Data analysis method and device based on feature decoupling and fusion, equipment and medium

Abstract

The application relates to the technical field of data analysis, and discloses a data analysis method, device, equipment and medium based on feature decoupling and fusion, which comprises the following steps: acquiring input data of multiple data channels and performing pretreatment, extracting a multi-phase fusion feature vector, constructing a multi-link adaptive attention matrix, performing context block processing and decoupling through a double-path convolution operation, performing cross-block information fusion, refining trend features and periodic features, and finally generating an analysis result through multi-channel hybrid reconstruction. Through the method of multi-scale feature decoupling and fusion, long-term dependence and local mutation features in time series data can be accurately captured, the deficiencies in the prior art in processing complex multivariate data, time scale concurrency and noise interference are overcome, the prediction accuracy and generalization ability are significantly improved, and the needs of high-precision prediction and real-time fault diagnosis are met.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a data analysis method, apparatus, device, and storage medium based on feature decoupling and fusion. Background Technology

[0002] Modern industrial production increasingly relies on improvements in automation and intelligence, requiring enterprises to monitor production processes in real time, make accurate predictions, and make intelligent decisions. However, current technologies have many shortcomings when processing complex time-series data in industrial scenarios. Traditional time series analysis methods, such as ARIMA and exponential smoothing, while widely used in simple scenarios, often reveal several key problems when facing industrial environments with multivariate coupling, multi-timescale concurrency, and complex noise interference.

[0003] First, many existing models cannot effectively handle complex industrial production data that are high-dimensional, multi-scale, and multivariate. Existing statistical models and basic machine learning models typically assume that the data has strong stationarity or a single periodicity, but real-world industrial data often exhibits complex nonlinear dynamic characteristics. This leads to incomplete feature extraction by the models, making it impossible to capture long-term dependencies between data points. Furthermore, these methods have poor adaptability to data, especially in complex industrial systems, and cannot flexibly cope with the complex and variable nature of signals.

[0004] Secondly, many industrial fault monitoring and prediction methods rely on expert experience and simple rule-based judgments, or alarm mechanisms based on fixed thresholds. While these methods may be effective in certain specific situations, they are relatively slow and insensitive when faced with the frequent occurrence of overlapping multi-frequency components, periodic changes, and sudden abnormal events in modern industrial production. These methods are less effective at detecting minor fault characteristics in the early stages and lack the ability to predict new fault modes, greatly limiting the potential of intelligent manufacturing.

[0005] Furthermore, although deep learning has demonstrated strong potential in processing complex time-series data, especially on large-scale datasets, and some studies have applied deep learning to industrial production monitoring, several technical limitations still exist. Many existing time-series prediction models tend to adopt a single type of feature extraction method or overemphasize long-term dependencies while neglecting important local features in the short time domain. This results in models being unable to effectively capture complex periodic fluctuations, seasonal behavioral patterns, and sudden abnormal events in multi-scale data and dynamically changing environments. These problems directly affect the real-time performance and reliability of deep learning-based prediction models in high-value areas such as industrial fault diagnosis and production optimization decision-making.

[0006] Finally, although large deep learning models have achieved remarkable results in some fields, in industrial scenarios, the models have high computational resource requirements, low training efficiency, poor flexibility, and a large number of model parameters, making it difficult for them to efficiently and economically complete intelligent monitoring and accurate prediction in some industrial enterprises with massive amounts of data.

[0007] Therefore, designing an intelligent analysis method that can effectively extract multi-scale dynamic features from industrial time-series data, accurately capture long-term dependencies and local mutation information between signals, and possess high-precision prediction, high-efficiency analysis, and good adaptability to different industrial scenarios has become an urgent need for current technological development. Summary of the Invention

[0008] The main objective of this invention is to provide a data analysis method, apparatus, device, and storage medium based on feature decoupling and fusion, aiming to solve the technical problem that existing technologies are unable to effectively decouple multi-scale periodic factors and trend features in time series data and achieve dynamic adaptive fusion, resulting in insufficient modeling of complex time dependencies.

[0009] To achieve the above objectives, this invention provides a data analysis method based on feature decoupling and fusion, comprising: The system acquires input data from multiple data channels and performs preprocessing operations on the input data to generate preprocessed data. The preprocessed data is subjected to phase domain joint transformation processing to extract multi-phase fusion feature vectors of each data channel; Based on the multi-phase fusion feature vector, a multi-link adaptive attention matrix is ​​constructed to obtain the fusion attention matrix; The multi-phase fusion feature vector and the fusion attention matrix are processed by context block segmentation, and decoupled within each context block by dual-path convolution operation to generate decoupled periodic factor features and trend slope features. Cross-block information fusion is performed on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The fused decoupled features are refined through trend-based attention pathways and periodic attention pathways to obtain refined trend features and refined periodic features. The refining trend characteristics and the refining periodic characteristics are reconstructed using a multi-channel hybrid method to obtain the analysis results.

[0010] Furthermore, to achieve the above objectives, the present invention provides a data analysis device based on feature decoupling and fusion, comprising: The data acquisition and preprocessing module is used to acquire input data from multiple data channels and perform preprocessing operations on the input data to generate preprocessed data. The feature extraction module is used to perform phase domain joint transformation processing on the preprocessed data to extract multi-phase fusion feature vectors of each data channel; The attention mechanism module is used to construct a multi-link adaptive attention matrix based on the multi-phase fusion feature vector to obtain the fusion attention matrix; The feature decoupling module is used to perform context block processing on the multi-phase fusion feature vector and the fusion attention matrix, and decouple them in each context block through dual-path convolution operation to generate decoupled periodic factor features and trend slope features. The information fusion module is used to perform cross-block information fusion on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The feature refinement module is used to refine the fused decoupled features through a trend attention path and a periodic attention path, respectively, to obtain refined trend features and refined periodic features. The result reconstruction module is used to perform multi-channel hybrid reconstruction of the refining trend characteristics and the refining periodic characteristics to obtain analysis results.

[0011] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a data analysis program based on feature decoupling and fusion stored in the memory and executable on the processor, wherein when the data analysis program based on feature decoupling and fusion is executed by the processor, it implements the steps of the data analysis method based on feature decoupling and fusion as described above.

[0012] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a data analysis program based on feature decoupling and fusion, wherein when the data analysis program based on feature decoupling and fusion is executed by a processor, it implements the steps of the data analysis method based on feature decoupling and fusion as described above.

[0013] Beneficial Effects: This invention relates to the field of data analysis technology, and discloses a data analysis method, apparatus, device, and medium based on feature decoupling and fusion. The method includes: acquiring input data from multiple data channels and preprocessing it; extracting multi-phase fusion feature vectors; constructing a multi-link adaptive attention matrix; performing contextual block processing and decoupling through dual-path convolution operations; performing cross-block information fusion; refining trend features and periodic features; and finally generating analysis results through multi-channel hybrid reconstruction. This invention, through a multi-scale feature decoupling and fusion method, can accurately capture long-term dependencies and local mutation features in time-series data, overcoming the shortcomings of existing technologies in handling complex multivariate data, concurrent time scales, and noise interference. This significantly improves prediction accuracy and generalization ability, meeting the needs of high-precision prediction and real-time fault diagnosis. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a data analysis method based on feature decoupling and fusion according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the data analysis method based on feature decoupling and fusion according to the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data analysis device based on feature decoupling and fusion of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0016] The data analysis method based on feature decoupling and fusion provided in this invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can obtain input data from multiple data channels through the user terminal and perform preprocessing, extract multi-phase fusion feature vectors, construct a multi-link adaptive attention matrix, perform context block processing and decouple through dual-path convolution operations, perform cross-block information fusion, refine trend features and periodic features, and finally generate analysis results through multi-channel hybrid reconstruction. This invention, through a multi-scale feature decoupling and fusion method, can accurately capture long-term dependencies and local mutation features in time-series data, overcoming the shortcomings of existing technologies in handling complex multivariate data, time-scale concurrency, and noise interference, thereby significantly improving prediction accuracy and generalization ability, and meeting the needs of high-precision prediction and real-time fault diagnosis. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster composed of multiple servers. The invention will be described in detail below through specific embodiments.

[0017] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the data analysis method based on feature decoupling and fusion provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0018] like Figure 2 As shown, the data analysis method based on feature decoupling and fusion proposed in this invention includes the following steps: S10: Acquire input data from multiple data channels and perform preprocessing operations on the input data to generate preprocessed data; In this embodiment, acquiring input data from multiple data channels and preprocessing it to generate preprocessed data is a fundamental step in time-series data analysis, ensuring that subsequent feature extraction and analysis can be performed on a high-quality data basis. First, the input data from multiple data channels typically originates from different types of sensors, monitoring devices, or historical data records. This data may come from industrial equipment, sensor networks, smart meters, etc., and the data dimensions and types may include temperature, humidity, pressure, vibration, current, etc. This data is sometimes presented in the form of time-series data, meaning it contains information that changes over time and can reflect the status and performance of the equipment or system.

[0019] Preprocessing is the process of transforming input data into clean, standardized, denoised, and missing value-impairing data that can be used for analysis. Specifically, preprocessing steps include data cleaning, missing value impairing, outlier detection, denoising, and standardization. Data cleaning involves removing invalid, duplicate, or erroneous data records. For example, if a data point was erroneous during acquisition, the preprocessing step deletes or corrects it to ensure data reliability. Missing value impairing refers to filling in the missing parts of partially missing time series data using appropriate interpolation methods to ensure data continuity. Common interpolation methods include linear interpolation and spline interpolation; the most suitable method can be selected based on actual needs. Outlier detection identifies and handles outliers that deviate from the normal data distribution; commonly used techniques include statistical methods such as standard deviation and IQR.

[0020] The purpose of data standardization is to transform input data at different scales into the same standard scale to eliminate the influence of different units, enabling fair comparison and fusion of data from different channels. For example, standard deviation standardization (Z-score) or minimum-maximum standardization can be used to transform the values ​​of each data channel to the same scale. Denoising, on the other hand, removes noise components from the data using filtering algorithms (such as Kalman filtering, low-pass filtering, etc.) to ensure that the data is closer to the true signal.

[0021] These preprocessing operations ultimately generate clean, continuous, standardized, and denoised preprocessed data, providing a high-quality data foundation for subsequent feature extraction, model training, and prediction.

[0022] In industrial applications, data may originate from multiple sensors or devices, each with varying degrees of accuracy and measurement error. Therefore, data cleaning must begin by removing invalid data caused by equipment malfunctions or transmission errors. For example, if a sensor malfunctions during measurement, resulting in extreme output values, the cleaning process will automatically detect and process these extreme values, replacing them with appropriate values ​​(e.g., interpolation or removal). Industrial equipment may experience occasional data gaps, such as data loss due to communication interruptions. In such cases, interpolation techniques can be used to supplement the data. For equipment with strong periodic characteristics, interpolation methods based on periodic trends may be more effective than linear interpolation, ensuring that the supplemented data retains the characteristic fluctuations of the equipment. In complex industrial environments, outliers or noise may appear in the data. These outliers are often irregular and can affect subsequent analysis results. For example, if a pressure sensor reading suddenly spikes, it may be due to sensor malfunction or changes in the external environment. Statistical methods (such as box plots or standard deviation methods) can effectively detect and remove these outliers. For data from multiple channels, standardization must be performed during preprocessing to ensure that data from different sensors can be processed on the same scale. Through standardization, the values ​​of each data channel are transformed into the same range, eliminating scale bias between features. Denoising is then achieved by using appropriate filters, such as Kalman filters, to remove high-frequency noise while preserving the essential characteristics of the signal.

[0023] This embodiment effectively eliminates noise, repairs missing values, and eliminates scale differences by preprocessing the input data, thus providing clean and consistent input data. This preprocessed data provides a reliable foundation for subsequent feature extraction, modeling, and analysis, improving the accuracy and robustness of the entire time-series data analysis system.

[0024] S20, perform phase domain joint transformation processing on the preprocessed data to extract the multi-phase fusion feature vector of each data channel; In this embodiment, performing phase domain joint transform processing on the preprocessed data to extract multi-phase fusion feature vectors from each data channel is a key step in time series data analysis. This aims to enhance the signal representation capability through phase domain analysis and feature fusion, thereby extracting more accurate features for subsequent analysis. Phase domain joint transform processing can handle multivariate data from multiple channels and effectively capture periodicity, trends, and other dynamic features in the data using phase information.

[0025] Phase-domain joint transformation is the process of converting time-domain data into the phase-domain. This process separates the frequency and phase information of the data and performs joint processing across multiple channels. Through phase-domain transformation, periodic components and long-term trend components in the data can be identified and extracted, overcoming the limitations of traditional methods in multivariate and multi-scale time-series data analysis. For example, when processing vibration signals from equipment, frequency and phase components can represent different dynamic characteristics. Especially in cases of high noise or complex coupling between signals, phase-domain transformation can effectively separate and enhance these characteristics.

[0026] The extraction of multi-phase fusion feature vectors involves fusing the phase maps of each data channel to generate a feature vector that integrates information from multiple channels. This process not only preserves the independent information of each channel but also strengthens the correlation between signals through the fusion mechanism. Multi-phase fusion can extract richer and more accurate feature representations by integrating the phase information of each channel. For multiple data channels, each channel typically undergoes independent phase transformation processing, and then the phase maps of each channel are fused according to certain rules to generate the final multi-phase fusion feature vector. These feature vectors can be used for subsequent pattern recognition, predictive analysis, and anomaly detection.

[0027] For each data channel, phase-domain transformation techniques such as Continuous Wavelet Transform (CWT) or Fourier Transform are used to transform the time-series data from the time domain to the phase domain. Continuous Wavelet Transform is very common in time-frequency analysis; it enables multi-resolution analysis of signals based on frequencies at different scales, making it particularly suitable for processing non-stationary signals. By selecting an appropriate mother wavelet function (such as the Morlet wavelet or the Mexican Hat wavelet), local frequency information in the signal can be effectively extracted to obtain the phase diagram.

[0028] After multi-scale decomposition of the phase map of each data channel, a weighted fusion strategy is used to fuse the phase maps of different channels. The weighted fusion method can control the contribution of different channels to the final feature vector by setting weights (e.g., based on the importance of each channel's signal). For example, in industrial equipment monitoring, some sensors may be more important than others and should be assigned larger weights. The fused feature vector can more accurately reflect the overall state of the equipment.

[0029] Finally, by concatenating or weighting the fused feature vectors of each channel, a multi-phase fused feature vector that integrates information from multiple channels is generated. This feature vector can effectively represent various state information of the device or system and provides sufficient feature details for subsequent analysis and prediction.

[0030] This embodiment effectively extracts more refined and richer feature representations from multiple data channels by performing phase-domain joint transformation on the preprocessed data. These feature vectors not only preserve the independent information of each channel, but also strengthen the interrelationships between channels through the phase fusion process, which is helpful for subsequent prediction and fault diagnosis. Especially in complex industrial scenarios, the fusion of multiple sensor data can improve the robustness and accuracy of the model.

[0031] S30, construct a multi-link adaptive attention matrix based on the multi-phase fusion feature vector to obtain the fusion attention matrix; In this embodiment, a multi-link adaptive attention matrix is ​​constructed based on the multi-phase fusion feature vector to obtain the fusion attention matrix. This involves using a multi-link mechanism and an adaptive attention mechanism to process multi-channel data. The multi-phase fusion feature vector contains data features extracted from different angles and scales, and these features have different weights and importance in the temporal and spatial dimensions. By constructing the multi-link adaptive attention matrix, the model can automatically adjust the weights of different data channels and time periods according to the characteristics of the data, thereby achieving attention to key features and further improving the data representation capability. The purpose of this step is to combine features from different time scales and different data channels through an attention mechanism in a weighted manner to form a fusion matrix that can comprehensively consider multiple aspects of information.

[0032] The introduction of a multi-link mechanism allows each link to focus on learning specific features from the input data, while different links operate independently, each focusing on certain dimensions of the data (such as periodic changes, trend changes, etc.). Within each link, the attention mechanism dynamically adjusts weights based on the characteristics of the input data to emphasize the features most important to the current task. In this way, the attention matrix reflects the importance of different regions and features in the data and passes this information to subsequent processing. The fusion attention matrix combines the attention weights learned from each link to comprehensively reflect the multi-dimensional feature representation of the input data.

[0033] In terms of implementation, multiple parallel attention pathways are first configured. Each pathway independently learns different feature interaction relationships and generates its own attention matrix. Within each pathway, an adaptive mechanism automatically adjusts the weights of each feature based on data at different time scales. In this way, different pathways can focus on different levels of data features; for example, some pathways can focus on periodic fluctuations, while others focus on trend changes. Then, the attention matrices generated by each pathway are weighted and superimposed to generate a fused attention matrix. This matrix integrates the attention information from all pathways, enabling subsequent feature decoupling and fusion operations to more accurately identify key features.

[0034] This embodiment, by adaptively adjusting the weights of each link, can more flexibly handle complex data at different time scales. It can automatically adjust the model's focus based on dynamic changes in the data, thereby improving the model's ability to capture complex patterns. In this way, when processing multidimensional data with complex temporal relationships, the model can more accurately identify potential periodic fluctuations and trend changes, thus providing more reliable feature support for subsequent predictions and decisions.

[0035] S40, perform context block processing on the multi-phase fusion feature vector and the fusion attention matrix, and decouple them in each context block through dual-path convolution operation to generate decoupled periodic factor features and trend slope features; In this embodiment, the multi-phase fusion feature vector and the fusion attention matrix are processed through contextual block segmentation, and decoupled within each context block using a dual-path convolution operation. This generates decoupled periodic factor features and trend slope features, an operation aimed at accurately extracting and decoupling features at different levels in time-series data. The multi-phase fusion feature vector typically contains different feature information from multiple data channels, which exhibit complex variations in both the time and frequency domains. The fusion attention matrix, through a weighting mechanism, emphasizes important feature regions. The key to the contextual block segmentation operation is dividing the time-series data according to specific time windows. This allows for the processing of local features within each window, followed by further data processing using subsequent convolution operations.

[0036] During the segmentation process, the multi-phase fusion feature vector and the fusion attention matrix must first be segmented synchronously to ensure their consistency in the temporal dimension. Specifically, after segmentation, each context block contains a multi-phase fusion feature vector and a fusion attention matrix within the same time range. This segmentation ensures that subsequent processing will not result in information loss due to data misalignment. Typically, this segmentation employs a sliding window technique, where the window size and step size are adjusted according to the requirements of the actual application scenario. The step size is smaller than the window size to ensure partial overlap between adjacent blocks, thereby fully utilizing the temporal relationships in the data.

[0037] After contextual segmentation, the data within each context block is decoupled through a dual-path convolution operation. Dual-path convolution refers to extracting features through two separate convolutional paths: one path extracts periodic factor features, and the other extracts trend slope features. These two paths use convolutional kernels of different sizes, with the smaller kernel capturing high-frequency periodic changes and the larger kernel extracting low-frequency trend changes. This decoupling operation, by processing periodic and trend features separately, enables more accurate extraction of signals with different properties from the data, thereby improving the processing capability for time-series data.

[0038] In this step, every detail of the convolution operation is crucial. For extracting periodic factor features, using a smaller convolution kernel can efficiently capture frequently changing patterns within a short time window, while a larger convolution kernel is used to capture slowly changing trends. In practical applications, the size and stride of the convolution kernel can be adjusted according to the characteristics of the data and the requirements to obtain the best results in different scenarios.

[0039] The periodic factor features and trend slope features extracted through dual-path convolution operations are orthogonally constrained to ensure that these two features can independently express different dimensions of the data, thereby avoiding redundant information between features. After dimensionality reduction and compression, decoupled periodic factor features and trend slope features are generated. These features have stronger representativeness and discriminative power, providing clearer and more effective input data for subsequent analysis and prediction.

[0040] During implementation, optimizations can be made based on the specific needs of the application. For example, when processing operational data from different industrial equipment, the window size and step size can be adjusted according to the equipment's operating characteristics to ensure that key patterns in the data can be effectively extracted. In the healthcare field, specific configurations can be applied to data related to different diseases or symptoms to ensure that the model can accurately identify key features associated with changes in the condition.

[0041] This embodiment effectively separates short-term fluctuations and long-term trends in time series data by decoupling periodic factor features and trend slope features. This enables the model to more accurately identify different types of features when processing multidimensional data with complex time dependencies, thereby improving prediction accuracy and reliability.

[0042] S50, perform cross-block information fusion on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; In this embodiment, the operation of cross-block information fusion of the decoupled periodic factor features and trend slope features to obtain fused decoupled features aims to further improve the expressive power of features and enhance the accuracy of data analysis by integrating information from different context blocks. The decoupling process separates periodic changes and trend changes in the data, but in order to obtain a comprehensive feature representation, these decoupled features must be fused across blocks to capture global information and the correlation between local features.

[0043] The core of cross-block information fusion lies in how to effectively fuse local features from different time periods to simultaneously preserve both the long-term dependencies and local details of time-series data. First, cross-block fusion requires the effective connection and weighting of periodic factor features and trend slope features extracted within each context block. The key here is ensuring that these features form meaningful correlations over time without losing the details of the time-series data. In particular, when correlations exist across multiple time periods, cross-block information fusion ensures that features from different time periods can influence each other, thereby strengthening the model's ability to capture long-term dependencies.

[0044] In practice, different mechanisms can be used to integrate cross-block information, such as bidirectional recurrent neural network (Bi-RNN) structures or multi-head attention mechanisms, to capture the temporal dependencies between different context blocks. In this fusion process, Bi-RNN, through its forward and backward propagation capabilities, enables bidirectional information transfer in the temporal dimension, thus avoiding the limitations of traditional models that rely solely on unidirectional temporal information. Meanwhile, multi-head attention mechanisms further enhance the ability to model complex data by weighting and aligning features from different levels and time periods.

[0045] Cross-block fusion can also employ other techniques, such as using gating mechanisms to adjust the fusion weights of features across different time periods. Through gating mechanisms, the model can dynamically adjust the feature weights of each block, preventing some blocks from having too much or too little influence, and ensuring that feature fusion is more flexible and accurate.

[0046] The ultimate goal of cross-block information fusion is to generate fused decoupled features. These features can not only effectively represent local periodic fluctuations and trend changes, but also comprehensively consider global patterns across different time periods, thus providing more accurate and comprehensive input for subsequent prediction or analysis tasks. Compared to single periodic or trend features, fused decoupled features can better reflect the multidimensionality and complexity of data, especially when dealing with high-dimensional, multivariate, and noisy industrial data, significantly improving the accuracy of analysis and prediction.

[0047] This embodiment improves the model's performance when processing complex time-series data by effectively fusing information from different context blocks, reducing feature omissions and information loss, thereby enhancing the overall model's prediction accuracy and robustness. Especially in fields such as industrial production, equipment failure prediction, or health diagnosis, this cross-block fusion method can more accurately identify potential anomalies or trend changes, providing more timely and effective decision support.

[0048] S60, the fused decoupled features are refined through trend attention pathway and periodic attention pathway respectively to obtain refined trend features and refined periodic features; In this embodiment, the process of refining the fused decoupled features first involves processing trend and periodic features using different attention paths. The purpose of this process is to further extract and optimize features from multiple dimensions, making each type of feature more representative and able to more accurately reflect key patterns in the data, especially in practical applications such as industrial production and equipment monitoring. Through this refinement process, long-term trend changes and periodic fluctuations in the data can be captured more effectively.

[0049] Specifically, the trend-based attention pathway aims to extract long-term trend changes from data, particularly those features with a long time span and significant impact on overall behavioral patterns. To achieve this, the trend pathway first models the decoupled features using time decay, employing techniques such as time decay coefficients to ensure the model focuses on recent trend changes while avoiding excessive interference from older data. Through this method, trend features can accurately reflect long-term fluctuation patterns in the data, making it particularly suitable for detecting long-term dependent features such as equipment operating trends and seasonal changes in production processes.

[0050] The periodic attention pathway focuses on capturing short-term periodic fluctuations in data, especially those features that are highly periodic and regular. In practical applications, these features often represent repetitive events in a system, such as periodic load fluctuations in equipment or periodic operational processes in a production line. The periodic pathway uses methods such as the phase difference loss function to perform periodic phase alignment calibration on the features, enabling the model to effectively capture specific periodic change patterns. This process ensures the accuracy of periodic features, avoids interference from other irrelevant fluctuations, and improves the overall recognition capability of the model.

[0051] In both pathways, the attention mechanism plays a central role. By assigning dynamic learning weights to each pathway, the model can automatically adjust its focus on different features based on the specific characteristics of the input data, thereby optimizing information extraction. The key to the attention pathway lies in its ability to dynamically adjust the model's focus based on the different characteristics of the data, ensuring that the extraction of trend and periodic features is highly relevant and accurate.

[0052] Through these two independent refining processes, refined trend characteristics and refined periodic characteristics are obtained respectively. These characteristics can better reflect the long-term and short-term dynamics of the data and provide more detailed data information, providing more accurate and richer input for subsequent analysis and prediction tasks.

[0053] This embodiment, by independently processing trend and periodic features, can eliminate the interference between them and highlight their respective characteristics, thereby improving the model's performance when processing complex time-series data. Especially when facing highly dynamic and variable industrial data, it can effectively improve prediction accuracy, reduce noise interference, and ultimately improve the efficiency and reliability of tasks such as fault detection and production scheduling.

[0054] S70, perform multi-channel hybrid reconstruction of the refining trend characteristics and the refining periodic characteristics to obtain the analysis results.

[0055] In this embodiment, the process of multi-channel hybrid reconstruction of refined trend and periodic features aims to combine the refined features to further improve the model's expressive power and obtain more accurate analysis results. This process involves multi-channel data fusion technology, enabling the model to comprehensively process data from different dimensions, thereby better capturing the potential patterns in the data. Especially in multi-scale, multi-modal time series data, it can provide higher-precision analysis and prediction.

[0056] In this step, refined trend features and refined periodic features are first used as inputs and processed through a multi-channel fusion layer. Each input channel's features are assigned different weights in this layer, and the two types of features are synthesized through weighted summation and fusion. The core of this fusion process is to utilize the multi-channel fusion layer to integrate features from different channels, thereby maximizing the extraction of information contained in each channel while avoiding information loss.

[0057] By employing a multi-head adaptive weighted fusion layer, refined features are weighted and fused, allowing each channel to be assigned a learnable, dynamic weight. This process enables the model to automatically adjust the contribution of different channel features to the final blended features based on their importance. Specifically, during training, the model learns the impact of each channel on the final task and dynamically adjusts the weight of each channel according to different input data, thereby improving the accuracy of the final result.

[0058] After fusion, the feature vectors are further processed through fully connected regression or classification layers to map the multi-channel fused features to the desired predicted values ​​or classification labels. During this process, non-linear activation functions (such as ReLU, Sigmoid, or Softmax) map the features to a high-dimensional space, allowing the data to better fit the model's output requirements.

[0059] This embodiment, through multi-channel hybrid reconstruction, fully leverages the advantages of different feature dimensions, integrating refined trend and periodic information to provide more accurate analysis results. It is suitable for processing complex time-series data, especially in scenarios requiring the simultaneous capture of long-term trends and short-term periodic changes, significantly improving the accuracy of analysis and prediction. Furthermore, utilizing learnable dynamic weights to adjust the contributions of different channel features helps eliminate noise influences, improving model stability and robustness. In practical applications, this process can effectively improve production efficiency, the accuracy of equipment monitoring, and decision support capabilities under different environments.

[0060] This invention relates to the field of data analysis technology, and discloses a data analysis method, apparatus, device, and medium based on feature decoupling and fusion. The method includes: acquiring input data from multiple data channels and preprocessing it; extracting multi-phase fusion feature vectors; constructing a multi-link adaptive attention matrix; performing contextual block processing and decoupling through dual-path convolution operations; performing cross-block information fusion; refining trend features and periodic features; and finally generating analysis results through multi-channel hybrid reconstruction. This invention, through a multi-scale feature decoupling and fusion method, can accurately capture long-term dependencies and local mutation features in time-series data, overcoming the shortcomings of existing technologies in handling complex multivariate data, concurrent time scales, and noise interference. This significantly improves prediction accuracy and generalization ability, meeting the needs of high-precision prediction and real-time fault diagnosis.

[0061] In one embodiment, step S20 above includes: S201, Perform continuous wavelet transform on the preprocessed data, and generate time spectrum diagrams for each data channel by adjusting the scaling and translation parameters; S202, Select a mother wavelet function with time-frequency localization characteristics, perform multi-scale decomposition on the preprocessed data, and generate a phase map covering different frequency ranges; S203, perform beamforming convolution on the phase map, and enhance the sensitivity of periodic texture features and trend energy distribution features in the phase map through directional filtering; S204 performs channel-level cascade fusion of phase maps corresponding to different scale parameters to generate multi-phase fusion feature vectors for each data channel.

[0062] In this embodiment, the purpose of performing phase-domain joint transform processing on the preprocessed data is to extract multi-phase fusion feature vectors for each data channel. The core of this process is to improve the expressive power of the features through multi-scale wavelet transform, beamforming convolution, and phase map fusion, ensuring that both periodic textures and trend changes in the data are captured simultaneously. Throughout the process, multiple steps work together to ensure that information from different time scales and frequency ranges is fully extracted and integrated, thereby providing more valuable feature inputs for subsequent data analysis, prediction, and decision-making.

[0063] First, the preprocessed data is processed using continuous wavelet transform, a time-frequency analysis tool that generates a time-spectrum graph by changing the scaling and translation parameters. This transform converts the time-domain signal into a time-frequency domain signal, thus reflecting the signal's changes in both time and frequency. In the wavelet transform, the scaling parameter determines the frequency range of the analysis, while the translation parameter controls the position of the wavelet basis functions on the time axis. By adjusting these parameters, different frequency components of the signal at different time points can be captured, which is particularly important for processing complex time-series data.

[0064] Next, the preprocessed data is decomposed at multiple scales by selecting a mother wavelet function with time-frequency localization characteristics. The choice of the mother wavelet function is crucial for the time-frequency localization of the data, enabling accurate decomposition of the signal at different time and frequency scales. The generated phase map covers different frequency ranges, with the low-frequency portion mainly corresponding to the trend changes of the signal, while the high-frequency portion mainly reflects the periodic fluctuations of the signal. In this way, different levels and details of the signal can be extracted, thus providing multi-dimensional information for subsequent feature analysis.

[0065] Next, beamforming convolution is performed on the generated phase map. This step enhances the sensitivity to periodic texture features and trend energy distribution features in the phase map by introducing directional filtering. Beamforming convolution allows the model to convolve the signal in different directions, thereby more accurately capturing periodic patterns and trend changes in the data. The introduction of directional filtering further strengthens the sensitivity to signal patterns in specific directions, making it particularly suitable for extracting periodic and trend signals.

[0066] Finally, the phase maps generated at different scales are channel-cascaded and fused, fusing the different frequency components of each data channel to generate a multi-phase fused feature vector. This fusion process not only preserves the feature information of different frequency ranges but also enhances the model's comprehensive understanding of time-series data by merging data from multiple channels. Each channel corresponds to a different feature level, and the fused feature vector has stronger expressive power, providing richer and more accurate input in subsequent models.

[0067] This embodiment, through the above process, can effectively extract multi-dimensional, multi-frequency range information from complex time-series data and fuse it into a higher-level feature vector. It can simultaneously capture the periodic fluctuations and trend changes of the signal, thus providing a comprehensive feature representation and improving the accuracy and robustness of subsequent analysis and prediction. Furthermore, the multi-scale wavelet transform and beamforming convolution operations employed give the model strong adaptability when processing multi-level, multi-mode signals, especially when dealing with complex industrial data, effectively handling noise interference and signal instability.

[0068] In one embodiment, step S30 above includes: S301, Configure multiple parallel attention links, each attention link independently learns the interaction relationship between different phase domain features in the multi-phase fusion feature vector, and generates initialized attention link parameters; S302, in each attention link, a global trend perception channel and a local time-varying perception channel are divided, and the feature interaction model of the multi-phase fusion feature vector is performed by a large receptive field attention head and a small receptive field attention head respectively to perform long-term dependence and short-term fluctuation feature interaction modeling. S303, perform multi-head attention weight calculation on the output of each attention link to generate the attention matrix corresponding to each link; S304, the attention matrices corresponding to each link are weighted and superimposed using learnable adaptive weights to generate the fused attention matrix.

[0069] In this embodiment, the core objective in constructing the multi-link adaptive attention matrix is ​​to generate a fused attention matrix by learning the interaction relationships between features in different phase domains of the signal through multiple parallel attention links based on multi-phase fusion feature vectors. In this way, the model can better capture long-term dependencies and short-term fluctuations in the signal, while adapting to different feature scales. The specific implementation involves multiple steps, including parallel link configuration, channel partitioning, long-term and short-term dependency modeling, and attention weight calculation.

[0070] First, multiple parallel attention links need to be configured, each responsible for independently learning the interaction relationships between features in different phase domains within the multi-phase fusion feature vector. This means that each link can focus on a specific feature dimension or time scale, thereby effectively capturing different patterns in the data. Each link generates initial attention link parameters, which are gradually adjusted through learning and optimization, ultimately helping the link to extract key features of the signal more accurately. The initialization process typically involves random initialization and continuous parameter adjustment using training data, enabling each link to independently process and learn specific feature relationships.

[0071] Next, within each attention path, a global trend-aware channel and a local time-varying awareness channel need to be defined. This division helps the model handle long-term dependent and short-term fluctuating features separately. Specifically, the global trend-aware channel primarily handles long-term dependent features in the signal, such as trend changes and seasonal variations; while the local time-varying awareness channel focuses on capturing short-term fluctuations in the signal, such as sudden events or periodic fluctuations. To achieve this, a large receptive field attention head and a small receptive field attention head are used in each channel, corresponding to the feature interaction modeling of long-term dependent and short-term fluctuating features, respectively. The large receptive field attention head can capture features over a longer time range, while the small receptive field attention head can capture short-term local fluctuations, which helps enhance the model's multi-scale feature capture capability.

[0072] Then, multi-head attention weights are calculated on the output of each attention link. The advantage of the multi-head attention mechanism is that it can weight features from different perspectives in parallel by multiple attention heads, thereby improving the model's expressive power. In each link, the attention weights can be calculated based on the dot product of the input feature vector and the weight matrix, and then normalized through a softmax operation to obtain the attention weights at each position. In this way, the model can effectively focus on important feature regions and suppress the influence of irrelevant features.

[0073] Finally, the attention matrices generated by each link are weighted and superimposed using learnable adaptive weights. This weighting process involves assigning different weights to the attention matrix output by each link; these weights are automatically adjusted through the learning process so that the model can adaptively allocate attention values ​​to different links. The goal of this step is to generate the final fused attention matrix, which contains weighted information from each link across different feature dimensions and effectively captures multi-level dependencies in the signal.

[0074] This embodiment constructs a multi-link adaptive attention matrix, enabling multi-dimensional and comprehensive modeling of multi-phase fused feature vectors. Each parallel attention link independently learns the interaction relationships of features in different phase domains, effectively capturing long-term dependencies and short-term fluctuations in the data. By dividing the attention heads into large and small receptive fields, features with long-term dependencies and short-term fluctuations can be processed simultaneously, enhancing the model's ability to handle multi-scale signals. The resulting fused attention matrix effectively improves the accuracy of subsequent analysis and prediction, exhibiting stronger robustness and adaptability, especially when dealing with complex time-series data.

[0075] In one embodiment, step S40 above includes: S401, according to the preset time window length, the multi-phase fusion feature vector and the fusion attention matrix are divided into continuous and partially overlapping context blocks according to the same time window; S402, within each context block, a convolution kernel of a first preset size is configured through the first path, and the multi-phase fusion feature vector is multiplied with the corresponding fusion attention matrix block to extract high-frequency periodic texture features and generate periodic factor features. S403, within each context block, a convolution kernel of a second preset size is configured through the second path to perform a dot product operation between the multi-phase fusion feature vector and the corresponding fusion attention matrix block, extracting low-frequency trend slope changes and generating trend slope features; S404, apply orthogonal constraints to the periodic factor features generated by the first path and the trend slope features generated by the second path to suppress feature redundancy information between the first path and the second path; S405 uses max pooling to reduce the dimensionality of orthogonally constrained features, generating decoupled periodic factor features and trend slope features.

[0076] In this embodiment, the multi-phase fusion feature vector and the fusion attention matrix are first segmented into multiple context blocks. The purpose of this operation is to segment the data along the time dimension while maintaining data continuity and partial overlap. This provides a local time window for subsequent convolutional operations, allowing features within a specific time range to be learned and extracted independently. The specific segmentation operation is based on a preset time window length, ensuring that the multi-phase fusion feature vector and the fusion attention matrix are segmented with the same window size in each context block, thereby guaranteeing feature alignment. In practice, the window length can be adjusted according to the application scenario and data characteristics. Typically, the chosen window length considers the temporal dependence of the data to ensure that local change patterns are fully captured.

[0077] Next, within each context block, feature decoupling is achieved through a dual-path convolution operation. The first path uses a convolution kernel of a first preset size to perform a dot product operation on the multi-phase fused feature vector and the corresponding fused attention matrix block. The dot product operation weights the multi-phase fused feature vector through the fused attention matrix, thereby highlighting periodic features. In this way, high-frequency periodic texture features in the signal can be extracted, generating periodic factor features. These periodic factor features can reflect the regularity and pattern of periodic fluctuations in the data, which is crucial for capturing long-term or short-term periodic behavior.

[0078] Within the same context block, the second path uses a convolutional kernel of a second preset size, similarly performing a dot product operation on the multi-phase fusion feature vector and the corresponding fusion attention matrix block. Unlike the first path, the second path has a larger convolutional kernel size, enabling it to capture low-frequency features in the signal, especially trend changes and slope features in the data. The features extracted in this process are trend slope features, which help the model identify changes in long-term trends in the data, providing effective feature support, especially in the case of sudden changes or abnormal fluctuations.

[0079] After the dual-path convolution operation, orthogonal constraints are applied to the periodic factor feature and the trend slope feature. The purpose of the orthogonal constraints is to reduce redundant information between these two features by restricting their independence. In this way, it is ensured that each feature can independently reflect the periodic and trend changes in the data, thereby improving the accuracy and reliability of the model in subsequent stages.

[0080] Finally, max pooling is used to reduce the dimensionality of the orthogonally constrained features. Pooling helps reduce the dimensionality of features while retaining the most important information. Max pooling removes unimportant redundant information and retains the features most useful for classification or regression tasks, thus providing a more compact and efficient feature representation for subsequent analysis and decision-making.

[0081] This embodiment effectively extracts periodic factors and trend slope features from signals by contextually segmenting the multi-phase fused feature vector and fused attention matrix and decoupling them using dual-path convolution. Orthogonal constraints further suppress the interference of redundant information, allowing each feature to independently reflect different patterns in the data. Max pooling ensures effective feature compression, enabling the model to process high-dimensional features at a lower computational cost, ultimately improving the model's ability to process complex time-series data.

[0082] In one embodiment, step S50 above includes: S501, establish a bidirectional recurrent neural network connection between adjacent context blocks to perform time series context transfer on the decoupled periodic factor features and trend slope features. S502 performs multi-head attention weight alignment on the periodic factor features and trend slope features transmitted across blocks to generate periodic factor features and trend slope features after cross-block attention alignment. S503, the fusion weights of the periodic factor features and trend slope features after cross-block attention alignment are adjusted through a gating mechanism to generate initial fusion features; S504, perform layer normalization on the initial fused features to eliminate scale differences between features and generate the fused decoupled features.

[0083] In this embodiment, firstly, a bidirectional recurrent neural network (Bi-RNN) connection is established between adjacent context blocks. Bidirectional recurrent neural networks are a commonly used neural network architecture capable of capturing the contextual dependencies of sequential data. In this operation, Bi-RNN is used to process the decoupled periodic factor features and trend slope features, and to model the data through the contextual propagation of the time series. The introduction of the bidirectional structure allows the network to not only utilize forward sequence information but also to obtain contextual information from subsequent data, thereby providing a more comprehensive feature representation. Especially when the data exhibits nonlinear dynamic changes, it enhances the model's ability to understand the global structure of the data.

[0084] Next, for the periodic factor features and trend slope features transmitted across blocks, the alignment relationship between cross-block features is further optimized through multi-head attention weight alignment. The multi-head attention mechanism can enhance the interrelationship between features by focusing on different aspects of the features through multiple independent attention heads. In this way, the model can effectively focus on important information in different time windows and reduce the risk of losing local information. In the specific implementation, each attention head can independently learn and focus on different subspace features of the input data, thereby improving the model's sensitivity to complex data.

[0085] Furthermore, a gating mechanism is used to adjust the fusion weights of periodic factor features and trend slope features after cross-block attention alignment. Gating mechanisms are commonly used in deep learning models to dynamically adjust the weights of information flow. In this step, the gating mechanism adjusts the weights of different features based on contextual information, ensuring that important information is amplified while less important parts are suppressed. This mechanism helps the model rationally allocate the importance of features across different time periods, further optimizing the feature fusion effect.

[0086] Finally, the adjusted initial fused features are subjected to layer normalization. Layer normalization is a common technique in deep learning, primarily used to reduce instability during training. By standardizing the output of each layer, layer normalization eliminates scale differences between features, enabling the model to learn more stably during training. The core function of this step is to unify and standardize features at different scales, thereby improving the model's training efficiency and promoting effective feature fusion.

[0087] Through these steps, the resulting fused decoupled features have higher accuracy and stability, and can better capture complex patterns and long-term dependencies in time series. Especially when facing complex time series data with multiple dimensions and scales, they can effectively improve the accuracy of prediction and analysis.

[0088] This embodiment utilizes bidirectional recurrent neural networks, multi-head attention weight alignment, gating mechanism adjustment, and layer normalization to efficiently fuse and optimize periodic factor features and trend slope features, thereby improving the accuracy of feature extraction and the robustness of the model. It can accurately capture long-term dependencies and local mutation features, demonstrating strong adaptability and reliability in time series data analysis.

[0089] In one embodiment, step S60 above includes: S601, In the trend attention path, a time decay coefficient is introduced to perform long-term trend dependency modeling on the fused decoupled features to generate initial refined trend features. S602, In the periodic attention path, the fused decoupled features are periodically phase aligned and calibrated based on the phase difference loss function to generate initial refined periodic features; S603, the attention weights of the initial refining trend feature and the initial refining periodicity feature are adjusted by temperature coefficient to suppress abnormal weight interference and generate the adjusted initial refining trend feature and initial refining periodicity feature. S604, the fused decoupling features are superimposed with the adjusted initial refining trend features and initial refining periodicity features through residual connection to generate refining trend features and refining periodicity features.

[0090] In this embodiment, a time decay coefficient is first introduced into the trend attention path to model the long-term trend dependencies of the fused, decoupled features. The time decay coefficient is a common technique in deep learning used to simulate long-term trend changes by adjusting the temporal importance of input features. Specifically, over time, the model's focus on past information gradually decreases, allowing it to concentrate more on recent data. This design effectively captures long-term dependencies in the data, especially useful when analyzing long-term trends (such as market cyclical fluctuations and climate change). The use of the time decay coefficient ensures effective weighting of features over time, enhancing the model's sensitivity to trend changes.

[0091] Next, in the periodic attention path, periodic phase alignment calibration is performed on the fused decoupled features based on the phase difference loss function. Periodic phase alignment is a method for normalizing periodic data, aiming to eliminate signal offsets or misalignments in time series. The phase difference loss function ensures signal consistency across multiple periods by measuring and minimizing the phase difference between periodic features, thereby enhancing the model's adaptability to periodic fluctuations. Through this process, the model can better capture periodic changes, especially when dealing with data with complex periodic features (such as weather changes, periodic equipment failures, etc.), significantly improving the extraction accuracy of periodic features.

[0092] Based on this, the attention weights for the initial refined trend features and the initial refined periodic features are adjusted using temperature coefficients. Temperature coefficient adjustment is a common technique used to control the balance of information distribution in the attention mechanism. By adjusting the temperature coefficients, attention weights can be finely controlled, suppressing the interference of abnormal weights on model training, ensuring that important features are strengthened while unimportant features are weakened. This operation can enhance the model's focus on effective information and reduce the impact of noise or irrelevant information, thereby improving the model's robustness.

[0093] Finally, the fused decoupled features are superimposed with the adjusted initial refined trend features and initial refined periodic features via residual connections. Residual connections are a technique in deep learning used to solve the vanishing gradient problem and accelerate training. By preserving the original features through skip connections, the network can directly utilize unprocessed features during training. This design effectively retains important information from the input features and fuses them with the refined features to generate the final refined trend and periodic features. This fusion not only ensures feature richness but also improves model stability and output accuracy.

[0094] Through these steps, the refined trend and periodic features generated can more accurately capture long-term trend changes and periodic fluctuations in time series data, improving the model's performance on complex time series data. Whether in production process monitoring, equipment fault diagnosis, financial market forecasting, or meteorological data analysis, it can provide more reliable analytical results.

[0095] This embodiment combines various techniques, such as time decay coefficient, phase difference loss function, temperature coefficient adjustment, and residual connection, to effectively improve the refinement of trend and periodic features, ensuring the independence and robustness of features, thereby providing more accurate prediction and decision support in highly complex time series data analysis tasks.

[0096] In one embodiment, step S70 above includes: S701, In the multi-head adaptive weighted fusion layer, learnable dynamic weights are assigned to the refined trend feature and the refined periodic feature respectively, generating a weighted fusion hybrid feature vector; S702, the mixed feature vector is input into a fully connected regression or classification layer, and the mixed feature vector is mapped to a predicted value or a multi-class probability distribution through a non-linear activation function; S703, a confidence quantification module is added to the output layer to model the uncertainty of the predicted value or multi-class probability distribution and generate analysis results containing confidence intervals.

[0097] In this embodiment, firstly, a multi-head adaptive weighted fusion layer is used to assign learnable dynamic weights to refined trend features and refined periodic features. This step, through adaptive adjustment during training, allows different feature channels to dynamically adjust their influence weights on the final result according to the actual task requirements. The multi-head weighting mechanism means multiple parallel weight allocation paths, each focusing on a subset of different features, optimizing the fusion effect through its own weighting operation. The dynamic adjustment of weights is based on training data and continuously adjusted by optimizing the loss function to maximize prediction accuracy. This approach allows the model to re-evaluate the influence of features at each time point based on the different importance of the input data, thereby ensuring that the model has higher adaptability and accuracy when facing complex multi-dimensional data.

[0098] Next, the weighted and fused hybrid feature vector is input into a fully connected regression or classification layer. This process uses a non-linear activation function (such as ReLU, Sigmoid, or Softmax) to map the hybrid feature vector to the final output. The fully connected layer processes the feature vector, utilizing the complex non-linear transformations of the neural network to map the input features to the target output space (e.g., for regression tasks, outputting a continuous predicted value; for classification tasks, outputting the probability distribution of different classes). Non-linear activation functions are crucial tools for ensuring the network can capture the non-linear relationships between input features, enabling the network to handle more complex patterns and avoid outputting overly simplistic linear relationships, thus providing more flexible and accurate predictive capabilities.

[0099] At the output layer, a confidence quantification module is added to model the uncertainty of predicted values ​​or multi-class probability distributions. The role of this module is to quantify the uncertainty of the output results through the model, generating analysis results that include confidence intervals. This module typically uses techniques such as Bayesian methods or Monte Carlo sampling to estimate the confidence range of the predicted results based on variations in the training data and the model's uncertainty. This allows the model to provide not only the predicted results themselves but also the range of reliability of the results, thereby helping decision-makers take different action strategies based on different confidence levels. For example, in financial forecasting, a high confidence level in the model's prediction may encourage a more aggressive investment strategy; while a low confidence level may suggest a more conservative decision.

[0100] Through this series of processing steps, the final analysis results not only include the core output of the prediction but also have a clear confidence interval, providing more background information and operability for decision-making. This processing method enables the system to operate more robustly in complex environments, providing more reliable and trustworthy predictions.

[0101] Example Description: In modern manufacturing, the precision machining operations of CNC machine tools not only have a crucial impact on product quality but also require efficient and reliable equipment monitoring systems to ensure production continuity and equipment safety. This system addresses equipment monitoring and fault prediction in CNC machine tool operations, enabling real-time monitoring, accurate prediction, and anomaly detection of the machine tool's status. The core hardware of this system includes a high-end CNC machine tool equipped with multiple different types of sensors capable of real-time monitoring of various key machine tool data, such as spindle temperature, servo motor current of each axis, machine tool pressure, and acoustic signals during operation. These sensors include PT100 temperature sensors, Hall effect current sensors, piezoelectric pressure sensors, and acoustic sensor arrays. Data from all sensors is synchronously acquired at a frequency of 1kHz through a data acquisition system and transmitted to edge computing nodes via industrial Ethernet. The edge computing nodes are configured with high-performance multi-core processors (such as Intel Core i7 or equivalent processors) and at least 16GB of memory, supporting powerful data analysis capabilities. The data analysis platform integrates an algorithm library for multi-scale feature decoupling and fusion, enabling real-time data processing and analysis.

[0102] For the raw sensor data collected, a sliding window Hampel filter is first used to identify and correct abnormal isolated points or sudden spikes in the data, in order to remove instantaneous spurious signals caused by electromagnetic interference or sensor malfunction. The corrected data is then standardized using a min-max normalization method, mapping the data values ​​to the interval [0,1] for subsequent processing. Subsequently, a continuous wavelet transform is performed on the data using the Morlet wavelet function to obtain the time-spectrum, and then the multi-phase fusion features of each data channel are extracted. These features reflect the signal energy distribution at different frequencies, which helps to reveal the periodic and trend changes in machine tool operation.

[0103] Based on the data features after phase domain transformation, a multi-link adaptive attention matrix was constructed. This matrix models the interaction relationships between different signals through multiple parallel attention links, particularly enhancing the interaction analysis of specific frequency band features between current and acoustic signals, and modeling the low-frequency trend features between temperature and pressure signals. During training, the adaptive weights of the links are optimized through backpropagation to ensure dynamic adjustment of the weights for different signal interactions, thereby providing more effective feature information for fault prediction.

[0104] In the next stage, the temporal data is segmented into multiple context blocks and features are decoupled using dual-path convolution. Within each context block, two convolutional kernels of different sizes are used to extract high-frequency periodic texture features and low-frequency trend slope change features, respectively. Short-term high-frequency changes are captured by small convolutional kernels, while long-term trend changes are extracted by large convolutional kernels. This step effectively helps the model distinguish between rapidly changing vibrational patterns and slowly changing trends, while maintaining the global information transmission of the decoupled features and avoiding information loss due to over-segmentation.

[0105] Subsequently, information between context blocks is fused using a cross-block information fusion mechanism (such as a Transformer encoder) to ensure the consistency of global features. In the extraction stage of seasonal and trend features, a dual-path adaptive attention mechanism is employed to refine the decoupled features. The seasonal attention path focuses on the impact of daily operating patterns and shift changes on the machine tool, while the trend attention path focuses on long-term wear and tear and usage. Through time decay coefficients and phase difference loss functions, long-term dependency modeling and periodic phase alignment calibration are performed on these features to ensure a balance between long-term and short-term variations in the refined features.

[0106] In the final stage, a multi-channel hybrid reconstruction network is used to fuse feature information from seasonal and trend-based attention pathways. This network weights and fuses these features through an adaptive weighting layer, and then generates a predicted tool remaining service life (RUL) through a fully connected regression layer. The system's RUL prediction results demonstrate excellent accuracy and meet the requirements for real-time monitoring in practical operation. Experimental results show that the system can accurately predict the remaining service life of equipment and provide early warnings before potential equipment failures occur.

[0107] This embodiment combines a multi-head adaptive weighted fusion layer, a fully connected regression or classification layer, and a confidence quantification module. This approach not only provides accurate predictions when dealing with multi-dimensional and complex time-series data but also quantifies the uncertainty of the predictions, enhancing the model's practicality and reliability. This method enables the prediction system to more accurately capture data features and provide more reliable decision support when handling complex dynamic changes and diverse tasks.

[0108] In one embodiment, a data analysis apparatus based on feature decoupling and fusion is provided, which corresponds one-to-one with the data analysis method based on feature decoupling and fusion described in the above embodiments. (Refer to...) Figure 3 , Figure 3This is a schematic diagram of the functional modules of a preferred embodiment of the data analysis device based on feature decoupling and fusion according to the present invention. The modules include a data acquisition and preprocessing module 10, a feature extraction module 20, an attention mechanism module 30, a feature decoupling module 40, an information fusion module 50, a feature refinement module 60, and a result reconstruction module 70. Detailed descriptions of each functional module are as follows: The data acquisition and preprocessing module 10 is used to acquire input data from multiple data channels and perform preprocessing operations on the input data to generate preprocessed data. Feature extraction module 20 is used to perform phase domain joint transformation processing on the preprocessed data to extract multi-phase fusion feature vectors of each data channel; Attention mechanism module 30 is used to construct a multi-link adaptive attention matrix based on the multi-phase fusion feature vector to obtain a fusion attention matrix; The feature decoupling module 40 is used to perform context block processing on the multi-phase fusion feature vector and the fusion attention matrix, and to decouple them in each context block through dual-path convolution operation to generate decoupled periodic factor features and trend slope features. The information fusion module 50 is used to perform cross-block information fusion on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The feature refinement module 60 is used to refine the fused decoupled features through a trend attention path and a periodic attention path, respectively, to obtain refined trend features and refined periodic features. The result reconstruction module 70 is used to perform multi-channel hybrid reconstruction of the refining trend characteristics and the refining periodic characteristics to obtain analysis results.

[0109] In one embodiment, the feature extraction module 20 is specifically used for: The preprocessed data is subjected to continuous wavelet transform, and the time spectrum of each data channel is generated by adjusting the scaling and translation parameters. Select a mother wavelet function with time-frequency localization characteristics to perform multi-scale decomposition on the preprocessed data and generate phase maps covering different frequency ranges; Beamforming convolution is performed on the phase map, and directional filtering is used to enhance the sensitivity of periodic texture features and trend energy distribution features in the phase map. Phase maps corresponding to different scale parameters are cascaded and fused to generate multi-phase fusion feature vectors for each data channel.

[0110] In one embodiment, the attention mechanism module 30 is specifically used for: Configure multiple parallel attention links, each attention link independently learns the interaction relationship between features in different phase domains in the multi-phase fusion feature vector, and generates initialized attention link parameters; In each attention link, a global trend perception channel and a local time-varying perception channel are divided, and the feature interaction model of the multi-phase fusion feature vector is performed by using a large receptive field attention head and a small receptive field attention head respectively to model the long-term dependence and short-term fluctuation of the feature. Multi-head attention weights are calculated for the output of each attention link to generate the attention matrix corresponding to each link; The attention matrix is ​​generated by weighting and superimposing the attention matrices corresponding to each link using learnable adaptive weights.

[0111] In one embodiment, the feature decoupling module 40 is specifically used for: Based on a preset time window length, the multi-phase fusion feature vector and the fusion attention matrix are divided into continuous and partially overlapping context blocks according to the same time window. Within each context block, a convolution kernel of a first preset size is configured through the first path, and the multi-phase fusion feature vector is multiplied with the corresponding fusion attention matrix block to extract high-frequency periodic texture features and generate periodic factor features. Within each context block, a convolution kernel of a second preset size is configured through a second path to perform a dot product operation between the multi-phase fusion feature vector and the corresponding fusion attention matrix block, thereby extracting low-frequency trend slope changes and generating trend slope features. Orthogonal constraints are applied to the periodic factor features generated by the first path and the trend slope features generated by the second path to suppress feature redundancy between the first path and the second path; Max pooling is used to reduce the dimensionality of the orthogonally constrained features, generating decoupled periodic factor features and trend slope features.

[0112] In one embodiment, the information fusion module 50 is specifically used for: A bidirectional recurrent neural network connection is established between adjacent context blocks to perform time series context passing on the decoupled periodic factor features and trend slope features. Multi-head attention weight alignment is performed on the periodic factor features and trend slope features transmitted across blocks to generate periodic factor features and trend slope features after cross-block attention alignment; The initial fused features are generated by adjusting the fusion weights of the periodic factor features and trend slope features after cross-block attention alignment through a gating mechanism. The initial fused features are subjected to layer normalization to eliminate scale differences between features, thereby generating the fused decoupled features.

[0113] In one embodiment, the feature refinement module 60 is specifically used for: In the trend attention path, a time decay coefficient is introduced to model the long-term trend dependence of the fused decoupled features, generating initial refined trend features; In the periodic attention path, the fused decoupled features are periodically phase aligned and calibrated based on the phase difference loss function to generate initial refined periodic features; The attention weights of the initial refining trend feature and the initial refining periodicity feature are adjusted by temperature coefficients to suppress abnormal weight interference and generate adjusted initial refining trend features and initial refining periodicity features. The fused decoupling features are superimposed with the adjusted initial refining trend features and initial refining periodicity features through residual connection to generate refining trend features and refining periodicity features.

[0114] In one embodiment, the result reconstruction module 70 is specifically used for: In the multi-head adaptive weighted fusion layer, learnable dynamic weights are assigned to the refined trend features and the refined periodic features respectively, generating a weighted fused hybrid feature vector. The hybrid feature vector is input into a fully connected regression or classification layer, and a non-linear activation function is used to map the hybrid feature vector into a predicted value or a multi-class probability distribution. A confidence quantification module is added to the output layer to model the uncertainty of the predicted value or multi-class probability distribution and generate analysis results containing confidence intervals.

[0115] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements server-side functions or steps of a data analysis method based on feature decoupling and fusion.

[0116] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the user-side functions or steps of a data analysis method based on feature decoupling and fusion. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The system acquires input data from multiple data channels and performs preprocessing operations on the input data to generate preprocessed data. The preprocessed data is subjected to phase domain joint transformation processing to extract multi-phase fusion feature vectors of each data channel; Based on the multi-phase fusion feature vector, a multi-link adaptive attention matrix is ​​constructed to obtain the fusion attention matrix; The multi-phase fusion feature vector and the fusion attention matrix are processed by context block segmentation, and decoupled within each context block by dual-path convolution operation to generate decoupled periodic factor features and trend slope features. Cross-block information fusion is performed on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The fused decoupled features are refined through trend-based attention pathways and periodic attention pathways to obtain refined trend features and refined periodic features. The refining trend characteristics and the refining periodic characteristics are reconstructed using a multi-channel hybrid method to obtain the analysis results.

[0117] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The system acquires input data from multiple data channels and performs preprocessing operations on the input data to generate preprocessed data. The preprocessed data is subjected to phase domain joint transformation processing to extract multi-phase fusion feature vectors of each data channel; Based on the multi-phase fusion feature vector, a multi-link adaptive attention matrix is ​​constructed to obtain the fusion attention matrix; The multi-phase fusion feature vector and the fusion attention matrix are processed by context block segmentation, and decoupled within each context block by dual-path convolution operation to generate decoupled periodic factor features and trend slope features. Cross-block information fusion is performed on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The fused decoupled features are refined through trend-based attention pathways and periodic attention pathways to obtain refined trend features and refined periodic features. The refining trend characteristics and the refining periodic characteristics are reconstructed using a multi-channel hybrid method to obtain the analysis results.

[0118] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0121] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A data analysis method based on feature decoupling and fusion, characterized in that, Includes the following steps: The system acquires input data from multiple data channels and performs preprocessing operations on the input data to generate preprocessed data. The preprocessed data is subjected to phase domain joint transformation processing to extract multi-phase fusion feature vectors of each data channel; Based on the multi-phase fusion feature vector, a multi-link adaptive attention matrix is ​​constructed to obtain the fusion attention matrix; The multi-phase fusion feature vector and the fusion attention matrix are processed by context block segmentation, and decoupled within each context block by dual-path convolution operation to generate decoupled periodic factor features and trend slope features. Cross-block information fusion is performed on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The fused decoupled features are refined through trend-based attention pathways and periodic attention pathways to obtain refined trend features and refined periodic features. The refining trend characteristics and the refining periodic characteristics are reconstructed using a multi-channel hybrid method to obtain the analysis results.

2. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, The preprocessed data is subjected to phase domain joint transform processing to extract multi-phase fusion feature vectors for each data channel, including: The preprocessed data is subjected to continuous wavelet transform, and the time spectrum of each data channel is generated by adjusting the scaling and translation parameters. Select a mother wavelet function with time-frequency localization characteristics to perform multi-scale decomposition on the preprocessed data and generate phase maps covering different frequency ranges; Beamforming convolution is performed on the phase map, and directional filtering is used to enhance the sensitivity of periodic texture features and trend energy distribution features in the phase map. Phase maps corresponding to different scale parameters are cascaded and fused to generate multi-phase fusion feature vectors for each data channel.

3. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, Based on the multi-phase fusion feature vector, a multi-link adaptive attention matrix is ​​constructed to obtain the fusion attention matrix, including: Configure multiple parallel attention links, each attention link independently learns the interaction relationship between different phase domain features in the multi-phase fusion feature vector, and generates initialized attention link parameters; In each attention link, a global trend perception channel and a local time-varying perception channel are divided, and the feature interaction model of the multi-phase fusion feature vector is performed by using a large receptive field attention head and a small receptive field attention head respectively to model the long-term dependence and short-term fluctuation of the feature. Multi-head attention weights are calculated for the output of each attention link to generate the attention matrix corresponding to each link; The attention matrix is ​​generated by weighting and superimposing the attention matrices corresponding to each link using learnable adaptive weights.

4. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, The multi-phase fusion feature vector and the fusion attention matrix are subjected to contextual block processing, and decoupled within each context block by a dual-path convolution operation to generate decoupled periodic factor features and trend slope features, including: Based on a preset time window length, the multi-phase fusion feature vector and the fusion attention matrix are divided into continuous and partially overlapping context blocks according to the same time window. Within each context block, a convolution kernel of a first preset size is configured through the first path, and the multi-phase fusion feature vector is multiplied with the corresponding fusion attention matrix block to extract high-frequency periodic texture features and generate periodic factor features. Within each context block, a convolution kernel of a second preset size is configured through a second path to perform a dot product operation between the multi-phase fusion feature vector and the corresponding fusion attention matrix block, thereby extracting low-frequency trend slope changes and generating trend slope features. Orthogonal constraints are applied to the periodic factor features generated by the first path and the trend slope features generated by the second path to suppress feature redundancy between the first path and the second path; Max pooling is used to reduce the dimensionality of the orthogonally constrained features, generating decoupled periodic factor features and trend slope features.

5. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, Cross-block information fusion is performed on the decoupled periodic factor features and trend slope features to obtain fused decoupled features, including: A bidirectional recurrent neural network connection is established between adjacent context blocks to perform time series context passing on the decoupled periodic factor features and trend slope features. Multi-head attention weight alignment is performed on the periodic factor features and trend slope features transmitted across blocks to generate periodic factor features and trend slope features after cross-block attention alignment; The initial fused features are generated by adjusting the fusion weights of the periodic factor features and trend slope features after cross-block attention alignment through a gating mechanism. The initial fused features are subjected to layer normalization to eliminate scale differences between features, thereby generating the fused decoupled features.

6. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, The fused decoupled features are refined through trend-based attention pathways and periodic attention pathways to obtain refined trend features and refined periodic features, including: In the trend attention path, a time decay coefficient is introduced to model the long-term trend dependence of the fused decoupled features, generating initial refined trend features; In the periodic attention path, the fused decoupled features are periodically phase aligned and calibrated based on the phase difference loss function to generate initial refined periodic features; The attention weights of the initial refining trend feature and the initial refining periodicity feature are adjusted by temperature coefficients to suppress abnormal weight interference and generate adjusted initial refining trend features and initial refining periodicity features. The fused decoupling features are superimposed with the adjusted initial refining trend features and initial refining periodicity features through residual connection to generate refining trend features and refining periodicity features.

7. The data analysis method based on feature decoupling and fusion as described in claim 1, characterized in that, The refining trend characteristics and the refining periodic characteristics are reconstructed using a multi-channel hybrid method to obtain the analysis results, including: In the multi-head adaptive weighted fusion layer, learnable dynamic weights are assigned to the refined trend features and the refined periodic features respectively, generating a weighted fused hybrid feature vector. The hybrid feature vector is input into a fully connected regression or classification layer, and a non-linear activation function is used to map the hybrid feature vector into a predicted value or a multi-class probability distribution. A confidence quantification module is added to the output layer to model the uncertainty of the predicted value or multi-class probability distribution and generate analysis results containing confidence intervals.

8. A data analysis device based on feature decoupling and fusion, characterized in that, The data analysis device based on feature decoupling and fusion includes: The data acquisition and preprocessing module is used to acquire input data from multiple data channels and perform preprocessing operations on the input data to generate preprocessed data. The feature extraction module is used to perform phase domain joint transformation processing on the preprocessed data to extract multi-phase fusion feature vectors of each data channel; The attention mechanism module is used to construct a multi-link adaptive attention matrix based on the multi-phase fusion feature vector to obtain the fusion attention matrix; The feature decoupling module is used to perform context block processing on the multi-phase fusion feature vector and the fusion attention matrix, and decouple them in each context block through dual-path convolution operation to generate decoupled periodic factor features and trend slope features. The information fusion module is used to perform cross-block information fusion on the decoupled periodic factor features and trend slope features to obtain the fused decoupled features; The feature refinement module is used to refine the fused decoupled features through a trend attention path and a periodic attention path, respectively, to obtain refined trend features and refined periodic features. The result reconstruction module is used to perform multi-channel hybrid reconstruction of the refining trend characteristics and the refining periodic characteristics to obtain analysis results.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a data analysis program based on feature decoupling and fusion stored in the memory and executable on the processor. When executed by the processor, the data analysis program based on feature decoupling and fusion implements the steps of the data analysis method based on feature decoupling and fusion as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a data analysis program based on feature decoupling and fusion, which, when executed by a processor, implements the steps of the data analysis method based on feature decoupling and fusion as described in any one of claims 1-7.

Images