An industrial steam pipeline temperature prediction method based on deep learning

Abstract

This invention discloses a deep learning-based method for predicting the temperature of industrial steam pipelines, comprising: collecting multi-source operating condition data and preprocessing it to form a multivariate time series; employing multivariate variational mode decomposition to decompose it into multivariate modal components; extracting amplitude and phase features to obtain a basic feature map; constructing an improved MobileViT network to generate preliminary temperature prediction results; calculating the loss value to obtain a fixed parameter set; and solidifying the control platform to obtain the final temperature prediction result. This invention, through multivariate variational mode decomposition and an improved MobileViT network, achieves high-precision real-time prediction of the temperature field at key locations and the overall temperature field of industrial steam pipelines under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a method for predicting the temperature of industrial steam pipelines based on deep learning. Background Technology

[0002] Existing industrial steam pipeline networks generally employ distributed control devices, field instruments, and monitoring platforms to monitor and regulate pipeline operation. Key parameters include temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature at various measuring points. Due to factors such as large fluctuations in operating conditions, frequent load switching, complex start-up and shutdown conditions, long distances with multiple branches in the pipeline layout, and uneven insulation, the heat transfer and flow processes of steam in pipelines exhibit strong nonlinearity, strong coupling, and significant time-varying characteristics. The temperature field distribution along the pipeline is difficult to directly reflect using a limited number of measuring points. Current engineering projects typically rely on empirical formulas, simplified mechanistic models, or interpolation of data from a few measuring points to estimate pipeline temperature. Some solutions use static safety margins for coarse control. Current methods are heavily reliant on model parameters and experience, and have poor adaptability to complex operating conditions and structural changes. When the number of measuring points is limited, on-site noise is high, or sensors malfunction or drift, it is difficult to accurately and promptly characterize the temperature field distribution at key locations, easily leading to problems such as localized overheating, thermal stress concentration, high energy consumption, or fluctuations in steam quality.

[0003] With the improvement of industrial automation and informatization, temperature prediction methods based on traditional time series models or ordinary machine learning algorithms have begun to be introduced in some situations to model and predict multi-source operating data. For example, linear regression, autoregressive moving average models, or simple neural networks are used to make short-term predictions of temperature at a single measuring point. However, these methods often focus on a single channel or a few variables, making it difficult to fully utilize the coupling relationship between multiple variables such as temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature. They also have limited processing capabilities for non-stationary, multi-frequency mixed operating signals and struggle to characterize transmission delays and spatial correlations along the pipeline direction. Under conditions of limited computing power on-site and the need for compatibility with existing control platforms, general-purpose deep learning models are often complex in structure and have high computational overhead, which is not conducive to achieving high-precision real-time online prediction at the edge and makes it difficult to provide reliable data for steam pipeline network over-temperature early warning, valve regulation, load distribution, and energy-saving optimization control in a timely manner.

[0004] Therefore, how to provide a deep learning-based method for predicting the temperature of industrial steam pipelines is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for predicting the temperature of industrial steam pipelines. This invention achieves high-precision prediction of pipeline temperature fields under complex operating conditions through multivariate variational mode decomposition and improved MobileViT network collaborative modeling. Multi-source operating condition data, including temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature, are collected and preprocessed to form multivariate time series. Multivariate variational mode decomposition, incorporating time evolution constraints and phase consistency complex domain constraints, yields multivariate mode components that possess frequency band concentration, time evolution stability, and spatial phase consistency. A basic feature map reflecting the relationship between pipeline spatial location and operating condition characteristics is constructed. In terms of network structure, a dual-branch block combining axis-decoupled local coding, convolutional branches, and MobileViT branches, along with a temperature field prediction decoding head, is used to extract features from the basic feature map and reconstruct the temperature field. Compared with existing technologies, this invention can improve the prediction accuracy and robustness of the temperature field at key locations and the overall temperature field of industrial steam pipelines under conditions of limited measurement points, variable operating conditions, and high noise levels. It is suitable for real-time deployment and application on industrial field control platforms.

[0006] A deep learning-based method for predicting the temperature of industrial steam pipelines according to an embodiment of the present invention includes:

[0007] Multi-source operating condition data of industrial steam pipelines are collected, preprocessed, and segmented to form a multivariate time series.

[0008] Multivariate time series are processed by multivariate variational mode decomposition, which introduces time evolution constraints and complex domain constraints for phase consistency, decomposing the multi-channel time series into multiple multivariate mode components with frequency band concentration, time evolution stability and spatial phase consistency.

[0009] Amplitude and phase features are extracted from the multimodal components and combined with the multimodal time series. A two-dimensional feature map is constructed according to the pipeline spatial location and multimodal operating condition data to obtain the basic feature map.

[0010] An improved MobileViT network is constructed, which extracts and jointly represents the basic feature map through axis-decoupled local coding, introduces dual-branch blocks for parallel modeling and fusion processing, and introduces a temperature field decoding head for upsampling and mapping to generate preliminary temperature prediction results.

[0011] The preliminary temperature prediction results are compared with the actual measured temperatures at each measuring point. The loss value is calculated, the multivariate variational mode decomposition and MobileViT network parameters are adjusted, and the prediction and error calculation are repeated to obtain a fixed parameter set.

[0012] By embedding a fixed set of parameters into the control platform at the industrial site, and pre-setting a multivariate variational mode decomposition calculation program and an improved MobileViT network temperature field decoding calculation program, the final temperature prediction result is obtained.

[0013] Optionally, the multi-source operating condition data includes temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature at each measuring point.

[0014] Optionally, the preprocessing includes time alignment, outlier removal, missing value imputation, and normalization.

[0015] Optionally, forming a multivariate time series includes:

[0016] Multi-source operating condition data are collected synchronously at the industrial steam pipeline site according to a unified sampling cycle. The time stamp of the collection time is bound to the measurement value of each collection channel to form a multi-source operating condition raw data sequence with time stamp.

[0017] Time alignment processing is performed on the original data sequence of multi-source operating conditions. Data of various types from different acquisition channels at the same sampling time are paired according to time labels. Missing data points are filled with missing values ​​by interpolation. Data points that exceed the physical reasonable range are removed as outliers. The result is a multi-source operating condition data sequence that is time aligned and has completed missing value filling and outlier removal.

[0018] The multi-source operating condition data sequence, after time alignment and completion of missing value imputation and outlier removal, is normalized. The normalized multi-source operating condition data sequence is segmented according to the time window length and sliding step size. The temperature data, pressure data, flow data, valve opening data, boiler load data and ambient temperature data of each measuring point in each time window are arranged in chronological order to form a multivariate time series containing multiple physical quantities and multiple measuring points.

[0019] Optionally, the decomposition of the multi-channel time series into multiple multi-modal components with frequency band concentration, temporal evolution stability, and spatial phase consistency includes:

[0020] Multivariate variational mode decomposition is performed on the original time series according to each time window. Initial values ​​for the number of modes and the center frequency of each mode are set, and the corresponding modal components are initialized for each operating condition variable and each measuring point.

[0021] By introducing time evolution constraints, the modal components of each mode at adjacent time steps are constrained according to the first-order linear time evolution relationship. The first-order linear time evolution relationship is jointly represented by the time evolution coefficient and the time bias parameter. By alternately updating the modal components, time evolution coefficient and time bias parameter, the changes of each mode in the whole time window are formed into a unified first-order linear time evolution structure.

[0022] The multivariate time series is converted into a complex analytical form, and the modal components of each operating condition variable at each measuring point are represented by both amplitude and phase. The amplitude vector and phase vector corresponding to each mode are jointly iteratively updated in the complex domain.

[0023] Phase consistency constraints are introduced during the complex domain iteration process. The phase difference of the same mode between different measurement points is constrained according to the spatial position of the steam pipeline. The phase difference changes with the increase of spatial position according to a predetermined rule. Under the combined action of frequency band concentration constraints, time evolution constraints and phase consistency constraints, the modal components and center frequency are updated to obtain multi-mode components with frequency band concentration, time evolution stability and spatial phase consistency.

[0024] Optionally, obtaining the basic feature map includes:

[0025] Based on the multivariate modal components, amplitude and phase characteristics are calculated for each mode and each measurement point within each time window. The amplitude characteristics include instantaneous amplitude, average amplitude within the time window, and amplitude variance within the time window. The phase characteristics include instantaneous phase and phase difference within the time window.

[0026] In the time dimension, the amplitude and phase features are aligned with the multi-source operating condition data, and the amplitude features, phase features and multivariate time series of the same measurement point within the same time window are spliced ​​into a multidimensional feature vector sequence.

[0027] Using the pipeline spatial location index as the first dimension and the modal number and operating condition variable type as the second dimension, the multidimensional feature vector sequence is filled into each row and column of the two-dimensional matrix according to the spatial location and feature category. The features of multiple time steps within the same time window are superimposed on the channel dimension to form the basic feature map for temperature prediction.

[0028] Optionally, generating preliminary temperature prediction results includes:

[0029] An improved MobileViT network was constructed, consisting of an axis-decoupled local coding layer, a dual-branch block feature extraction layer, and a temperature field prediction decoding head layer.

[0030] The axis-decoupled local coding layer receives the basic feature map, performs a one-dimensional convolution operation in the pipeline spatial position dimension to obtain the first local feature map, and performs a two-dimensional convolution operation in the working condition and modal feature dimensions to obtain the second local feature map. The first local feature map and the second local feature map are added element-wise in the channel dimension, and the addition result is input into a one-to-one convolution layer for channel linear transformation and compression to generate the axis-decoupled local coding feature map.

[0031] The dual-branch block feature extraction layer is divided into a convolutional branch and a MobileViT branch. The convolutional branch performs convolution operations on the axis-decoupled local coding feature map through a depthwise separable convolutional layer to obtain the convolutional branch feature map. The MobileViT branch performs feature extraction and global dependency modeling on the axis-decoupled local coding feature map through a local convolutional sub-layer, a feature sequence unrolling sub-layer, a self-attention Transformer encoding sub-layer, and a feature reconstruction sub-layer to obtain the MobileViT branch feature map. The convolutional branch feature map and the MobileViT branch feature map are concatenated in the channel dimension and input into a 1x1 convolutional layer for channel fusion to form the backbone output feature map.

[0032] The temperature field prediction decoding head layer transforms the backbone output feature map into a two-dimensional temperature field feature map with the spatial location of the pipeline as one dimension and the prediction time step as the other dimension through an upsampling layer and a two-dimensional convolutional layer. A two-dimensional convolutional layer with linear activation is set at the end of the temperature field feature map to output the corresponding temperature prediction value for each spatial location and each prediction time step in the two-dimensional temperature field feature map, generating a preliminary temperature prediction result.

[0033] Optionally, obtaining the fixed parameter set includes:

[0034] Select multi-source operating condition data from historical operation phases and measured temperature data from each measuring point within the corresponding time period. Generate basic feature maps based on the improved MobileViT network. Use the basic feature maps corresponding to each time window as input samples. Arrange the measured temperatures of each measuring point within a predetermined number of prediction time steps after the current time window in spatial location and time order to form a supervised output temperature matrix.

[0035] The basic feature maps are sequentially input into the temperature prediction process consisting of multivariate variational mode decomposition and improved MobileViT network. Forward operation is performed to obtain the corresponding preliminary predicted temperature matrix. The predicted temperature value at each position in the preliminary predicted temperature matrix is ​​compared with the measured temperature value at the corresponding position in the supervised output temperature matrix. The errors of all spatial positions and all prediction time steps are weighted and summed and normalized according to the weighted square error method to obtain the loss value.

[0036] Based on the loss value, the decomposition parameters, time evolution constraint parameters, and phase consistency constraint parameters in the multivariate variational mode decomposition, as well as the convolution kernel coefficients, channel transformation weights, self-attention weights, and temperature field prediction decoder connection weights in the improved MobileViT network, are updated according to the gradient. The forward operation, loss calculation, and parameter update are repeatedly performed until the loss value is less than the preset threshold or the number of iterations reaches the preset upper limit. The current values ​​of all parameters are recorded as a fixed parameter set.

[0037] Optionally, obtaining the final temperature prediction result includes:

[0038] Write a fixed set of parameters into the industrial field control platform, load an executable program containing multivariate variational mode decomposition calculation and improved MobileViT temperature field prediction calculation into the control platform, and generate the operating configuration for industrial steam pipeline temperature prediction.

[0039] In the industrial field control platform, data input and output channels are configured between the steam pipeline field acquisition device, process control device, and monitoring device. Real-time multi-source operating condition data are received according to a unified sampling period. The multivariate variational mode decomposition operation associated with the fixed parameter set and the improved MobileViT temperature field prediction operation are called to perform mode decomposition, basic feature map construction, and temperature field prediction in sequence. The final temperature prediction results of each pipeline spatial location and multiple prediction time steps are sent to the monitoring interface and safety early warning unit through the data output channel.

[0040] The beneficial effects of this invention are:

[0041] This invention significantly enhances the modeling and prediction capabilities of temperature fields in industrial steam pipelines through a collaborative design of multivariate variational mode decomposition and an improved MobileViT network. Utilizing multivariate variational mode decomposition with introduced temporal evolution constraints and phase consistency complex domain constraints, it performs fine decomposition of multi-source operating condition data, including temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature. This enables the extraction of multivariate mode components with frequency band concentration, temporal evolution stability, and spatial phase consistency under strongly nonlinear, strongly coupled, and non-stationary conditions. By combining axis-decoupled local coding, a dual-branch block with convolutional and MobileViT branches in parallel, and a temperature field prediction decoding head, this invention maintains a lightweight network structure while enhancing its ability to express spatial correlation, operating condition coupling relationships, and the evolution of temperature fields over multiple future time points. Compared to traditional empirical models, simple time series models, and general deep networks, it exhibits higher prediction accuracy and robustness under conditions of limited measurement points, high noise, and variable operating conditions.

[0042] The temperature prediction process proposed in this invention can be embedded in an industrial field control platform with a fixed set of parameters and calculation programs. This enables online processing of real-time multi-source operating data and continuous prediction of temperature fields at various spatial locations in pipelines and at multiple future time steps. It facilitates integration with existing data acquisition devices, control devices, and monitoring interfaces. By predicting overheating trends at key locations in advance and providing a detailed characterization of the overall temperature field distribution, this invention supports safety early warning, valve regulation, and load distribution optimization in steam pipeline networks. It reduces the risk of localized overheating and energy waste, improves the operational efficiency and safety margin of steam transmission and distribution processes, and overcomes the technical bottlenecks of existing technologies, such as insufficient utilization of multi-source operating data, poor adaptability to complex operating conditions, and difficulty in achieving real-time prediction under limited computing power. Attached Figure Description

[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0044] Figure 1 This is a flowchart of a deep learning-based method for predicting the temperature of industrial steam pipelines proposed in this invention.

[0045] Figure 2 This is a structural block diagram of the multivariate variational mode decomposition of a deep learning-based industrial steam pipeline temperature prediction method proposed in this invention.

[0046] Figure 3 This is a functional diagram of the improved MobileViT network for an industrial steam pipeline temperature prediction method based on deep learning proposed in this invention. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0048] refer to Figure 1 , Figure 2 and Figure 3 A deep learning-based method for predicting the temperature of industrial steam pipelines, comprising:

[0049] Multi-source operating condition data of industrial steam pipelines are collected, preprocessed, and segmented to form a multivariate time series.

[0050] Multivariate time series are processed by multivariate variational mode decomposition, which introduces time evolution constraints and complex domain constraints for phase consistency, decomposing the multi-channel time series into multiple multivariate mode components with frequency band concentration, time evolution stability and spatial phase consistency.

[0051] Amplitude and phase features are extracted from the multimodal components and combined with the multimodal time series. A two-dimensional feature map is constructed according to the pipeline spatial location and multimodal operating condition data to obtain the basic feature map.

[0052] An improved MobileViT network is constructed, which extracts and jointly represents the basic feature map through axis-decoupled local coding, introduces dual-branch blocks for parallel modeling and fusion processing, and introduces a temperature field decoding head for upsampling and mapping to generate preliminary temperature prediction results.

[0053] The preliminary temperature prediction results are compared with the actual measured temperatures at each measuring point. The loss value is calculated, the multivariate variational mode decomposition and MobileViT network parameters are adjusted, and the prediction and error calculation are repeated to obtain a fixed parameter set.

[0054] By embedding a fixed set of parameters into the control platform at the industrial site, and pre-setting a multivariate variational mode decomposition calculation program and an improved MobileViT network temperature field decoding calculation program, the final temperature prediction result is obtained.

[0055] In this embodiment, the multi-source operating condition data includes temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature at each measuring point.

[0056] In this embodiment, the preprocessing includes time alignment, outlier removal, missing value imputation, and normalization.

[0057] In this embodiment, the formation of a multivariate time series includes:

[0058] Multi-source operating condition data are collected synchronously at the industrial steam pipeline site according to a unified sampling cycle. The time stamp of the collection time is bound to the measurement value of each collection channel to form a multi-source operating condition raw data sequence with time stamp.

[0059] Time alignment processing is performed on the original data sequence of multi-source operating conditions. Data of various types from different acquisition channels at the same sampling time are paired according to time labels. Missing data points are filled with missing values ​​by interpolation. Data points that exceed the physical reasonable range are removed as outliers. The result is a multi-source operating condition data sequence that is time aligned and has completed missing value filling and outlier removal.

[0060] The multi-source operating condition data sequence, after time alignment and completion of missing value imputation and outlier removal, is normalized. The normalized multi-source operating condition data sequence is segmented according to the time window length and sliding step size. The temperature data, pressure data, flow data, valve opening data, boiler load data and ambient temperature data of each measuring point in each time window are arranged in chronological order to form a multivariate time series containing multiple physical quantities and multiple measuring points.

[0061] In this embodiment, the decomposition of the multi-channel time series into multiple multi-modal components with frequency band concentration, temporal evolution stability, and spatial phase consistency includes:

[0062] Multivariate variational mode decomposition is performed on the original time series for each time window. Initial values ​​for the number of modes and the center frequency of each mode are set, and corresponding modal components are initialized for each operating condition variable and each measuring point.

[0063] The initial values ​​for the number of modes and the initial values ​​for the center frequencies of each mode are set as follows: a discrete Fourier transform is performed on the time series of each channel within each time window to obtain a complex spectrum. The amplitude of the spectrum is calculated within the frequency range and the local peak position of the amplitude is found. The peak frequency positions obtained from each channel are merged and deduplicated. The initial value for the number of modes is the number of peaks after merging. The initial value for the center frequency of each mode is the position of the peak frequency after merging. When the initial value for the number of modes is greater than the number of peaks, the frequency range is divided into equal intervals according to the initial value for the number of modes, and the remaining initial values ​​for the center frequencies are supplemented with the center frequencies of each interval.

[0064] The initialization of the corresponding modal components for each operating condition variable and each measuring point is specifically as follows: First, perform a Discrete Fourier Transform on the time series of the operating condition variable at the measuring point to obtain a complex spectrum. Then, randomly determine the frequency band range with an initial center frequency as the center. The frequency band range is obtained by extending one half-bandwidth to the left and right of the center frequency. The original complex values ​​of the frequency points in the complex spectrum that fall within the frequency band range are retained, and the complex values ​​of the frequency points that fall outside the frequency band range are set to zero to obtain a band-limited complex spectrum. Perform an Inverse Discrete Fourier Transform on the band-limited complex spectrum to obtain a time-domain sequence. Use the current time-domain sequence as the initial modal components of the operating condition variable at the measuring point.

[0065] A time evolution constraint is introduced, constraining the modal components of each mode at adjacent time steps according to a first-order linear time evolution relationship. This first-order linear time evolution relationship is jointly represented by time evolution coefficients and time bias parameters. By alternately updating the modal components, time evolution coefficients, and time bias parameters, the changes of each mode within the entire time window are formed into a unified first-order linear time evolution structure. Specifically, the first-order linear time evolution relationship, represented by time evolution coefficients and time bias parameters, is as follows:

[0066] For the same mode, at any two adjacent sampling times within the time window, the value of the mode component at the later sampling time is obtained by multiplying the time evolution coefficient by the value of the mode component at the previous sampling time, and adding the time offset parameter. The time evolution coefficient is used to describe the amplification, attenuation or maintenance ratio of the mode over time, and the time offset parameter is used to describe the overall translation of the mode between adjacent times. The time evolution coefficient is the covariance of the mode components at adjacent times divided by the variance of the mode components at the previous time, and the time offset parameter is the mean of the mode components at the later time minus the time evolution coefficient multiplied by the mean of the mode components at the previous time.

[0067] The multivariate time series is converted into a complex analytical form, where amplitude and phase are used to represent the modal components of each operating condition variable at each measurement point. In the complex domain, the amplitude vector and phase vector corresponding to each mode are jointly iteratively updated. Specifically, the conversion of the multivariate time series into a complex analytical form involves:

[0068] For each operating condition variable at each measuring point, the Hilbert transform of the time series is calculated to obtain an orthogonal component of the same length as the original series. Then, the original time series is used as the real part and the orthogonal component obtained by the Hilbert transform is used as the imaginary part to form the corresponding complex time series. In the complex time series, the amplitude at each sampling time is taken as the square root of the sum of the squares of the real part and the squares of the imaginary part, and the phase is taken as the arctangent of the ratio of the imaginary part to the real part, thus obtaining the complex analytical form characterized by amplitude and phase.

[0069] The Hilbert transform is a method that uses linear integration on a real-valued time series to generate a 90-degree phase shift component orthogonal to the original series. The orthogonal component and the original series together form an analytic signal.

[0070] Phase consistency constraints are introduced during the complex domain iteration process. The phase difference of the same mode between different measurement points is constrained according to the spatial position order of the steam pipeline. The phase difference changes with the increase of spatial position according to a predetermined rule. Under the combined action of frequency band concentration constraints, time evolution constraints, and phase consistency constraints, the modal components and center frequency are updated, resulting in multi-mode components with frequency band concentration, time evolution stability, and spatial phase consistency. The predetermined rule is as follows:

[0071] For the same mode, the measuring points are arranged according to their spatial positions along the steam flow direction. The phase difference between any two adjacent measuring points is required to be non-negative and not less than zero. At the same time, the difference between the phase differences of two adjacent segments is required to not exceed a preset threshold. The preset threshold is the product of the mode center frequency and the maximum allowable transmission delay between adjacent measuring points. The maximum allowable transmission delay is obtained by dividing the distance between adjacent measuring points by the minimum steam flow velocity.

[0072] In this embodiment, obtaining the basic feature map includes:

[0073] Based on the multivariate modal components, amplitude and phase characteristics are calculated for each mode and each measurement point within each time window. The amplitude characteristics include instantaneous amplitude, average amplitude within the time window, and amplitude variance within the time window. The phase characteristics include instantaneous phase and phase difference within the time window.

[0074] In the time dimension, the amplitude and phase features are aligned with the multi-source operating condition data, and the amplitude features, phase features and multivariate time series of the same measurement point within the same time window are spliced ​​into a multidimensional feature vector sequence.

[0075] Using the pipeline spatial location index as the first dimension and the modal number and operating condition variable type as the second dimension, the multidimensional feature vector sequence is filled into each row and column of a two-dimensional matrix according to spatial location and feature category. Features from multiple time steps within the same time window are superimposed on the channel dimension to form the basic feature map for temperature prediction. Specifically, filling the two-dimensional matrix with the multidimensional feature vector sequence according to spatial location and feature category involves:

[0076] All measuring points are numbered according to the steam flow direction, and the row number of the two-dimensional matrix is ​​determined, with each row corresponding to one measuring point. A unique column number is assigned to the amplitude feature, phase feature, and the value of each type of operating condition variable within the time window for each mode. The column number is generated according to the rules of mode number from small to large, amplitude first and then phase under the same mode, and operating condition variables arranged in a preset order. For each measuring point, each component in the multi-dimensional feature vector is extracted step by step within the same time window, and written into the row position corresponding to the measuring point according to the corresponding column number. Different time steps are written into different channels, so that the two-dimensional matrix on any channel represents the full feature distribution of each measuring point in the same time step.

[0077] In this embodiment, generating preliminary temperature prediction results includes:

[0078] An improved MobileViT network is constructed, consisting of an axis-decoupled local coding layer, a dual-branch block feature extraction layer, and a temperature field prediction decoding head layer, wherein:

[0079] The axis-decoupled local coding layer replaces the original MobileViT local convolution coding position. It adopts one-dimensional convolution along the pipeline spatial position dimension and two-dimensional convolution along the working condition and modal feature dimensions, and fuses them through one-to-one convolution to obtain local coding features. The dual-branch block feature extraction layer replaces the original trunk serial block structure. In each block, the convolutional branch and the MobileViT branch are connected in parallel and spliced ​​and fused at the end of the block for output. The temperature field prediction decoding head layer replaces the original pooling output method. After the trunk output, it is connected to upsampling and two-dimensional convolution to map to the temperature field grid and outputs the temperature prediction result through linear convolution.

[0080] The axis-decoupled local coding layer receives the basic feature map, performs a one-dimensional convolution operation in the pipeline spatial position dimension to obtain the first local feature map, and performs a two-dimensional convolution operation in the working condition and modal feature dimensions to obtain the second local feature map. The first local feature map and the second local feature map are added element-wise in the channel dimension, and the addition result is input into a one-to-one convolution layer for channel linear transformation and compression to generate the axis-decoupled local coding feature map.

[0081] The dual-branch block feature extraction layer consists of a convolutional branch and a MobileViT branch. The convolutional branch performs convolution operations on the axis-decoupled local encoded feature map using depthwise separable convolutional layers to obtain the convolutional branch feature map. The MobileViT branch extracts features from the axis-decoupled local encoded feature map and models global dependencies using local convolutional sub-layers, feature sequence unrolling sub-layers, self-attention Transformer encoding sub-layers, and feature reconstruction sub-layers to obtain the MobileViT branch feature map. The convolutional branch feature map and the MobileViT branch feature map are concatenated along the channel dimension and input into a 1x1 convolutional layer for channel fusion to form the backbone output feature map, where:

[0082] The local convolutional sublayer and the feature sequence unfolding sublayer are as follows: the local convolutional sublayer performs convolution operations on the axis-decoupled local encoded feature map to obtain a local representation and adjusts the channel dimension; the feature sequence unfolding sublayer divides the local representation into non-overlapping small blocks of a fixed size in the pipeline spatial position dimension and the feature dimension, and flattens each small block into a set of one-dimensional feature sequences in row priority order; the fixed block size is set to the fixed length of the pipeline spatial position dimension multiplied by the fixed length of the feature dimension.

[0083] The self-attention Transformer encoding sublayer and feature reconstruction sublayer are as follows: The self-attention Transformer encoding sublayer performs multi-head self-attention operation, feedforward operation, residual connection and normalization operation on the unfolded feature sequence to obtain a sequence representation containing global dependencies. The feature reconstruction sublayer folds the encoded sequence representation back into a two-dimensional feature map according to the position and arrangement of small blocks, and performs channel alignment and fusion output on the reconstructed feature map through convolution operation.

[0084] The temperature field prediction decoding head layer transforms the backbone output feature map into a two-dimensional temperature field feature map with the spatial location of the pipeline as one dimension and the prediction time step as the other dimension through an upsampling layer and a two-dimensional convolutional layer. A two-dimensional convolutional layer with linear activation is set at the end of the temperature field feature map to output the corresponding temperature prediction value for each spatial location and each prediction time step in the two-dimensional temperature field feature map, generating a preliminary temperature prediction result.

[0085] In this embodiment, obtaining the fixed parameter set includes:

[0086] Multi-source operating condition data from historical operation phases and measured temperature data from various measuring points within corresponding time periods are selected. A basic feature map is generated based on the improved MobileViT network. The basic feature map corresponding to each time window is used as input samples. The measured temperatures of each measuring point within a predetermined number of prediction time steps following the current time window are arranged according to spatial location and temporal order to form a supervised output temperature matrix. The predetermined number of prediction time steps specifically includes:

[0087] Determine the total future duration that the temperature prediction needs to cover, divide the total duration by the sampling period to obtain the prediction steps, and set the prediction time steps to be multiple consecutive moments starting from the end of the current time window and increasing by one sampling period. The total future duration is taken as the time length that does not exceed the control and regulation cycle of the steam pipeline and does not exceed the time length corresponding to the characteristic time scale of pipeline temperature change.

[0088] The basic feature maps are sequentially input into the temperature prediction process consisting of multivariate variational mode decomposition and improved MobileViT network. Forward operation is performed to obtain the corresponding preliminary predicted temperature matrix. The predicted temperature value at each position in the preliminary predicted temperature matrix is ​​compared with the measured temperature value at the corresponding position in the supervised output temperature matrix. The errors of all spatial positions and all prediction time steps are weighted and summed and normalized according to the weighted square error method to obtain the loss value.

[0089] Based on the loss value, the decomposition parameters, time evolution constraint parameters, and phase consistency constraint parameters in the multivariate variational mode decomposition, as well as the convolution kernel coefficients, channel transformation weights, self-attention weights, and temperature field prediction decoder head connection weights in the improved MobileViT network, are updated according to the gradient. The forward operation, loss calculation, and parameter update are repeatedly performed until the loss value is less than a preset threshold or the number of iterations reaches a preset upper limit. The current values ​​of all parameters are recorded as a fixed parameter set. Specifically, the preset threshold and preset upper limit are:

[0090] When the loss value is calculated by summing the weighted squared errors of all spatial locations and all prediction time steps and then dividing by the number of effective elements, the threshold is set to 0.25. The update stops when the average root mean square error is no greater than 0.5. The maximum number of iterations for parameter updates is 300 rounds. In each round, the forward operation, loss calculation, and gradient update are performed sequentially for all historical time windows. When the number of iterations reaches 300 rounds, the update stops and a fixed set of parameters is recorded.

[0091] In this embodiment, obtaining the final temperature prediction result includes:

[0092] Write a fixed set of parameters into the industrial field control platform, load an executable program containing multivariate variational mode decomposition calculation and improved MobileViT temperature field prediction calculation into the control platform, and generate the operating configuration for industrial steam pipeline temperature prediction.

[0093] In the industrial field control platform, data input and output channels are configured between the steam pipeline field acquisition device, process control device, and monitoring device. Real-time multi-source operating condition data are received according to a unified sampling period. The multivariate variational mode decomposition operation associated with the fixed parameter set and the improved MobileViT temperature field prediction operation are called to perform mode decomposition, basic feature map construction, and temperature field prediction in sequence. The final temperature prediction results of each pipeline spatial location and multiple prediction time steps are sent to the monitoring interface and safety early warning unit through the data output channel.

[0094] Example 1:

[0095] To verify the feasibility of this invention in practice, it was applied to a centralized steam supply network in a chemical industrial park, covering two boiler rooms and six vehicle workshops. The main pipeline is approximately 8.4 kilometers long, with branch lines of approximately 12 kilometers. Only 26 temperature monitoring points, 18 pressure monitoring points, and 14 flow monitoring points were installed along the pipelines. Valve opening and boiler load data were obtained from historical records on the control platform, with a sampling period of 10 seconds. From March to May 2025, the park experienced frequent plant start-ups and shutdowns and load switching, resulting in large fluctuations in steam flow, frequent valve operations, and significant non-stationarity and high-frequency noise in the temperature signal. Some monitoring points experienced short-term data loss and drift, causing traditional interpolation and empirical models to exhibit delayed predictions and false alarms at the end of the branch lines. This prevented the reliable early warning of key locations and the overall temperature field. Furthermore, the system could only be run on an edge industrial control computer, limiting computing power and network bandwidth.

[0096] In the current scenario, temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature are aligned by time and organized into a multivariate time series using a sliding time window. First, multivariate variational mode decomposition, incorporating time evolution constraints and complex domain phase consistency constraints, decomposes the multi-channel data of each window, obtaining multivariate mode components that simultaneously possess concentrated frequency bands, stable time evolution, and spatial phase consistency along the pipeline. Then, amplitude and phase features are extracted from the mode components and fused with multi-source operating condition data within the window, concatenating them into a basic feature map by multiplying spatial location by feature category. This basic feature map is input into an improved MobileViT network. It first undergoes axis-decoupled local encoding to extract spatial correlation and operating condition coupling local features, then a dual-branch block combining convolutional and MobileViT branches fuses local steady-state and global dependency information. Finally, a temperature prediction decoder outputs the temperature prediction result by multiplying the spatial location by the future time step. The prediction result is written back to the control platform screen in real time and linked to an early warning threshold strategy to indicate over-temperature trends and insufficient steam supply risks at the branch end.

[0097] The method of this invention was deployed in an industrial control computer, outputting a continuous multi-step temperature field with a prediction range of ten minutes in advance. It ran continuously in the field for sixty days, receiving signals from eight different operating conditions, with an average of approximately 500,000 valid samples per day and a missing rate between 0.5% and 1.2%. Noise mainly came from high-frequency fluctuations caused by flow meter jitter and valve position jitter. The actual operating temperature range was approximately 160 to 240 degrees Celsius. After going live, under four typical load surges, two valve malfunctions, and one short-term drop in steam supply pressure, the method consistently provided trend warnings before a significant temperature rise appeared at key end measuring points and simultaneously output the temperature field distribution changes along the line, helping operators adjust valve positions and load distribution in advance. Furthermore, it maintained continuous temperature field output even when a single temperature measuring point experienced a short-term loss of data, preventing interruptions in the monitoring interface.

[0098] Table 1 Comparison of Industrial Steam Pipeline Temperature Prediction Methods

[0099] index This invention Mechanism Model ARIMA LSTM CNN-LSTM MobileViT Mean absolute error (°C) 1.08 3.25 2.74 1.86 1.62 1.44 Root mean square error (°C) 1.52 4.38 3.71 2.54 2.23 2.06 Mean relative error (%) 0.72 2.18 1.86 1.22 1.05 0.96 Goodness of fit 0.982 0.845 0.881 0.934 0.948 0.956 Ninth percentile error (°C) 2.55 7.90 6.60 4.10 3.62 3.28 Inference delay (ms) 18 6 9 24 31 21 Number of parameters (M) 2.6 0.1 0.2 1.9 3.4 3.1 Online update time (s) 0.42 0.12 0.18 0.95 1.30 0.78

[0100] As shown in Table 1, in terms of accuracy metrics, this invention performs best in mean absolute error, root mean square error, mean relative error, and goodness of fit, with a mean absolute error of 1.08%, a root mean square error of 1.52%, a mean relative error of 0.72%, and a goodness of fit of 0.982. Compared to the mechanistic model and ARIMA, this invention exhibits significantly lower errors under complex conditions, indicating a more thorough characterization of nonlinearity and multi-source coupling. Compared to LSTM, CNN-LSTM, and the original MobileViT, this invention maintains a stable advantage among similar deep learning methods, demonstrating the role of multivariate variational mode decomposition and the improved MobileViT network structure in enhancing prediction accuracy.

[0101] In terms of stability and extreme errors, the 9th percentile error of this invention is 2.55℃, significantly lower than the original MobileViT's 3.28℃, CNN-LSTM's 3.62℃, and LSTM's 4.10℃, and also far lower than ARIMA's 6.60℃ and the mechanistic model's 7.90℃. The results indicate that under disturbances such as sudden load changes and frequent valve operations, this invention can effectively suppress the occurrence of large error samples, making temperature field prediction more robust and suitable for overheating early warning at key locations and determining temperature distribution along the pipeline.

[0102] From an engineering deployment perspective, this invention achieves an inference latency of 18 milliseconds, superior to LSTM and CNN-LSTM, and lower than the original MobileViT's 21 milliseconds, maintaining good real-time performance while ensuring accuracy. The number of parameters is 2.6 million, lower than CNN-LSTM and the original MobileViT, and the online update time is 0.42 seconds, significantly faster than LSTM and CNN-LSTM. Although the mechanistic model and ARIMA have lower latency and online update time, their overall accuracy and robustness are insufficient. This invention achieves a more balanced performance in terms of accuracy, real-time performance, and scalability under limited computing power.

[0103] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the temperature of industrial steam pipelines based on deep learning, characterized in that, include: Multi-source operating condition data of industrial steam pipelines are collected, preprocessed, and segmented to form a multivariate time series. Multivariate time series are processed by multivariate variational mode decomposition, which introduces time evolution constraints and complex domain constraints for phase consistency, decomposing the multi-channel time series into multiple multivariate mode components with frequency band concentration, time evolution stability and spatial phase consistency. Amplitude and phase features are extracted from the multimodal components and combined with the multimodal time series. A two-dimensional feature map is constructed according to the pipeline spatial location and multimodal operating condition data to obtain the basic feature map. An improved MobileViT network is constructed, which extracts and jointly represents the basic feature map through axis-decoupled local coding, introduces dual-branch blocks for parallel modeling and fusion processing, and introduces a temperature field decoding head for upsampling and mapping to generate preliminary temperature prediction results. The preliminary temperature prediction results are compared with the actual measured temperatures at each measuring point. The loss value is calculated, the multivariate variational mode decomposition and MobileViT network parameters are adjusted, and the prediction and error calculation are repeated to obtain a fixed parameter set. By embedding a fixed set of parameters into the control platform at the industrial site, and pre-setting a multivariate variational mode decomposition calculation program and an improved MobileViT network temperature field decoding calculation program, the final temperature prediction result is obtained.

2. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The multi-source operating condition data includes temperature, pressure, flow rate, valve opening, boiler load, and ambient temperature at each measuring point.

3. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The preprocessing includes time alignment, outlier removal, missing value imputation, and normalization.

4. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The formation of multivariate time series includes: Multi-source operating condition data are collected synchronously at the industrial steam pipeline site according to a unified sampling cycle. The time stamp of the collection time is bound to the measurement value of each collection channel to form a multi-source operating condition raw data sequence with time stamp. Time alignment processing is performed on the original data sequence of multi-source operating conditions. Data of various types from different acquisition channels at the same sampling time are paired according to time labels. Missing data points are filled with missing values ​​by interpolation. Data points that exceed the physical reasonable range are removed as outliers. The result is a multi-source operating condition data sequence that is time aligned and has completed missing value filling and outlier removal. The multi-source operating condition data sequence, after time alignment and completion of missing value imputation and outlier removal, is normalized. The normalized multi-source operating condition data sequence is segmented according to the time window length and sliding step size. The temperature data, pressure data, flow data, valve opening data, boiler load data and ambient temperature data of each measuring point in each time window are arranged in chronological order to form a multivariate time series containing multiple physical quantities and multiple measuring points.

5. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The process of decomposing a multi-channel time series into multiple multi-modal components with frequency band concentration, temporal evolution stability, and spatial phase consistency includes: Multivariate variational mode decomposition is performed on the original time series according to each time window. Initial values ​​for the number of modes and the center frequency of each mode are set, and the corresponding modal components are initialized for each operating condition variable and each measuring point. By introducing time evolution constraints, the modal components of each mode at adjacent time steps are constrained according to the first-order linear time evolution relationship. The first-order linear time evolution relationship is jointly represented by the time evolution coefficient and the time bias parameter. By alternately updating the modal components, time evolution coefficient and time bias parameter, the changes of each mode in the whole time window are formed into a unified first-order linear time evolution structure. The multivariate time series is converted into a complex analytical form, and the modal components of each operating condition variable at each measuring point are represented by both amplitude and phase. The amplitude vector and phase vector corresponding to each mode are jointly iteratively updated in the complex domain. Phase consistency constraints are introduced during the complex domain iteration process. The phase difference of the same mode between different measurement points is constrained according to the spatial position of the steam pipeline. The phase difference changes with the increase of spatial position according to a predetermined rule. Under the combined action of frequency band concentration constraints, time evolution constraints and phase consistency constraints, the modal components and center frequency are updated to obtain multi-mode components with frequency band concentration, time evolution stability and spatial phase consistency.

6. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The obtained basic feature map includes: Based on the multivariate modal components, amplitude and phase characteristics are calculated for each mode and each measurement point within each time window. The amplitude characteristics include instantaneous amplitude, average amplitude within the time window, and amplitude variance within the time window. The phase characteristics include instantaneous phase and phase difference within the time window. In the time dimension, the amplitude and phase features are aligned with the multi-source operating condition data, and the amplitude features, phase features and multivariate time series of the same measurement point within the same time window are spliced ​​into a multidimensional feature vector sequence. Using the pipeline spatial location index as the first dimension and the modal number and operating condition variable type as the second dimension, the multidimensional feature vector sequence is filled into each row and column of the two-dimensional matrix according to the spatial location and feature category. The features of multiple time steps within the same time window are superimposed on the channel dimension to form the basic feature map for temperature prediction.

7. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The generation of preliminary temperature prediction results includes: An improved MobileViT network was constructed, consisting of an axis-decoupled local coding layer, a dual-branch block feature extraction layer, and a temperature field prediction decoding head layer. The axis-decoupled local coding layer receives the basic feature map, performs a one-dimensional convolution operation in the pipeline spatial position dimension to obtain the first local feature map, and performs a two-dimensional convolution operation in the working condition and modal feature dimensions to obtain the second local feature map. The first local feature map and the second local feature map are added element-wise in the channel dimension, and the addition result is input into a one-to-one convolution layer for channel linear transformation and compression to generate the axis-decoupled local coding feature map. The dual-branch block feature extraction layer is divided into a convolutional branch and a MobileViT branch. The convolutional branch performs convolution operations on the axis-decoupled local coding feature map through a depthwise separable convolutional layer to obtain the convolutional branch feature map. The MobileViT branch performs feature extraction and global dependency modeling on the axis-decoupled local coding feature map through a local convolutional sub-layer, a feature sequence unrolling sub-layer, a self-attention Transformer encoding sub-layer, and a feature reconstruction sub-layer to obtain the MobileViT branch feature map. The convolutional branch feature map and the MobileViT branch feature map are concatenated in the channel dimension and input into a 1x1 convolutional layer for channel fusion to form the backbone output feature map. The temperature field prediction decoding head layer transforms the backbone output feature map into a two-dimensional temperature field feature map with the spatial location of the pipeline as one dimension and the prediction time step as the other dimension through an upsampling layer and a two-dimensional convolutional layer. A two-dimensional convolutional layer with linear activation is set at the end of the temperature field feature map to output the corresponding temperature prediction value for each spatial location and each prediction time step in the two-dimensional temperature field feature map, generating a preliminary temperature prediction result.

8. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The obtained fixed parameter set includes: Select multi-source operating condition data from historical operation phases and measured temperature data from each measuring point within the corresponding time period. Generate basic feature maps based on the improved MobileViT network. Use the basic feature maps corresponding to each time window as input samples. Arrange the measured temperatures of each measuring point within a predetermined number of prediction time steps after the current time window in spatial location and time order to form a supervised output temperature matrix. The basic feature maps are sequentially input into the temperature prediction process consisting of multivariate variational mode decomposition and improved MobileViT network. Forward operation is performed to obtain the corresponding preliminary predicted temperature matrix. The predicted temperature value at each position in the preliminary predicted temperature matrix is ​​compared with the measured temperature value at the corresponding position in the supervised output temperature matrix. The errors of all spatial positions and all prediction time steps are weighted and summed and normalized according to the weighted square error method to obtain the loss value. Based on the loss value, the decomposition parameters, time evolution constraint parameters, and phase consistency constraint parameters in the multivariate variational mode decomposition, as well as the convolution kernel coefficients, channel transformation weights, self-attention weights, and temperature field prediction decoder connection weights in the improved MobileViT network, are updated according to the gradient. The forward operation, loss calculation, and parameter update are repeatedly performed until the loss value is less than the preset threshold or the number of iterations reaches the preset upper limit. The current values ​​of all parameters are recorded as a fixed parameter set.

9. The method for predicting the temperature of industrial steam pipelines based on deep learning according to claim 1, characterized in that, The process of obtaining the final temperature prediction result includes: Write a fixed set of parameters into the industrial field control platform, load an executable program containing multivariate variational mode decomposition calculation and improved MobileViT temperature field prediction calculation into the control platform, and generate the operating configuration for industrial steam pipeline temperature prediction. In the industrial field control platform, data input and output channels are configured between the steam pipeline field acquisition device, process control device, and monitoring device. Real-time multi-source operating condition data are received according to a unified sampling period. The multivariate variational mode decomposition operation associated with the fixed parameter set and the improved MobileViT temperature field prediction operation are called to perform mode decomposition, basic feature map construction, and temperature field prediction in sequence. The final temperature prediction results of each pipeline spatial location and multiple prediction time steps are sent to the monitoring interface and safety early warning unit through the data output channel.

Images