Time-frequency bi-clustering steel corrosion prediction method, device, equipment and medium
By using a time-frequency dual clustering method to extract features and cluster steel corrosion monitoring data, the problems of low prediction accuracy, non-stationary time series offset, and insufficient high-frequency signal capture in existing technologies are solved, achieving higher accuracy and robust corrosion prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-13
- Publication Date
- 2026-03-31
AI Technical Summary
Existing methods for predicting time-series data of steel corrosion monitoring suffer from problems such as low prediction accuracy, insufficient handling of non-stationary time-series distribution offsets, complex and difficult-to-explore inter-channel dependencies, inability of single-channel modeling strategies to balance independence and correlation, and insufficient capture of high-frequency abrupt signals.
The time-frequency dual clustering method is adopted. By extracting features and clustering multi-channel time-domain and frequency-domain data, clustering is performed along the channel dimension and time dimension respectively to generate collaborative clusters and coupled clusters. Feature fusion is performed by combining self-attention mechanism and cross-segment self-attention mechanism to realize the collaborative utilization and prediction of multi-dimensional information.
It improves the accuracy and robustness of steel corrosion prediction, can more accurately capture the temporal and spectral evolution of corrosion processes, reduces noise interference, enhances the ability to collaboratively utilize multi-channel data, and meets the needs of precise early warning in engineering scenarios.
Smart Images
Figure CN121765276A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of steel corrosion prediction technology, and in particular to a time-frequency bi-clustering method, apparatus, equipment and medium for predicting steel corrosion. Background Technology
[0002] In the current field of time-series data prediction for steel corrosion monitoring, the mainstream techniques mainly rely on traditional time-series models, including random forests, long short-term memory artificial neural networks (LSTM), and improved Transformer-based models (such as PatchTST and iTransformer). These models have been widely used in relevant scenarios and have become the core technical support for corrosion time-series data prediction.
[0003] However, steel corrosion monitoring data exhibits significant non-stationary temporal distribution shifts. These shifts, coupled with the complex dependencies between multi-channel monitoring data (temperature, humidity, corrosion rate, etc.), result in low prediction accuracy for steel corrosion monitoring time-series data using traditional square models. Channel-independent analysis strategies completely ignore the interactive correlation information between different monitoring channels, making it difficult to uncover the coupling patterns behind the data. Channel-related analysis strategies are highly susceptible to environmental noise and data fluctuations, failing to effectively extract useful correlation features between channels and thus failing to meet the actual needs of accurate early warning and assessment of corrosion status in engineering scenarios. Summary of the Invention
[0004] Therefore, it is necessary to provide a time-frequency bi-clustering method, apparatus, equipment, and medium for predicting steel corrosion in order to address the aforementioned technical problems and solve the deficiency of low prediction accuracy of time-series data for steel corrosion monitoring in the existing technology, thereby achieving more accurate prediction of steel corrosion.
[0005] The following technical solution is adopted in this specification: This specification provides a time-frequency biclustering method for predicting steel corrosion, including: Multi-channel time-domain data of the target steel is acquired from multiple monitoring channels, and frequency-domain transformation is performed on the multi-channel time-domain data to obtain the corresponding multi-channel frequency-domain data of the target steel. Temporal features were extracted from multi-channel time-domain data to obtain multi-channel steel corrosion temporal features; frequency domain features were extracted from multi-channel frequency-domain data to obtain multi-channel steel corrosion frequency-domain features. Clustering of the temporal features of multi-channel steel corrosion along the channel dimension and the time dimension respectively yields temporal channel features including multiple cooperative clusters and temporal features including multiple temporal evolution clusters. Each cooperative cluster represents a cooperative response mode of multiple monitoring channels, and each temporal evolution cluster represents a temporal evolution mode of corrosion. The frequency domain features of multi-channel steel corrosion are clustered along both the channel and time dimensions to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters. Each coupling cluster represents a spectral coupling mode of multiple monitoring channels, and each spectral evolution cluster represents a spectral evolution mode of corrosion. The time-domain channel features, time-domain time features, frequency-domain channel features, and frequency-domain time features are fused to obtain fused features; the corrosion prediction results of the target steel are obtained based on the fused features.
[0006] Furthermore, the clustering of multi-channel steel corrosion time-domain features along both the channel and time dimensions, and the clustering of multi-channel steel corrosion frequency-domain features along both the channel and time dimensions, are both implemented through a clustering module in a pre-trained steel corrosion prediction model; the clustering module includes: The channel-dimensional clustering unit is used to cluster the time-domain features of multi-channel steel corrosion along the channel dimension using the K-means clustering algorithm, generating time-domain channel features including multiple co-clusters; it is also used to cluster the frequency-domain features of multi-channel steel corrosion along the channel dimension using the K-means clustering algorithm, generating frequency-domain channel features including multiple coupled clusters; wherein, the K-means clustering algorithm uses the Euclidean distance between channels as a similarity measure; The time-dimensional clustering unit is used to cluster the time-domain features of multi-channel steel corrosion along the time dimension using a hierarchical clustering algorithm, generating time-domain features including multiple time-series evolution clusters; it is also used to cluster the frequency-domain features of multi-channel steel corrosion along the time dimension using a hierarchical clustering algorithm, generating frequency-domain features including multiple spectral evolution clusters; wherein, the hierarchical clustering algorithm uses the Dynamic Time Warping (DTW) distance as a similarity metric.
[0007] Furthermore, the frequency domain feature extraction of the multi-channel frequency domain data is achieved through the FilterNet module in the pre-trained steel corrosion prediction model. The FilterNet module extracts spectral features from the multi-channel frequency domain data through the PaiFilter filter kernel, which learns general spectral modeling characteristics during pre-training, and the TexFilter filter kernel, which learns the spatiotemporal coupling characteristics of the corrosion spectrum, and outputs multi-channel steel corrosion frequency domain features. The multi-channel steel corrosion frequency domain features include N sets of single-channel steel corrosion frequency domain features corresponding to N monitoring channels, where N is a positive integer. Each set of single-channel steel corrosion frequency domain features maintains the same number of time windows as the multi-channel frequency domain data in the time window dimension, so as to preserve the time-varying spectral characteristics of the corrosion process.
[0008] Furthermore, the extraction of temporal features from the multi-channel time-domain data is achieved through the Transformer module in the pre-trained steel corrosion prediction model. The Transformer module captures the long-range dependencies of the multi-channel time-domain data in the time dimension through a self-attention mechanism, and outputs multi-channel steel corrosion time-domain features. The multi-channel steel corrosion time-domain features include N sets of single-channel steel corrosion time-domain features corresponding to N monitoring channels, where N is a positive integer, and each set of single-channel steel corrosion time-domain features maintains the same time step as the multi-channel time-domain data in the time dimension.
[0009] Furthermore, the fusion of time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features is achieved through a multi-modal feature fusion module in a pre-trained steel corrosion prediction model; the multi-modal feature fusion module includes: The channel dimension alignment unit is used to align the time-domain channel features and frequency-domain channel features in dimensions, generating channel representation vectors with consistent dimensions. The time dimension alignment unit is used to align the time-domain time features and the frequency-domain time features in dimensions, and generate time-series representation vectors with consistent dimensions. The mask attention fusion unit is used to perform attention-weighted fusion of channel representation vectors and temporal representation vectors through the mask attention mechanism to obtain fused features that contain multi-dimensional representations of the erosion process.
[0010] Furthermore, the corrosion prediction result of the target steel obtained based on the fused features is achieved through the PatchTST module in the pre-trained steel corrosion prediction model. The PatchTST module divides the fused features into time-series segments and captures the nonlinear evolution law of corrosion contained in the time-series segments through a cross-segment self-attention mechanism to obtain the corrosion prediction result of the target steel. The cross-segment self-attention mechanism enhances the ability to identify the turning point of steel corrosion evolution through local-global feature interaction enhancement.
[0011] This specification provides a time-frequency biclustering device for predicting steel corrosion, comprising: The data acquisition module is used to acquire multi-channel time-domain data of the target steel from multiple monitoring channels, and to perform frequency domain transformation on the multi-channel time-domain data to obtain the corresponding multi-channel frequency-domain data of the target steel. The feature extraction module is used to extract time-series features from multi-channel time-domain data to obtain multi-channel steel corrosion time-domain features; and to extract frequency-domain features from multi-channel frequency-domain data to obtain multi-channel steel corrosion frequency-domain features. The temporal clustering module is used to cluster the temporal features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain temporal channel features including multiple cooperative clusters and temporal time features including multiple time-series evolution clusters; wherein, each cooperative cluster corresponds to a cooperative response mode of multiple monitoring channels, and each time-series evolution cluster corresponds to a time evolution mode of corrosion. The frequency domain clustering module is used to cluster the frequency domain features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters; wherein, each coupling cluster corresponds to a spectral coupling mode of multiple monitoring channels; and each spectral evolution cluster corresponds to a spectral evolution mode of corrosion. The prediction module is used to fuse time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features to obtain fused features; and to obtain corrosion prediction results for the target steel based on the fused features.
[0012] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described time-frequency biclustering method for predicting steel corrosion.
[0013] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described time-frequency biclustering method for predicting steel corrosion.
[0014] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This invention clusters time-domain and frequency-domain features along both the channel and time dimensions, accurately uncovering collaborative response patterns and spectral coupling patterns among multiple monitoring channels. It also clearly captures the temporal and spectral evolution patterns of the corrosion process, achieving feature complementarity and information gain across channels, between the time and frequency domains, and along the time dimension. For non-stationary temporal distribution shifts, time-dimensional clustering categorizes the temporal evolution features of different stages in the corrosion process into temporal evolution clusters and spectral evolution clusters. By extracting feature clusters with consistent evolutionary patterns, it filters out temporal distribution shift interference caused by non-stationarity, providing a stable feature foundation for the alignment and effective modeling of non-stationary time-series data. Addressing the pain point of utilizing information across channels, channel-dimensional clustering overcomes the limitations of independent channel strategies. By uncovering collaborative response patterns and spectral coupling patterns among multiple monitoring channels, it fully releases the value of cross-channel interactive information. Furthermore, the aggregation effect of intra-cluster features weakens the interference of noise on channel-related strategies, achieving in-depth mining and collaborative utilization of the value of multi-dimensional monitoring data. By fusing these clustered differential features, corrosion correlation information from different dimensions can be effectively integrated, constructing a three-dimensional feature space that can comprehensively characterize the overall corrosion of steel. This avoids the problems of information loss or misjudgment when analyzing single features, and enables in-depth mining and collaborative utilization of the value of multi-dimensional monitoring data, thereby obtaining more accurate prediction results. Attached Figure Description
[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0016] Figure 1 This is one of the flowcharts illustrating a time-frequency bi-clustering method for predicting steel corrosion provided in this specification. Figure 2 This is a schematic diagram of a steel corrosion prediction model provided in this specification; Figure 3 This is the second flowchart illustrating a time-frequency biclustering method for predicting steel corrosion provided in this specification. Figure 4 A schematic diagram of a time-frequency biclustering steel corrosion prediction device provided in this specification; Figure 5 This is a schematic diagram of a computer device provided for this specification. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0018] The technical solution provided by this invention can be applied to steel corrosion prediction scenarios. Currently, steel corrosion monitoring time-series data prediction mainly relies on traditional time-series models, such as random forests, long short-term memory (LSTM) artificial neural networks, and improved Transformer-based models, such as PatchTST and iTransformer. These methods face three core problems in practical applications: insufficient sensitivity to high-frequency signals in corrosion data, limited utilization of full-spectrum information, leading to insufficient capture of corrosion mutation features; significant non-stationary temporal distribution shifts in steel corrosion data, making it difficult for traditional models to achieve data alignment and effective modeling; and complex dependencies between channel monitoring data (temperature, humidity, corrosion rate, etc.), with existing channel-independent (CI) strategies ignoring cross-channel interactions and channel-dependent (CD) strategies being susceptible to noise interference and unable to accurately mine useful correlations between channels. These problems result in low prediction accuracy, making it difficult to meet the needs of accurate corrosion early warning in engineering scenarios. The closest existing solutions are the iTransformer-based multivariate time series prediction method, which uses inverted embedding to process corrosion monitoring time series data and a self-attention mechanism to model temporal dependencies; and the DUET-based bidirectional clustering-enhanced multivariate time series prediction method. However, these solutions have significant drawbacks:
[0019] 1. Heuristic shortcomings of traditional frequency domain feature selection. Existing solutions employ top-K or random K-frequency selection strategies, which are essentially heuristic screenings based on empirical thresholds. This method has two major problems: ① Low utilization of full-spectrum information: The importance of frequency domain features is not absolutely ranked. Top-K may miss frequency components that are strongly correlated with the target but ranked lower. For example, high-frequency abrupt signals around 30Hz in steel corrosion data may be filtered out due to their low ranking, resulting in significant loss of full-spectrum information utilization. ② Introduction of suboptimal correlations: The selected K frequencies may contain redundant or negatively correlated components. For example, low-frequency components of environmental humidity signals have a spurious correlation with high-frequency components of corrosion rates. Forcibly retaining these components will interfere with model learning and reduce the effective contribution of frequency domain features.
[0020] 2. Insufficient handling of non-stationary temporal distribution shifts. Existing solutions rely solely on the implicit alignment capability of self-attention mechanisms, lacking a dedicated module for handling distribution shifts: ① Data alignment failure: During seasonal fluctuations in steel corrosion data (e.g., rainy season vs. dry season) or equipment aging, the temporal distribution will experience significant shifts. Self-attention mechanisms can only adapt the distribution through global weight adjustments, failing to differentiate the modeling of different distribution clusters, resulting in numerous data alignment errors. ② Accumulated prediction errors: In long-term prediction scenarios, the lack of handling of distribution shifts leads to an exponential accumulation of errors. For example, in long-term corrosion prediction, iTransformer's prediction error increases by an additional 0.02 mm / a (millimeters per year) every 24 hours due to distribution shifts.
[0021] 3. Imbalance in Single-Channel Modeling. Existing solutions employ a single strategy of "full-channel dependency" or "full-channel independence," which fails to balance channel relationships: ① Information loss due to channel independence: If channels such as temperature, humidity, and corrosion rate are modeled independently, strong causal relationships such as "humidity → corrosion rate" and "temperature → corrosion rate" are ignored, resulting in a loss of feature representation ability. ② Channel-related noise interference: If full-channel dependency is forced, noise from redundant channels such as equipment vibration and electromagnetic interference will propagate through correlation, reducing the signal-to-noise ratio of the target channel (such as corrosion rate) by more than 10dB, significantly decreasing prediction accuracy.
[0022] 4. Insufficient ability to capture high-frequency abrupt changes. Steel corrosion data contains high-frequency abrupt changes such as pitting corrosion and stress corrosion (manifesting as spikes in the time domain and concentrated above 20Hz in the frequency domain). Existing solutions are insufficient in capturing these signals due to the following reasons: ① Model structure limitations: The self-attention mechanism of iTransformer has weak perception of local high-frequency features, and the bidirectional clustering of the DUET framework is not optimized for frequency domain characteristics, resulting in lower feature responsivity for high-frequency signals compared to this invention. ② Long-term accuracy degradation: In long-term prediction, the loss of high-frequency features leads to the model's failure to predict abrupt changes.
[0023] In summary, existing technologies have significant shortcomings in four dimensions: frequency domain information utilization, non-stationary processing, channel relationship modeling, and high-frequency signal capture. They cannot meet the engineering requirements of high-precision and high-robust time series prediction for steel corrosion monitoring, thus providing a clear direction for improvement for the innovative design of this invention.
[0024] The core objective of this invention is to address the key technical deficiencies of existing time-series data prediction methods for steel corrosion monitoring in practical engineering applications. Specifically, these include: addressing the problems of low utilization rate of full-spectrum information in corrosion data, insufficient capture of high-frequency abrupt signals, difficulties in data alignment and modeling caused by non-stationary time-series distribution offsets, and the complex dependencies between multi-channel monitoring data and the difficulty in balancing independence and correlation in single-channel modeling strategies. Ultimately, this invention provides a high-precision, highly robust multivariate time-series prediction scheme to meet the needs of accurate early warning of steel corrosion in engineering scenarios.
[0025] The following is combined Figures 1-3 The present invention describes a time-frequency bi-clustering method for predicting steel corrosion.
[0026] Figure 1 This is one of the flowcharts of the time-frequency biclustering method for predicting steel corrosion provided by the present invention, such as... Figure 1 As shown, the method includes the following: S101. Obtain multi-channel time-domain data of the target steel from multiple monitoring channels, perform frequency domain transformation on the multi-channel time-domain data, and obtain the corresponding multi-channel frequency-domain data of the target steel.
[0027] S102. Perform time-series feature extraction on the multi-channel time-domain data to obtain the multi-channel steel corrosion time-domain features; perform frequency-domain feature extraction on the multi-channel frequency-domain data to obtain the multi-channel steel corrosion frequency-domain features. S103. Cluster the time-domain features of multi-channel steel corrosion along the channel dimension and the time dimension respectively to obtain time-domain channel features including multiple cooperative clusters and time-domain time features including multiple time-series evolution clusters; wherein, each cooperative cluster corresponds to a cooperative response mode of multiple monitoring channels, and each time-series evolution cluster corresponds to a time evolution mode of corrosion. S104. Cluster the frequency domain features of multi-channel steel corrosion along the channel dimension and time dimension respectively to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters; wherein, each coupling cluster corresponds to a spectral coupling mode of multiple monitoring channels; each spectral evolution cluster corresponds to a spectral evolution mode of corrosion. S105. The time-domain channel features, time-domain time features, frequency-domain channel features, and frequency-domain time features are fused to obtain fused features; the corrosion prediction results of the target steel are obtained based on the fused features.
[0028] based on Figure 1In the illustrated embodiment, for example, in step S101 above, the multiple monitoring channels include monitoring channels for physical quantities directly related to the corrosion behavior of steel, such as temperature monitoring channels, humidity monitoring channels, corrosion rate monitoring channels, stress-strain monitoring channels, and environmental salt spray concentration monitoring channels. This application does not impose any limitations on these channels. The multi-channel time-domain data includes the original time-series data collected from multiple monitoring channels of the steel to be predicted under continuous time-series conditions. It can intuitively reflect the real-time dynamic characteristics of each monitoring indicator changing over time. The multi-channel frequency-domain data is obtained by performing frequency-domain transformations on the time-series data of multiple monitoring channels. The frequency-domain transformation method can be Fast Fourier Transform (FFT), or commonly used frequency-domain analysis methods such as Discrete Wavelet Transform (DWT) and Hilbert-Huang Transform (HHT). The appropriate method can be selected based on the actual monitoring data characteristics and analysis needs. This application does not impose any limitations on these methods. In one specific embodiment, the dimensions of the multi-channel time-domain data are [batch B, time step T, number of channels N], where B is the number of sample batches, representing the total number of steel corrosion sample groups analyzed in parallel; T is the time-series acquisition length of a single sample, corresponding to the number of time-series data acquisition points of a single steel sample within the monitoring period; and N is the total number of monitoring channels, corresponding one-to-one with the types of monitoring channels actually deployed. This three-dimensional data structure can simultaneously support the sample parallelism of the batch dimension, the evolutionary continuity of the time dimension, and the index diversity of the channel dimension, providing a structured data foundation for subsequent feature extraction and cluster analysis.
[0029] After obtaining the multi-channel time-domain and multi-channel frequency-domain data of the target steel, the data analysis in steps S102-S105 can be performed. It should be noted that steps S102-S105 are implemented using a pre-trained steel corrosion prediction model. Figure 2 This is a schematic diagram of a steel corrosion prediction model provided in this specification, such as... Figure 2 As shown, the steel corrosion prediction model includes a time-domain encoding transformer module, a FilterNet module (including PaiFilter time term FT and TexFilter channel term FT), a clustering module (including K-means clustering and hierarchical clustering), a multimodal feature fusion module (including channel / cross-domain processing, dimension alignment, and fully connected mapping), and a PatchTST module. The model execution process also includes data preprocessing (jittering, scaling, permutation, etc.) and data augmentation (RevNIN normalization and feature exchange).
[0030] In step S102, the multi-channel time domain data (i.e. Figure 2In step S102, time-series features are extracted from the corrosion current time-series data in the pre-trained steel corrosion prediction model using the time-domain encoding transformer module. In step S103 and S104, clustering operations are performed along the channel and time dimensions, respectively, using the clustering module (K-means clustering for the channel dimension and hierarchical clustering based on DTW distance) in the pre-trained steel corrosion prediction model. In step S105, the time-domain channel features, time-domain time features, frequency-domain channel features, and frequency-domain time features are fused using the multi-modal feature fusion module (including channel / cross-domain feature processing, dimension alignment, and fully connected mapping) in the pre-trained steel corrosion prediction model. In step S106, the corrosion prediction result of the target steel is obtained based on the fused features using the PatchTST module in the pre-trained steel corrosion prediction model.
[0031] Based on any of the above embodiments, for example, in step S102 above, extracting time-series features from the multi-channel time-domain data to obtain multi-channel steel corrosion time-domain features specifically includes: Multi-channel time-domain data is input into a pre-trained Transformer model for feature extraction and encoding. This Transformer model captures the long-range dependencies of the multi-channel time-domain data in the time dimension through a self-attention mechanism, and outputs multi-channel steel corrosion time-domain features that retain the time dimension. The dimensions of these multi-channel steel corrosion time-domain features are [batch B, time step T, number of channels N, feature dimension d1] (where d1 is the feature encoding dimension of the Transformer model). The features are output separately for each of the N monitoring channels, forming N sets of single-channel time-domain features with dimensions [batch B, time step T, feature dimension d1]. The final multi-channel steel corrosion time-domain features include N sets of single-channel steel corrosion time-domain features corresponding to the N monitoring channels, and each set of single-channel steel corrosion time-domain features maintains the same time step as the multi-channel time-domain data in the time dimension.
[0032] Based on any of the above embodiments, for example, in step S102 above, frequency domain feature extraction is performed on the multi-channel frequency domain data to obtain multi-channel steel corrosion frequency domain features, specifically including: Multi-channel time-domain data is input into the FilterNet module for frequency domain feature extraction and encoding. Based on the FilterNet model, the FilterNet module adaptively extracts key frequency patterns related to steel corrosion through learnable frequency filtering kernels—PaiFilter and TexFilter. The PaiFilter kernel learns general spectrum modeling characteristics through pre-training, and the TexFilter kernel learns the spatiotemporal coupling characteristics of the corrosion spectrum through pre-training. The output is a multi-channel steel corrosion frequency domain feature that retains the time window dimension. The dimension of this multi-channel steel corrosion frequency domain feature is [batch B, number of time windows T, number of channels N, feature dimension d2] (where d2 is the feature encoding dimension of the FilterNet model). Feature extraction is performed for each of the N monitoring channels, thus forming N sets of single-channel steel corrosion frequency domain features with the dimension [batch B, number of time windows T, feature dimension d2]. Each set of single-channel steel corrosion frequency domain features maintains the same number of time windows as the multi-channel frequency domain data in the time window dimension, thereby preserving the time-varying spectral characteristics of the corrosion process.
[0033] Based on any of the above embodiments, for example, in step S103, the multi-channel steel corrosion time-domain features are clustered along the channel dimension and the time dimension respectively to obtain time-domain channel features including multiple cooperative clusters and time-domain time features including multiple time-series evolution clusters, specifically including: The temporal feature representation requires clustering along both the channel and time dimensions: For channel-based clustering, single-channel temporal features corresponding to N channels are used as clustering samples. The K-means clustering algorithm (using Euclidean distance between channels as a similarity metric) is applied to cluster the multi-channel steel corrosion temporal features, generating temporal channel features comprising multiple co-clusters. Each co-cluster represents a co-response pattern across multiple monitoring channels. For time-based clustering, features corresponding to T time steps are used as clustering samples. A hierarchical clustering algorithm (using Dynamic Time Warping (DTW) distance as a similarity metric) is applied to cluster the multi-channel steel corrosion temporal features, generating temporal features comprising multiple temporal evolution clusters. Each temporal evolution cluster represents a temporal evolution pattern of corrosion. This allows for the separate acquisition of temporal channel features representing aggregated index similarity and temporal features representing aggregated temporal pattern similarity.
[0034] Based on any of the above embodiments, for example, in step S104, the multi-channel steel corrosion frequency domain features are clustered along the channel dimension and the time dimension respectively to obtain frequency domain channel features including multiple coupled clusters and frequency domain time features including multiple spectral evolution clusters, specifically including: Frequency domain feature representation requires clustering along both the channel and time dimensions: When clustering along the channel dimension, the single-channel frequency domain features corresponding to N channels are used as clustering samples. The K-means clustering algorithm (using the Euclidean distance between channels as a similarity measure) is used to cluster the multi-channel steel corrosion frequency domain features, generating frequency domain channel features including multiple coupled clusters. Each coupled cluster represents a spectral coupling mode of multiple monitoring channels. When clustering along the time dimension, the features corresponding to T time windows are used as clustering samples. The hierarchical clustering algorithm (using the Dynamic Time Warping (DTW) distance as a similarity measure) is used to cluster the multi-channel steel corrosion frequency domain features, generating frequency domain time features including multiple spectral evolution clusters. Each spectral evolution cluster represents a spectral evolution mode of corrosion. In this way, the frequency domain channel features of aggregated index similarity and the frequency domain time features of aggregated time-frequency pattern similarity are obtained respectively.
[0035] Based on any of the above embodiments, for example, in step S105, the time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features are fused to obtain fused features. Specifically, this is achieved through the multimodal feature fusion module in the pre-trained steel corrosion prediction model. This module includes a channel dimension alignment unit, a time dimension alignment unit, and a mask attention fusion unit. The channel dimension alignment unit is used to align the time-domain channel features and frequency-domain channel features in dimensions to generate a channel representation vector with consistent dimensions. The time dimension alignment unit is used to align the time-domain temporal features and frequency-domain temporal features in dimensions to generate a temporal representation vector with consistent dimensions. The mask attention fusion unit then uses a mask attention mechanism to perform attention-weighted fusion of the channel representation vector and the temporal representation vector to finally obtain fused features containing multi-dimensional representations of the corrosion process.
[0036] Based on any of the above embodiments, for example, in step S105, the corrosion prediction result of the target steel is obtained according to the fusion features, which is achieved by the PatchTST module in the pre-trained steel corrosion prediction model; the PatchTST module divides the fusion features into time-series segments, and captures the nonlinear evolution law of corrosion contained in the time-series segments through a cross-segment self-attention mechanism to obtain the corrosion prediction result of the target steel; wherein, the cross-segment self-attention mechanism enhances the ability to identify the turning point of steel corrosion evolution through local-global feature interaction enhancement.
[0037] This invention achieves time-frequency domain collaborative optimization, non-stationary data alignment, and accurate modeling of channel relationships through learnable frequency filtering, temporal clustering, and sparse channel clustering, significantly improving prediction accuracy and environmental adaptability. Specific advantages are as follows: In terms of noise resistance and robustness, frequency domain filtering can prioritize high-frequency noise or suppress irrelevant frequency band interference, effectively addressing the problems of insufficient sensitivity to high-frequency signals and limited utilization of full-spectrum information in corrosion data. This allows for more thorough capture of corrosion mutation features, while filtering redundant interference signals, resulting in smoother and more stable prediction results, maintaining high performance even under noise interference in complex engineering environments. In terms of data alignment and effective modeling, core technologies such as temporal clustering are employed... This approach precisely addresses the significant non-stationary temporal distribution shift in steel corrosion data, laying a solid foundation for the alignment and effective modeling of non-stationary time-series data. It ensures the model more accurately captures the evolutionary patterns of corrosion time and avoids modeling biases caused by data distribution shifts. Regarding multi-channel collaboration, sparse channel clustering is used to uniformly model multi-channel sequences, breaking the limitations of traditional independent channel analysis. This allows for the simultaneous output of future sequences from each channel, accurately capturing relevant frequency band information between channels, fully exploring the coupling correlation between multi-channel monitoring data, and achieving synergistic utilization of multi-channel information. This further enhances the model's comprehensive characterization of corrosion states and improves overall prediction accuracy and environmental adaptability.
[0038] In another specific embodiment, the implementation steps of this method are as follows: Figure 3 As shown, Figure 3 The second flowchart of a time-frequency biclustering method for predicting steel corrosion provided in this specification includes: Data acquisition: Real-time multi-channel time-series data of steel corrosion is acquired through a sensor array. The sampling frequency is set to once every 10 minutes to obtain data such as temperature, humidity, and corrosion current for consecutive days (more than 10 days) to form the original multi-channel data matrix.
[0039] Data augmentation: Time-domain data augmentation performs three types of preprocessing operations on the original time-series data of steel corrosion (such as corrosion current, temperature, humidity, and other multi-channel data): random jittering applies ±5% numerical fluctuations to the data to simulate small fluctuations in actual monitoring; amplitude scaling adjusts the amplitude of the data by a factor of 0.8-1.2 to enhance the diversity of data distribution; and data point replacement shuffles the order of data points within a local time window (such as a continuous 24-hour data segment) to improve the model's robustness to changes in time sequence. Frequency-domain data augmentation performs a Fast Fourier Transform (FFT) on the original time-domain data to convert it to frequency-domain data, sets a frequency-domain amplitude threshold (such as retaining frequency components with amplitude > 0.1), filters out high-frequency noise interference, and obtains clean frequency-domain features.
[0040] Data encoding processing: Time-domain encoding inputs the time-domain enhanced data into the Transformer network for tokenization. A self-attention mechanism is used to capture long-range dependencies in the time dimension, generating a time-domain feature representation. Frequency-domain encoding inputs the time-series data into the PaiFilter in FilterNet for frequency filtering in the time dimension, and the TexFilter for frequency filtering in the channel dimension. A learnable frequency filter kernel adaptively extracts key frequency patterns related to steel corrosion (such as frequency components corresponding to high-frequency abrupt signals), generating a frequency-domain feature representation.
[0041] Temporal clustering: Hierarchical clustering using K-means clustering and DTW distance is employed to filter strongly correlated channels and cluster temporal features, dividing non-stationary temporal distribution clusters. The frequency domain also includes temporal and channel-specific clustering units, clustering the frequency-domain temporal features output by the PaiFilter and TeXFilter modules in FilterNet to align with frequency domain distribution offsets. Channel-specific clustering units: Cluster frequency-domain channel features to filter redundant frequency channels.
[0042] Feature fusion and prediction: The clustering results in the time domain and frequency domain are fused with the temporal clustering features and the channel mask matrix through a mask attention mechanism to generate the final feature representation; the improved PatchTST model is input, with 100 iterations and a learning rate of 0.001, to complete the prediction of steel corrosion time series data.
[0043] Results output and early warning: The prediction results are displayed visually. An early warning is triggered when the predicted corrosion rate exceeds a preset threshold. At the same time, the entire process data is stored in the database.
[0044] Experimental data show that this method has been validated on 15 real datasets in 6 fields including Electricity, ETT, Weather, and a dedicated dataset for steel corrosion. Compared with baseline models such as iTransformer and PatchTST, it reduces MSE by an average of 7.1% and MAE by 6.5%.
[0045] The time-frequency binarization steel corrosion prediction device provided by the present invention will be described below. The time-frequency binarization steel corrosion prediction device described below can be referred to in correspondence with the time-frequency binarization steel corrosion prediction method described above.
[0046] Figure 4 A schematic diagram of a time-frequency biclustering steel corrosion prediction device provided by the present invention is shown below. For example, please refer to... Figure 4 As shown, the time-frequency biclustering steel corrosion prediction device may include: The data acquisition module is used to acquire multi-channel time-domain data of the target steel from multiple monitoring channels, and to perform frequency domain transformation on the multi-channel time-domain data to obtain the corresponding multi-channel frequency-domain data of the target steel.
[0047] The feature extraction module is used to extract time-series features from multi-channel time-domain data to obtain multi-channel steel corrosion time-domain features; and to extract frequency-domain features from multi-channel frequency-domain data to obtain multi-channel steel corrosion frequency-domain features.
[0048] The temporal clustering module is used to cluster the temporal features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain temporal channel features including multiple cooperative clusters and temporal time features including multiple temporal evolution clusters. Each cooperative cluster corresponds to a cooperative response mode of multiple monitoring channels, and each temporal evolution cluster corresponds to a temporal evolution mode of corrosion.
[0049] The frequency domain clustering module is used to cluster the frequency domain features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters. Each coupling cluster corresponds to a spectral coupling mode of multiple monitoring channels, and each spectral evolution cluster corresponds to a spectral evolution mode of corrosion.
[0050] The prediction module is used to fuse time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features to obtain fused features; and to obtain corrosion prediction results for the target steel based on the fused features.
[0051] Specific limitations regarding the time-frequency binarization steel corrosion prediction device can be found in the limitations on steel corrosion prediction described above, and will not be repeated here. Each module in the aforementioned time-frequency binarization steel corrosion prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0052] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method is a time-frequency bi-clustering method for predicting steel corrosion.
[0053] This instruction manual also provides Figure 5 The schematic diagram of the computer device shown is as follows: Figure 5At the hardware level, the computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The provided method is a time-frequency bi-clustering method for predicting steel corrosion.
[0054] 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 methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0055] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A time-frequency biclustering method for predicting steel corrosion, characterized in that, include: Multi-channel time-domain data of the target steel is acquired from multiple monitoring channels, and frequency-domain transformation is performed on the multi-channel time-domain data to obtain the corresponding multi-channel frequency-domain data of the target steel. Temporal features were extracted from multi-channel time-domain data to obtain the temporal features of steel corrosion in multiple channels. Frequency domain features are extracted from multi-channel frequency domain data to obtain multi-channel steel corrosion frequency domain features; Clustering of the temporal features of multi-channel steel corrosion along the channel dimension and the time dimension respectively yields temporal channel features including multiple cooperative clusters and temporal features including multiple temporal evolution clusters. Each cooperative cluster represents a cooperative response mode of multiple monitoring channels, and each temporal evolution cluster represents a temporal evolution mode of corrosion. The frequency domain features of multi-channel steel corrosion are clustered along both the channel and time dimensions to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters. Each coupling cluster represents a spectral coupling mode of multiple monitoring channels, and each spectral evolution cluster represents a spectral evolution mode of corrosion. The time-domain channel features, time-domain time features, frequency-domain channel features, and frequency-domain time features are fused to obtain fused features; the corrosion prediction results of the target steel are obtained based on the fused features.
2. The time-frequency bi-clustering method for predicting steel corrosion as described in claim 1, characterized in that, The clustering of multi-channel steel corrosion time-domain features along both the channel and time dimensions, and the clustering of multi-channel steel corrosion frequency-domain features along both the channel and time dimensions, are both implemented through a clustering module in a pre-trained steel corrosion prediction model; the clustering module includes: The channel-dimensional clustering unit is used to cluster the time-domain features of multi-channel steel corrosion along the channel dimension using the K-means clustering algorithm, generating time-domain channel features including multiple co-clusters; it is also used to cluster the frequency-domain features of multi-channel steel corrosion along the channel dimension using the K-means clustering algorithm, generating frequency-domain channel features including multiple coupled clusters; wherein, the K-means clustering algorithm uses the Euclidean distance between channels as a similarity measure; The time-dimensional clustering unit is used to cluster the time-domain features of multi-channel steel corrosion along the time dimension using a hierarchical clustering algorithm, generating time-domain features including multiple time-series evolution clusters; it is also used to cluster the frequency-domain features of multi-channel steel corrosion along the time dimension using a hierarchical clustering algorithm, generating frequency-domain features including multiple spectral evolution clusters; wherein, the hierarchical clustering algorithm uses the Dynamic Time Warping (DTW) distance as a similarity metric.
3. The time-frequency biclustering method for predicting steel corrosion as described in claim 1, characterized in that, The frequency domain feature extraction of multi-channel frequency domain data is achieved through the FilterNet module in the pre-trained steel corrosion prediction model. The FilterNet module extracts spectral features from the multi-channel frequency domain data using the PaiFilter filter kernel, which learns general spectral modeling characteristics during pre-training, and the TexFilter filter kernel, which learns the spatiotemporal coupling characteristics of corrosion spectrum. The multi-channel steel corrosion frequency domain features include N sets of single-channel steel corrosion frequency domain features corresponding to N monitoring channels, where N is a positive integer. Each set of single-channel steel corrosion frequency domain features maintains the same number of time windows as the multi-channel frequency domain data in the time window dimension to preserve the time-varying spectral characteristics of the corrosion process.
4. The time-frequency biclustering method for predicting steel corrosion as described in claim 1, characterized in that, The extraction of time-series features from multi-channel time-domain data is achieved through the Transformer module in a pre-trained steel corrosion prediction model. The Transformer module captures the long-range dependencies of multi-channel time-domain data in the time dimension through a self-attention mechanism, and outputs multi-channel steel corrosion time-domain features. The multi-channel steel corrosion time-domain features include N sets of single-channel steel corrosion time-domain features corresponding to N monitoring channels, where N is a positive integer, and each set of single-channel steel corrosion time-domain features maintains the same time step as the multi-channel time-domain data in the time dimension.
5. The time-frequency biclustering method for predicting steel corrosion as described in claim 1, characterized in that, The fusion of time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features is achieved through a multimodal feature fusion module in a pre-trained steel corrosion prediction model; the multimodal feature fusion module includes: The channel dimension alignment unit is used to align the time-domain channel features and frequency-domain channel features in dimensions, generating channel representation vectors with consistent dimensions. The time dimension alignment unit is used to align the time-domain time features and the frequency-domain time features in dimensions, and generate time-series representation vectors with consistent dimensions. The mask attention fusion unit is used to perform attention-weighted fusion of channel representation vectors and temporal representation vectors through the mask attention mechanism to obtain fused features that contain multi-dimensional representations of the erosion process.
6. The time-frequency biclustering method for predicting steel corrosion as described in claim 1, characterized in that, The corrosion prediction result of the target steel is obtained based on the fused features, which is achieved through the PatchTST module in the pre-trained steel corrosion prediction model. The PatchTST module divides the fused features into time-series segments and captures the nonlinear evolution law of corrosion contained in the time-series segments through a cross-segment self-attention mechanism to obtain the corrosion prediction result of the target steel. The cross-segment self-attention mechanism enhances the ability to identify the turning point of steel corrosion evolution through local-global feature interaction enhancement.
7. A time-frequency biclustering device for predicting steel corrosion, characterized in that, include: The data acquisition module is used to acquire multi-channel time-domain data of the target steel from multiple monitoring channels, and to perform frequency domain transformation on the multi-channel time-domain data to obtain the corresponding multi-channel frequency-domain data of the target steel. The feature extraction module is used to extract time-series features from multi-channel time-domain data to obtain multi-channel steel corrosion time-domain features; Frequency domain features are extracted from multi-channel frequency domain data to obtain multi-channel steel corrosion frequency domain features; The temporal clustering module is used to cluster the temporal features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain temporal channel features including multiple cooperative clusters and temporal time features including multiple time-series evolution clusters; wherein, each cooperative cluster corresponds to a cooperative response mode of multiple monitoring channels, and each time-series evolution cluster corresponds to a time evolution mode of corrosion. The frequency domain clustering module is used to cluster the frequency domain features of multi-channel steel corrosion along the channel dimension and the time dimension, respectively, to obtain frequency domain channel features including multiple coupling clusters and frequency domain time features including multiple spectral evolution clusters; wherein, each coupling cluster corresponds to a spectral coupling mode of multiple monitoring channels; and each spectral evolution cluster corresponds to a spectral evolution mode of corrosion. The prediction module is used to fuse time-domain channel features, time-domain temporal features, frequency-domain channel features, and frequency-domain temporal features to obtain fused features; and to obtain corrosion prediction results for the target steel based on the fused features.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the time-frequency biclustering method for predicting steel corrosion as described in any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the time-frequency biclustering method for predicting steel corrosion as described in any one of claims 1 to 6.