the parameters of the multilayer perceptron are updated according to the gradient boosting machine, the multidimensional statistical features, and the filtered feature vectors.

5. The method for monitoring the status of a drifting buoy in profile according to claim 4, characterized in that, The calculation model for the cross-entropy loss of the gradient booster is as follows: in, For cross-entropy loss operator; For input Learnable parameters The prediction results of the gradient booster; This is a real label; The number of samples; Number of fault types; The indicator function is used to select the first... Sample Does it belong to the first Check for similar faults; For gradient boosting machines with learnable parameters The following is an example of input feature vectors Belongs to the The original prediction score for the class; For gradient boosting machines with learnable parameters The following is an example of input feature vectors Belongs to the The original prediction score for the class.

6. The method for monitoring the status of a drifting buoy in profile according to claim 5, characterized in that, The calculation model for the loss function of the multilayer perceptron is as follows: in, Let be the loss function of the multilayer perceptron; It is the first balance factor; It is the second balance factor; This is a causal loss operator.

7. The method for monitoring the status of a drifting buoy in profile according to claim 6, characterized in that, The computational model for the causal loss operator is as follows: in, Multidimensional statistical characteristics; For Gumbel's set of random variables; For the Gumbel random variable set, the first... One element; These are the parameters of the multilayer perceptron; This is an element-wise multiplication operation; The feature vector of a single sample; The probability values ​​generated by the multilayer perceptron; For gradient booster to the first Predicted probability of class-specific faults.

Images