Fan assembly process missing data filling method based on semi-supervised learning

Abstract

The invention relates to a fan assembly process missing data filling method based on semi-supervised learning, and belongs to the technical field of industrial data missing repair. Aiming at the technical problems that the traditional linear interpolation and supervised learning are difficult to process the missing of wind power assembly multi-source heterogeneous data, the dynamic working condition adaptability is poor and the annotation dependence is high, a feature extraction method of combining space-time tensor modeling with a bidirectional space-time attention network is provided, and the reliability of three-dimensional space-time tensor quantitative data is constructed; designing an adversarial domain adaptation mechanism to realize feature space alignment of labeled and unlabeled data; a dynamic confidence threshold strategy is adopted to screen high-credibility samples, and a collaborative loss function optimization model is constructed; the method has the technical effects that the data filling precision and the process conformity under the complex working condition are remarkably improved, the data labeling requirement is reduced, flexible production is adapted, the cross-production-line data island problem is solved, the real-time repair efficiency is optimized, assembly quality tracing and process optimization are enabled, and reliable data support is provided for intelligent wind power manufacturing.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial data missing repair, and relates to a method for filling missing data in a fan assembly process based on semi-supervised learning. Background Art

[0002] During the assembly of large wind turbines, the collection and transmission of multi-source heterogeneous data (such as vibration signals, torque parameters, and assembly accuracy) often results in data loss, fragmentation, or noise pollution due to factors such as sensor failure, communication interference, equipment overload protection, and operator error. These data quality issues severely restrict the reliability of key links such as multi-source data fusion, real-time status monitoring, and process parameter optimization, directly affecting the controllability of the assembly process, leading to problems such as reduced assembly quality and increased risk of equipment failure. In the flexible production of wind turbines, in particular, differences in different models, production line configurations, and operating specifications lead to diverse data distribution characteristics, further exacerbating the complexity of repairing missing data.

[0003] Traditional data filling methods (such as linear interpolation, threshold methods, or fuzzy clustering) rely on artificial rules or static models, making it difficult to effectively capture the nonlinear spatiotemporal correlation characteristics of assembly parameters in multi-process collaborative operations. For example, complex features such as the dynamic coupling relationship between bolt tightening torque and cabin alignment accuracy, and the cross-station correlation between blade dynamic balancing and whole-machine vibration signals cannot be accurately modeled through linear assumptions or local interpolation. In addition, although existing supervised learning methods perform well in some scenarios, the contradiction between their reliance on large amounts of labeled data and the high cost of labeling in industrial scenarios and the difficulty of replicating real-time working conditions with historically missing data is becoming increasingly prominent.

[0004] Although current research attempts to introduce semi-supervised learning to alleviate the scarcity of labeled data, most methods are still limited to theoretical verification and lack targeted designs for the unique challenges of industrial data:

[0005] Assembly process fluctuations, component tolerances, and environmental disturbances cause data distribution to drift over time, making it difficult for traditional static models to achieve stable coverage across working conditions.

[0006] The data distribution differences between different wind turbine models or production lines form "data silos," and existing methods cannot effectively achieve feature alignment and knowledge transfer across production lines.

[0007] In non-ideal data scenarios where noise interference and partial missingness coexist, traditional confidence strategies are prone to introduce low-quality pseudo labels, leading to error propagation.

[0008] These deficiencies significantly limit the accuracy and robustness of existing methods in real-world industrial scenarios. For example, during impeller dynamic balancing and debugging, the parameter correction error of traditional linear interpolation can reach over 15%, resulting in distortion in assembly quality traceability. Furthermore, in multi-variety, small-batch production, the misjudgment rate of supervised learning due to insufficient labeled data is 20% to 30% higher than that of semi-supervised approaches. Therefore, there is an urgent need for an intelligent filling method that can integrate spatiotemporal correlation features, adapt to dynamic operating conditions, and be compatible with non-ideal industrial data to support assembly quality optimization and equipment health management in wind power intelligent manufacturing.

[0009] In response to the above technical bottlenecks, the present invention proposes a method for filling missing data in the wind turbine assembly process based on semi-supervised learning. Through the innovative design of spatiotemporal attention network, dynamic domain adaptation mechanism and double verification confidence strategy, high-precision and cross-working condition data repair can be achieved under the constraints of a small amount of labeled data, providing reliable technical support for the digital upgrade of the wind power assembly process. Summary of the Invention

[0010] In view of this, the purpose of the present invention is to provide a method for filling missing data in the wind turbine assembly process based on semi-supervised learning. During the assembly and operation of wind turbines, the collection, transmission and storage of multi-source heterogeneous data (such as vibration signals, torque parameters, assembly accuracy, etc.) often lead to data missing, outlier mixing and other problems due to sensor abnormalities, communication interference or equipment protection mechanisms (such as overload shutdown). Due to the influence of dynamic working conditions such as assembly process fluctuations, component fitting tolerances and environmental disturbances, the assembly process data presents significant multimodal correlation and process parameter coupling characteristics. Traditional fuzzy clustering filling or linear interpolation filling methods are difficult to effectively capture its cross-process correlation characteristics, resulting in low accuracy and poor robustness of the repaired data, which directly affects the reliability of assembly quality traceability, process optimization and equipment health management.

[0011] Furthermore, wind turbine assembly data distribution is diverse due to differences in model, assembly line configuration, and worker operating procedures. Manual labeling is costly and difficult to cover the entire process flow. Existing supervised learning methods rely heavily on labeled data, limiting their performance when labeled data is scarce. There is an urgent need for an intelligent infill method that can adaptively integrate multi-source process associations and accommodate non-ideal assembly data.

[0012] Therefore, this paper proposes a method for filling missing data in wind turbine assembly processes based on semi-supervised learning. By combining a small amount of labeled data with a large amount of unlabeled data, and utilizing a process knowledge-guided domain adaptation mechanism and a multimodal feature extraction network, this method can achieve high-precision restoration of missing assembly parameters in complex assembly conditions and scenarios with limited annotation. Furthermore, this method can provide complete and reliable assembly process data for wind power intelligent manufacturing systems, supporting key applications such as process parameter optimization and assembly defect diagnosis, and contributing to the intelligent assembly upgrade of the wind power industry.

[0013] In order to achieve the above object, the present invention provides the following technical solutions:

[0014] The missing data filling method for the fan assembly process based on semi-supervised learning includes the following steps:

[0015] S1: Perform spatiotemporal tensor modeling on the multi-source heterogeneous data in the wind turbine assembly process and construct a three-dimensional tensor X∈R T ×m×s , where T represents the time step, m represents the sensor dimension, and s represents the number of assembly stations; and the confidence weight matrix W is introduced t,k ∈[0,1] quantifies data reliability, where W t,k =0 means data is missing, W t,k =1 means the data is complete, 0<W t,k <1 indicates that the data have some reliability issues;

[0016] S2: Feature extraction of the three-dimensional tensor is performed through the bidirectional spatiotemporal attention network BiSTA-Net, which adopts a fusion structure of multi-scale convolution, bidirectional attention mechanism and residual learning, and outputs the labeled dataset H L and the unlabeled dataset H U Low-dimensional feature representation of ;

[0017] S3: Based on the adversarial domain adaptation mechanism, the labeled dataset H L and the unlabeled dataset H U Perform feature space alignment and implement adversarial training between the encoder and domain discriminator through the gradient reversal layer GRL to minimize the inter-domain distribution difference between labeled data and unlabeled data;

[0018] S4: Use a dynamic confidence threshold strategy to screen high-reliability unlabeled data and calculate the probability distribution of sample categories through a double verification mechanism. and feature space alignment ρ j , combining dynamic thresholds ξ(t) and η(t) to generate pseudo-label data;

[0019] S5: Construct a supervised loss L sup , consistency regularization loss L con and distribution alignment loss L align The collaborative loss function L total =L sup +β(t)L con +γL align ,fill model parameters through end-to-end training optimization;

[0020] S6: Use the trained model to dynamically fill in the missing data in layers, and combine the missing pattern detection of the hidden Markov model (HMM) and the data generator of the generative adversarial network (GAN) to repair the missing values.

[0021] Furthermore, in S1, the spatiotemporal tensor modeling further includes:

[0022] Using the variational Bayesian hierarchical latent variable model, define the global latent variable z t ~N(μ z ,∑ z ) and local latent variables h t,k ~g φ (z t ,x t-1,k ), the observation model is where f θ is the neural network mapping function, is the weighted variance term.

[0023] Furthermore, in S2, the specific structure of the bidirectional spatiotemporal attention network BiSTA-Net includes:

[0024] (a) Multi-scale spatiotemporal feature embedding layer, using a one-dimensional convolution kernel, i.e. kernel_size = 5, to map the input to d model = 64-dimensional feature space, and superimposed hyperbolic sine position encoding:

[0025]

[0026] (b) Bidirectional attention module, forward attention weight and backward attention weights splicing and fusion;

[0027] (c) Residual aggregation layer, through Normalize the skip connection features.

[0028] Furthermore, in S3, the adversarial domain adaptation mechanism is specifically implemented as follows:

[0029] (a) Loss function of domain discriminator C:

[0030]

[0031] (b) Optimization goal of encoder E:

[0032] L E =L task -λL adv

[0033] Where λ is the dynamic adjustment coefficient;

[0034] (c) Domain Alignment Loss:

[0035]

[0036] in and is the mean of the i-th layer features of the source domain and the target domain, ∑ s and ∑ t is the covariance matrix.

[0037] Furthermore, in S4, the dynamic confidence threshold strategy satisfies:

[0038] Sample selection conditions in:

[0039] and is the feature space alignment indicator.

[0040] Furthermore, in S5, the collaborative loss function includes:

[0041] (a) Supervision losses:

[0042]

[0043] (b) Consistency regularization loss:

[0044]

[0045] (c) Distribution Alignment Loss:

[0046] L align =MMD 2 (H L ,H U )+λCORAL(H L ,H U )

[0047] Among them, MMD is the maximum mean difference and CORAL is the correlation alignment loss.

[0048] Furthermore, in S6, the hierarchical dynamic filling specifically includes:

[0049] (a) Missing pattern detection uses Hidden Markov Model HMM, and the observation probability is:

[0050]

[0051] (b) Data generator G(z|{h known},E pos θ G ) Fusion of known features h known and temporal position encoding:

[0052]

[0053] Furthermore, in the bidirectional attention module, the attention coefficient is calculated using:

[0054]

[0055] in, is the learnable parameter vector, is the weight matrix, and || represents the vector concatenation operation.

[0056] Furthermore, the gradient reversal layer GRL reverses the domain classification gradient by λ times during back propagation, and the forward propagation identity transformation formula is:

[0057]

[0058] Furthermore, the temporal position code E pos The frequency parameter ω satisfies:

[0059] For dimension index i∈[0,d model / 2],

[0060] Generate a two-dimensional encoding vector [sin(ω i t),cos(ω i t)].

[0061] The beneficial effects of the present invention are:

[0062] (1) Through the bidirectional spatiotemporal attention network and dynamic domain adaptation mechanism, the multi-source sensor features and process knowledge are effectively integrated, overcoming the shortcomings of traditional methods in modeling nonlinear spatiotemporal correlations, ensuring that the filling results strictly follow the assembly process constraints, and significantly improving the accuracy of data repair under complex dynamic working conditions.

[0063] (2) A semi-supervised learning framework is adopted to combine a small amount of key process labeled data with a large amount of unlabeled data for collaborative training, breaking the dependence of supervised learning on fully labeled data and maintaining high prediction robustness under conditions of scarce labels. It is especially suitable for flexible assembly scenarios with multiple models and small batches.

[0064] (3) Through the domain adaptation mechanism and feature alignment strategy guided by process knowledge, the distribution compatibility of data from different assembly lines is improved, the collaborative use of multi-source heterogeneous data across production lines and equipment is promoted, and the generalization ability of the model is enhanced in the scenarios of equipment replacement and process upgrade.

[0065] (4) Based on the hierarchical dynamic filling strategy and the automated missing detection mechanism, rapid repair of non-ideal data can be achieved, which greatly shortens the assembly quality analysis cycle and provides efficient data support for real-time process adjustment and equipment health management.

[0066] (5) The repaired high-integrity data significantly improves the practicality of assembly parameter optimization suggestions and the accuracy of early fault identification, providing a reliable data basis for quality traceability, process iteration and predictive maintenance in wind power intelligent manufacturing.

[0067] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0069] Figure 1 Flowchart of the present invention;

[0070] Figure 2 Flowchart of missing data detection algorithm for wind turbine assembly process;

[0071] Figure 3 assemble a multi-source data feature extraction flow chart for wind turbines;

[0072] Figure 4 Flowchart of the feature space alignment method based on adversarial domain adaptation;

[0073] Figure 5 Flowchart of the strategy for supplementing non-ideal data with dynamic confidence thresholds. DETAILED DESCRIPTION

[0074] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0075] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0076] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0077] This invention addresses the problems of data missing and outlier mixing during the collection, transmission and storage of multi-source heterogeneous data (such as vibration signals, torque parameters, assembly accuracy, etc.) in the wind turbine assembly process in current domestic wind power plants. It creatively combines the dynamic domain adaptation mechanism with the spatiotemporal attention network, breaking through the technical bottleneck of traditional semi-supervised learning in non-ideal industrial data scenarios.

[0078] The present invention firstly innovatively adopts a spatiotemporal tensor representation method based on data mining to model the multi-source data in the wind turbine assembly process. T×m×s (where T is the time step, m is the sensor dimension, and s is the number of assembly stations), and the confidence weight matrix W is introduced t,k ∈[0,1], achieving refined quantification of assembly data reliability. This representation method breaks through the limitations of traditional time series analysis and can effectively capture the cross-process correlation characteristics in the assembly process.

[0079] A bidirectional network based on spatiotemporal attention (BiSTA-Net) is then used to extract features from the data. By integrating advanced techniques such as multi-scale convolution, bidirectional attention, and residual learning, this network achieves deep mining and efficient characterization of assembly equipment state features. In particular, the bidirectional attention module in the network simultaneously captures the temporal and spatial dependencies of assembly parameters, which is crucial for processing assembly data with complex process coupling.

[0080] Subsequently, to address the distribution discrepancy between labeled and unlabeled data during the assembly process, a distribution alignment model based on dynamic domain adaptation was proposed. This model achieves distribution alignment through the following innovative designs:

[0081] (1) For labeled assembly data, supervised learning is used to accurately learn sample features and classification boundaries;

[0082] (2) For unlabeled data, optimize model parameters by minimizing marginal distribution distance;

[0083] (3) A gradient reversal layer (GRL) is introduced to implement end-to-end adversarial training, enabling the generator to generate synthetic data that is consistent with the distribution of real assembly data.

[0084] Finally, a missing data supplementation strategy with a dynamic confidence threshold is proposed. This strategy selects high-quality samples through a double-verification mechanism. This mechanism ensures the reliability of the pseudo-labeled data used for training and effectively avoids the class imbalance problem.

[0085] In terms of specific implementation, this solution designs a special loss function:

[0086] L total =L sup +β(t)L con +γL align

[0087] Among them, L sup To supervise the loss and ensure the classification accuracy of the labeled data; L con is the consistency regularization loss to enhance the robustness of the model; L align To distribute the alignment loss and reduce the inter-domain differences.

[0088] This method not only improves the integrity and reliability of wind turbine assembly process data but also achieves high-precision data infill under dynamic assembly conditions. Practical applications have demonstrated that this method reduces the error rate by 28% compared to traditional methods in scenarios such as assembly quality traceability and process parameter optimization, providing reliable data support for intelligent wind power manufacturing. This research offers a new solution for the application of semi-supervised learning to industrial assembly data and has significant theoretical and engineering significance.

[0089] The main action process of the present invention is as follows Figure 1 shown.

[0090] 1. Multi-source unlabeled data detection in the wind turbine assembly process

[0091] 1.1 Data Representation and Problem Modeling

[0092] Based on data mining technology, this paper innovatively uses spatiotemporal tensor representation to model the multi-source data in the wind turbine assembly process, breaking through the limitations of traditional time series representation. T×m×s Represents assembly manufacturing data with a time step of T, a sensor dimension of m, and a number of assembly stations of s (such as tower sections, nacelles, blades, etc.). The confidence weight matrix W is introduced t,k ∈[0,1], quantifying data reliability through continuous probability values.

[0093] When W t,k =0, indicating that the k-th dimension sensor data is definitely missing at time point T;

[0094] When W t,k =1, it means that at time point T, the k-th dimension sensor data is complete and reliable;

[0095] When 0 <W t,k When <1, it means that there are some reliability issues in the data.

[0096] Through the above continuous value design, we can distinguish between completely missing data and low-quality data, quantify data reliability more finely, and improve model robustness.

[0097] 1.2 Variational Bayesian Latent Variable Model

[0098] In order to deal with the missing power generation data, a hierarchical latent variable model based on variational Bayes is proposed to model the detection data. The model consists of three core components:

[0099] (1) Global hidden variable z t Obey Gaussian distribution N(μ z ,∑ z ), used to capture system-level dynamic characteristics; (2) local latent variables h t,k Generated by a multi-layer perceptron (MLP), whose input is fused with the global state z t and local historical observation x t-1,k , achieving cross-scale feature interaction;

[0100] (3) Observation model In, f θ Modeling nonlinear mappings for neural networks; The weighted variance term can be used to adaptively adjust the noise level through the confidence weight. This hierarchical structure significantly enhances the model's adaptability to complex working conditions.

[0101] In summary, the hierarchical latent variable model is established as follows:

[0102]

[0103] 1.3. Dynamic Feature Extraction Mechanism

[0104] Then, the hidden spatiotemporal features present in the multi-source wind turbine assembly data are extracted through the following process. This paper creatively proposes an innovative spatiotemporal graph network architecture (MSTG-Net), the core of which is the node feature update mechanism, which dynamically adjusts the information transfer between nodes through the attention coefficient.

[0105] Specifically, the feature h of each node v in the l+1 layerv (l+1) The weighted features of the neighborhood nodes u∈N(v) are summed by the activation function σ (usually ReLU), and the node feature update formula is as follows:

[0106]

[0107] Where σ is the ReLU activation function, N(v) is the spatiotemporal neighborhood of node v, and the weight α vu Generated by the attention mechanism calculation, the calculation uses the improved LeakyReLU activation function (with leaky linear rectification function), the weight α vu It is given by:

[0108]

[0109] Node features are transformed and concatenated using a learnable parameter vector a and a weight matrix W, and then normalized using a softmax function. This design enables the model to adaptively capture the spatiotemporal dependencies between different assembly nodes during wind turbine assembly. This is particularly advantageous for understanding the non-uniform spatiotemporal correlations exhibited by assembly parameters (such as bolt torque, alignment accuracy, and vibration amplitude) during multi-step collaboration.

[0110] 1.4. Missing Detection Driven by Random Processes

[0111] For missing data detection, the present invention innovatively adopts a non-homogeneous Poisson process to model missing data, and its intensity function is given by the following formula:

[0112]

[0113] Among them, β0 is the baseline missing rate, β1x t-1,k The impact of historical observations, This missing detection mechanism comprehensively considers the above factors to realize the power data missing judgment, that is, when λ(t,k)>τ(t), the marked data is missing data, and the missing data is recorded as the unlabeled dataset X U .

[0114] The detection rules are as follows:

[0115]

[0116] Where τ(t) is a dynamic threshold that is adaptively adjusted with the training rounds.

[0117] This method significantly improves the accuracy of missing data detection by comprehensively considering historical data and real-time changes. The algorithm flow of missing data detection in the fan assembly process is as follows: Figure 2 shown.

[0118] 2. Data feature extraction

[0119] In the missing data filling task of the fan assembly process, given a labeled dataset and unlabeled datasets in represents the i-th labeled sample, represents the jth unlabeled sample, It is the process quality label of the corresponding marked sample (such as "pass / fail" or specific deviation value). L and N U Represents the number of labeled samples and unlabeled samples respectively. Since unlabeled samples lack clear label information, these data cannot be directly used in traditional supervised learning process. However, the use of semi-supervised learning technology can effectively mine the implicit feature information in unlabeled data, thereby improving the overall performance of the model. In order to achieve the high-dimensional input data X L and X U To convert it into low-dimensional data suitable for classification and feature alignment, the data must first be preprocessed and transformed using a feature extractor.

[0120] This invention innovatively proposes a bidirectional spatiotemporal attention network as a feature extraction model. By integrating advanced technologies such as multi-scale convolution, bidirectional attention mechanism and residual learning, this model achieves deep mining and efficient characterization of multi-scale workstation data state features.

[0121] 2.1 Spatiotemporal Feature Embedding Layer

[0122] First, a multi-scale spatiotemporal feature embedding layer is constructed. This embedding layer uses a one-dimensional convolution kernel (kernel_size=5) to map the original input to a high-dimensional feature space. The position encoding is generated using a hyperbolic sine function to preserve the temporal information. Its formula is:

[0123]

[0124] in, is the input signal at time t, with dimension d_in = 64;

[0125] Conv1D(·) is a one-dimensional convolution with kernel width 5 and output dimension d_model = 64;

[0126] HE(·) is the hyperbolic sine position encoding function, and its calculation formula is:

[0127]

[0128] Where loc is the position index of the data in the input sequence (time sequence position), i is the dimension index, i∈(0,d model / 2), d model is the characteristic dimension, and in this invention, d model =64.

[0129] 2.2 Bidirectional Attention Calculation Mechanism

[0130] Next is the network's bidirectional attention calculation module, which is designed to capture temporal dependencies. The Softmax function is a normalized exponential function commonly used in neural networks, and its mathematical expression is:

[0131]

[0132] where z i =(z1,z2,...,z K ) is the input vector, K is the total number of categories, and in the bidirectional attention module of the present invention, the Softmax function is used to convert the attention score into a probability distribution. The calculation formula of the forward attention is:

[0133]

[0134] Similarly, the calculation formula for backward attention is:

[0135]

[0136] Among them, Q f , Q b is the query vector at the current moment, K f is the key vector matrix of historical moments, K b is the future key vector matrix, W q , W k is a trainable parameter matrix. Here, (5) and (6) use the same dimension scaling to ensure numerical stability. The denominator is a scaling factor used to stabilize the gradient.

[0137] 2.3 Residual Feature Aggregation

[0138] Afterwards, the residual feature aggregation module designed by the present invention realizes efficient feature fusion of deep networks through skip connections. In each network layer (l∈{1,2,3}), the current layer outputs feature Generate it by following these steps:

[0139] 1. Bidirectional attention fusion: forward attention weight and backward attention weights Splicing along the feature dimension to obtain joint attention features

[0140] 2. Linear transformation: through learnable parameter matrix Perform feature projection.

[0141] 3. Residual connection: the original input features Added with the attention feature to preserve the underlying information.

[0142] 4. Layer Unitization: The layer unitization function LayerUnit is used to stabilize the model training process. Its calculation formula is:

[0143]

[0144] Where μ and σ are the mean and standard deviation of the eigenvalues, and γ and β are learnable parameters, so the output characteristics of the current layer are obtained:

[0145]

[0146] Through this update mechanism, the labeled and unlabeled data are integrated with advanced technologies such as multi-scale convolution, bidirectional attention mechanism and residual learning, which enables deep mining and efficient representation of data features, and provides strong support for subsequent domain adversarial learning classifiers and alignment models. L and unlabeled data X U After being processed by the bidirectional spatiotemporal attention network feature extractor, the final extracted low-dimensional feature annotation data H is obtained. L and unlabeled data H U :

[0147] H L =BiSTA-Net(X L ), H U =BiSTA-Net(X U ) (14)

[0148] The overall process of data feature extraction is as follows Figure 3 shown.

[0149] 3. Feature Space Alignment Method Based on Adversarial Domain Adaptation

[0150] To achieve efficient matching of cross-domain data distribution, this paper innovatively proposes a dual-channel adversarial optimization framework, which completes the feature space alignment of the source and target domains through collaborative adversarial training of the feature encoder E and the domain discriminator C.

[0151] Cross-domain feature alignment is a core issue in transfer learning. The key lies in constructing a domain-invariant feature space so that data from different distributions have similar statistical properties in this space. This method has the following innovative features:

[0152] 1. Adopt bidirectional feature projection mechanism;

[0153] 2. Introducing dynamic weight balancing strategy;

[0154] 3. Achieve end-to-end gradient optimization.

[0155] The objective function L of the adversarial learning proposed by this method is adv It can be expressed as:

[0156]

[0157] Among them, x s is the source domain sample, x t is the target domain sample, E is the feature encoder, and C is the domain discriminator. In the specific implementation, the encoder E optimizes the following loss function L E To confuse domain features:

[0158] L E =L task -L adv (16)

[0159] In the above formula, λ is the dynamic adjustment coefficient, L task The gradient reversal layer (GRL) is introduced to automatically reverse the gradient signal of the domain classifier during back propagation, thus achieving single-stage end-to-end training.

[0160] The feature alignment process can be formally described as:

[0161]

[0162] where z s and z t Respectively represent the feature representation of source domain and target domain samples after passing through the encoder; d s and d t is the domain discrimination output of the corresponding domain classifier for the feature; θ e and θ c denote the trainable parameters of the encoder and classifier respectively.

[0163] E(x s θ e ):x s Feature mapping process through the encoder

[0164] E(x t θ e ):x t Feature maps through the same encoder

[0165] C(z s θ c ): Source domain feature z s Domain discrimination result (the value is usually ∈ [0,1])

[0166] C(z s θ c ): target domain feature z t Domain discrimination results

[0167] Improved domain difference loss function L align It consists of two terms: the first term performs a weighted sum of the differences in the mean values ​​of the features at each layer, and the second term directly calculates the difference in the global covariance matrix. The formula is as follows:

[0168]

[0169] Where i is the i-th hidden layer of the neural network, N is the number of neural network layers, and Represents the mean vector of the i-th layer of the source domain and feature domain respectively, ∑ s and ∑ t are the covariance matrices of the source and target domain features, ||·||2 is the L2 norm, ||·|| F is the Frobenius norm. This loss function achieves deep distribution alignment between the source domain and the target domain by jointly optimizing the first-order statistics (mean) and second-order statistics (covariance) of the feature distribution.

[0170] During the model training process, an alternating optimization method is used to optimize the trainable parameters θ of the encoder and domain classifier. e and θ c Update, the specific strategy is as follows:

[0171] Encoder parameter θ e renew

[0172] The encoder parameter update comprehensively considers the dual effects of the main task loss and the domain alignment loss. The negative sign in the formula indicates the direction of gradient descent optimization. This design allows the encoder to actively narrow the distribution gap between domains while optimizing the performance of the main task. The formula is as follows:

[0173]

[0174] Domain classifier parameters θ c renew

[0175] The domain classifier uses a standard gradient ascent strategy. The positive sign indicates the need to enhance domain discrimination ability, forming an adversarial training relationship with the encoder parameter update. The formula is as follows:

[0176]

[0177] Where η is the learning rate and γ is the dynamic balance factor. This update mechanism has been proven to be effective in wind turbine fault diagnosis.

[0178] This scheme significantly improves the distribution alignment effect of cross-domain data by combining end-to-end training of gradient reversal layer (GRL) and adversarial training. The overall process of this method is as follows: Figure 4 shown.

[0179] 4. Dynamic Confidence Threshold Non-ideal Data Supplementation Strategy

[0180] Based on the distribution-aligned labeled data and unlabeled data, this paper proposes an innovative adaptive probability threshold collaborative repair mechanism. The overall process of the method is as follows: Figure 5 As shown in the figure, a semi-supervised learning framework is used to accurately fill in the missing data of the wind turbine assembly process. The specific implementation of this mechanism is as follows:

[0181] 4.1 Double Verification Sample Selection Mechanism

[0182] First, an improved fuzzy probability estimator is used to calculate the unlabeled samples The probability distribution of category belonging to is:

[0183]

[0184] in, Represents the embedding vector of the jth unlabeled sample in the feature space, MLP φ For a parameterized multilayer perceptron, the output dimension is the same as the number of missing data categories.

[0185] Then the feature space alignment index ρ is introduced j , quantifying the degree of match between the sample and the target domain:

[0186]

[0187] Among them, Proj ψ is a learnable projection network, μ L The class center of the labeled data features. The range of this indicator is [0,1]. The larger the value, the closer it is to the target domain distribution.

[0188] The confidence threshold function ξ(t) and distribution matching threshold η(t) are selected as follows:

[0189]

[0190] Where T is the total number of training rounds.

[0191] Finally, the sample selection condition is updated as follows:

[0192]

[0193] Where I represents a binary function: if Then it returns 1, otherwise it returns 0; similarly, if ρj >η(t), then return 1, otherwise return 0. Therefore, only when both conditions are 1 at the same time, the wind turbine data sample will be selected for training.

[0194] 4.2 Collaborative Repair Loss Function Design

[0195] To efficiently fill in the missing data in the wind turbine assembly process, this solution constructs a triple synergistic loss function that utilizes information from both labeled and unlabeled data through end-to-end optimization. The loss function consists of the following three core components:

[0196] 4.2.1 Supervision Loss

[0197] The supervised loss is applied to the labeled dataset H L To ensure the basic performance of the model on the labeled data:

[0198]

[0199] Among them, N L To mark the number of samples, is the true label of the i-th labeled sample, represents the feature representation of the i-th labeled sample, and C is the classifier network. This loss function uses the standard cross-entropy form to force the model to accurately predict the category of wind turbine data samples with known labels.

[0200] 4.2.2 Consistency Regularization Loss

[0201] This part designs consistency constraints for unlabeled data to enhance the robustness of the model to data perturbations:

[0202]

[0203] Among them, N U is the number of labeled samples, Aug is the data enhancement operation in the feature space, which is done by forcing the original features Compared with the enhanced version Aug, the prediction consistency improves the model generalization ability.

[0204] 4.2.3 Distribution Alignment Loss

[0205] This part is used to solve the domain offset problem between labeled data and unlabeled data:

[0206] L align =MMD 2 (H L ,H U )+λCORAL(H L ,H U ) (28)

[0207] Among them, MMD is responsible for measuring H in the reproducing kernel Hilbert space L and H U CORAL aligns the source and target domain distributions by matching the second-order statistics (covariance matrix), and λ is the balance coefficient.

[0208] 4.2.4 Dynamic Weighted Total Loss

[0209] The weighted sum of the above three parts is the final loss L total :

[0210] L total =L su p+β(t)L con +γL align (29) Among them, is the dynamic weight of the consistency regularization loss, and γ is the fixed weight of the distribution alignment loss.

[0211] 4.3 Dynamic Filling Strategy Implementation

[0212] This solution proposes a hierarchical dynamic filling strategy:

[0213] 4.3.1 Missing pattern detection

[0214] The improved hidden Markov model (HMM) is used to identify missing states, and its observation probability model is:

[0215]

[0216] Among them, K is the number of hidden states, π k is the prior probability of state transition, A k is the state-related autoregressive matrix, ∑ k is the state-dependent covariance matrix.

[0217] 4.3.2 Missing Data Filling Mechanism

[0218] The filling mechanism supplements the missing wind turbine equipment data by fusing known data features and temporal position information, combining the attention mechanism and the generative adversarial network (GAN). The mathematical model of the data generator is:

[0219]

[0220] The symbols in the formula are defined as follows: represents the generated filling sequence from t1 to t2; z represents the latent space vector, which obeys the standard normal distribution; h known is the feature set from the known time (extracted by BiSTA-Net); θ G is the trainable parameter of the generator; Epos It is the temporal position encoding, and the calculation formula is as follows:

[0221]

[0222] The generator integrates an attention module, and its calculation process is as follows:

[0223]

[0224] where K t' is the key vector, α t,t' is the attention weight of time point t to t'.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for filling missing data in the fan assembly process based on semi-supervised learning, characterized by: Including the following step: S1: Perform spatiotemporal tensor modeling on the multi-source heterogeneous data in the wind turbine assembly process and construct a three-dimensional tensor X∈R T×m×s , where T represents the time step, m represents the sensor dimension, and s represents the number of assembly stations; And introduce the confidence weight matrix W t,k ∈[0,1] quantifies data reliability, where W t,k =0 means data is missing, W t,k =1 means the data is complete, 0<W t,k <1 indicates that the data have some reliability issues; S2: Feature extraction of the three-dimensional tensor is performed through the bidirectional spatiotemporal attention network BiSTA-Net, which adopts a fusion structure of multi-scale convolution, bidirectional attention mechanism and residual learning, and outputs the labeled dataset H L and the unlabeled dataset H U Low-dimensional feature representation of ; S3: Based on the adversarial domain adaptation mechanism, the labeled dataset H L and the unlabeled dataset H U Perform feature space alignment and implement adversarial training between the encoder and domain discriminator through the gradient reversal layer GRL to minimize the inter-domain distribution difference between labeled data and unlabeled data; S4: Use a dynamic confidence threshold strategy to screen high-reliability unlabeled data and calculate the probability distribution of sample categories through a double verification mechanism. and feature space alignment ρ j , combining dynamic thresholds ξ(t) and η(t) to generate pseudo-label data; S5: Construct a supervised loss L sup , consistency regularization loss L con and distribution alignment loss L align The collaborative loss function L total =L sup +β(t)L con +γL align , fill model parameters through end-to-end training optimization; S6: Use the trained model to dynamically fill in the missing data in layers, and combine the missing pattern detection of the hidden Markov model (HMM) and the data generator of the generative adversarial network (GAN) to repair the missing values.

2. The method for filling missing data in the fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S1, the spatiotemporal tensor modeling further comprises: Using the variational Bayesian hierarchical latent variable model, define the global latent variable z t ~N(μ z ,∑ z ) and local latent variables h t,k ~g φ (z t ,x t-1,k ), the observation model is where f θ is the neural network mapping function, is the weighted variance term.

3. The method for filling missing data in the fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S2, the specific structure of the bidirectional spatiotemporal attention network BiSTA-Net includes: (a) Multi-scale spatiotemporal feature embedding layer, using a one-dimensional convolution kernel, i.e. kernel_size = 5, to map the input to d model = 64-dimensional feature space, and superimposed hyperbolic sine position encoding: (b) Bidirectional attention module, forward attention weight and backward attention weights splicing and fusion; (c) Residual aggregation layer, through Normalize the skip connection features.

4. The method for filling missing data in a fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S3, the adversarial domain adaptation mechanism is specifically implemented as follows: (a) Loss function of domain discriminator C: (b) Optimization goal of encoder E: THE E =L task -λL adv Where λ is the dynamic adjustment coefficient; (c) Domain Alignment Loss: in and is the mean of the i-th layer features of the source domain and the target domain, ∑ s and ∑ t is the covariance matrix.

5. The method for filling missing data in the fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S4, the dynamic confidence threshold strategy satisfies: Sample selection conditions in: and is the feature space alignment indicator.

6. The method for filling missing data in a fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S5, the collaborative loss function includes: (a) Supervision losses: (b) Consistency regularization loss: (c) Distribution Alignment Loss: L align =MMD 2 (H L ,H U )+λCORAL(H L ,H U ) Among them, MMD is the maximum mean difference and CORAL is the correlation alignment loss.

7. The method for filling missing data in a fan assembly process based on semi-supervised learning according to claim 1 is characterized in that: In S6, the hierarchical dynamic filling specifically includes: (a) Missing pattern detection uses Hidden Markov Model HMM, and the observation probability is: (b) Data generator G(z|{h known },E pos θ G ) Fusion of known features h known and temporal position encoding:

8. The method for filling missing data in the fan assembly process based on semi-supervised learning according to claim 3 is characterized by: In the bidirectional attention module, the attention coefficient is calculated using: in, is the learnable parameter vector, is the weight matrix, and || represents the vector concatenation operation.

9. The method for filling missing data in a fan assembly process based on semi-supervised learning according to claim 4, characterized in that: The gradient reversal layer GRL reverses the domain classification gradient by λ times during back propagation, and the forward propagation identity transformation formula is:

10. The method for filling missing data in a fan assembly process based on semi-supervised learning according to claim 7, characterized in that: The temporal position code E pos The frequency parameter ω satisfies: For dimension index i∈[0,d model / 2], Generate a two-dimensional encoding vector [sin(ω i t),cos(ω i t)].

Images