Intelligent temperature control method and system for electric heat tracing band based on data analysis

Abstract

The application discloses an intelligent temperature control method and system for an electric heat tracing band based on data analysis, and belongs to the technical field of temperature control of electric heat tracing bands. The method obtains real-time and historical temperature data of multiple temperature monitoring points along the electric heat tracing band through a supervisory control and data acquisition system, performs time series decomposition on the temperature data, extracts a temperature trend component, a periodic component and a residual component, inputs the residual component into a pre-trained conditional variational autoencoder, generates a latent feature vector and fuses the latent feature vector with the temperature trend component to construct a dynamic temperature prediction model, calculates a predicted temperature change rate of each temperature monitoring point according to an output of the dynamic temperature prediction model, generates an adaptive heating power adjustment instruction and sends the instruction to a solid-state relay module when the predicted temperature change rate exceeds a set threshold, and thus the forward-looking adjustment of the heating power of the electric heat tracing band is realized.

Description

Technical Field

[0001] This invention relates to the field of electric heating cable temperature control technology, specifically to an intelligent temperature control method and system for electric heating cables based on data analysis. Background Technology

[0002] Electric heating cables are widely used in petrochemical pipeline insulation and pipeline antifreeze applications, and the accuracy of their temperature control directly affects process safety and energy consumption levels. Existing electric heating cable temperature control methods largely rely on supervisory control and data acquisition systems to collect real-time temperature values ​​from discrete temperature monitoring points along the pipeline. These values ​​are then compared with set temperature thresholds to directly generate on / off control or proportional adjustment commands. This approach only responds to temperature deviations at the current moment and cannot predict temperature change trends, resulting in heating power adjustments always lagging behind actual temperature fluctuations. This can easily lead to temperature overshoot and oscillations when environmental conditions change rapidly. Existing technologies, when processing temperature data from multiple monitoring points along the electric heating cable, often ignore the spatial correlation between temperature changes at each monitoring point and the complex patterns in the time series. Conventional methods only perform independent smoothing of single-point temperatures or use simple regression models for trend extrapolation, making it difficult to separate regular fluctuations caused by periodic environmental changes from non-stationary residuals caused by random disturbances from the temperature signal. This approach makes it impossible for prediction models to effectively distinguish the intrinsic driving factors of temperature changes, and prediction accuracy significantly decreases when facing complex operating conditions such as sensor noise and differences in local heat loss. In actual operation of electric heating cables, temperature data includes slowly changing long-term trends, periodic fluctuations caused by diurnal or seasonal changes, and residual disturbances caused by random events such as gusts and rainfall. Separating and utilizing these different components is crucial for improving prediction accuracy. However, with existing decomposition methods, residual components are often discarded as useless noise, ignoring the implicit inter-monitoring disturbance correlation characteristics and limiting the ability of prediction models to characterize complex temperature dynamics. Furthermore, existing control strategies fail to correlate the predicted rate of change with dynamic environmental conditions when generating power regulation, keeping the regulation threshold fixed and making it difficult to adapt to the differentiated needs of temperature change sensitivity under different operating conditions. Summary of the Invention

[0003] This paper presents a data analysis-based intelligent temperature control method and system for electric heating cables. It decomposes temperature data from multiple monitoring points along the cable into time series data and extracts implicit multi-point disturbance correlation features from the residual components using a conditional variational autoencoder. These features are then fused with trend components to construct a dynamic temperature prediction model, enabling accurate prediction of temperature change trends under complex operating conditions. Simultaneously, an adaptive power adjustment command is generated based on a comparison between the predicted temperature change rate and the dynamic adjustment threshold. This allows the heating power adjustment to respond in advance to impending temperature changes, applying correction before actual temperature deviations, thus improving the response speed and stability of the electric heating cable temperature control.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent temperature control method for electric heating cables based on data analysis, comprising: acquiring real-time temperature data and historical temperature data of multiple temperature monitoring points along the electric heating cable through a supervisory control and data acquisition system; performing time-series decomposition on the real-time temperature data and historical temperature data to extract temperature trend components, periodic components, and residual components; inputting the residual components into a pre-trained conditional variational autoencoder to generate latent feature vectors; fusing the latent feature vectors with the temperature trend components to construct a dynamic temperature prediction model; calculating the predicted temperature change rate for each temperature monitoring point based on the output of the dynamic temperature prediction model; and generating an adaptive heating power adjustment command when the predicted temperature change rate exceeds a set threshold, and sending the command to the solid-state relay module of the electric heating cable.

[0005] This method deeply mines temperature data from multiple points along the electric heating cable, separating macroscopic temperature change trends from complex random fluctuations. It utilizes a conditional variational autoencoder to extract high-level features from the difficult-to-model residual components, generating latent feature vectors that characterize the system's inherent uncertainties. By deeply fusing these latent feature vectors with temperature trend components, the constructed dynamic temperature prediction model can simultaneously capture both deterministic temperature evolution patterns and random disturbance characteristics, thus enabling high-precision predictions of temperature states at multiple future moments. Based on the predicted temperature change rate, it can detect abnormal temperature fluctuations in advance, generating refined heating power adjustment commands before the temperature deviates from the allowable range. Through solid-state relay modules, it implements independent and rapid power adjustment for each area of ​​the electric heating cable, forming a closed-loop control combining feedforward and feedback. This effectively avoids temperature overshoot and hysteresis, improving the temperature control accuracy and energy efficiency of the electric heating system in complex environments.

[0006] As a technical solution of the present invention, when acquiring real-time and historical temperature data from multiple temperature monitoring points along the electric heating cable through a monitoring, control, and data acquisition system, multiple thermocouple sensors are arranged at equal intervals on the surface of the electric heating cable as temperature monitoring points. The signal output terminal of each thermocouple sensor is connected to the data acquisition server of the monitoring, control, and data acquisition system through an analog input module. The monitoring, control, and data acquisition system polls the temperature values ​​reported by each thermocouple sensor according to a preset sampling period, and stores the temperature value collected at each sampling moment, along with the corresponding temperature monitoring point number and the timestamp of the sampling moment, into the real-time database of the monitoring, control, and data acquisition system. All historical temperature data for each temperature monitoring point within multiple consecutive sampling periods before the current sampling moment are extracted from the historical database of the monitoring, control, and data acquisition system. Linear interpolation is performed to fill missing values ​​in the historical temperature data, and outliers are removed based on the Laida criterion to obtain cleaned historical temperature data. Preferably, the thermocouple sensor is a K-type thermocouple, and a cold junction compensation circuit is connected in series between each thermocouple sensor and the analog input module. By performing interpolation and outlier removal on historical data, the continuity and consistency of input data are ensured, and outliers introduced by accidental sensor failures or signal transmission interference are eliminated from their influence on subsequent modeling and prediction. This lays a data foundation for improving the stability of time series decomposition and the reliability of prediction models.

[0007] Preferably, during the polling sampling of the monitoring control and data acquisition system, the data acquisition server sends a broadcast synchronization pulse to each analog input module. Upon receiving the broadcast synchronization pulse, each analog input module simultaneously latches the current temperature value of its connected thermocouple sensor. The data acquisition server sequentially reads the temperature values ​​latched by each analog input module in ascending order of temperature monitoring point number, and after reading, sends a synchronization pulse for the next sampling cycle to all analog input modules. The data acquisition server writes the mapping relationship between all temperature values ​​read each time and the sequence number of the current sampling cycle and the temperature monitoring point number into the real-time database. This synchronous acquisition mechanism ensures that the sampling times of all temperature monitoring points are strictly aligned, eliminating time deviations caused by time-division readings. This ensures that the acquired multi-point temperature data truly reflects the temperature field distribution along the electric heating tape at the same time, providing accurate spatial consistency assurance for multivariate time series analysis.

[0008] As a further improvement of the present invention, when performing time-series decomposition on the real-time temperature data and historical temperature data, the real-time temperature data of each temperature monitoring point at the current sampling time is concatenated with the cleaned historical temperature data to form the original temperature time series of that temperature monitoring point; the original temperature time series is iteratively decomposed using a local weighted regression scatter smoothing method. In each iteration, the periodic component is first extracted from the original temperature time series, then the temperature trend component is extracted from the remaining series, and finally the temperature trend component and periodic component are subtracted from the original temperature time series to obtain the residual component; the decomposition operation is performed on each temperature monitoring point to obtain the temperature trend component sequence, periodic component sequence, and residual component sequence corresponding to each temperature monitoring point. Preferably, each iteration includes: identifying the time interval between adjacent peaks and troughs from the original temperature time series, using the median of the time interval as the period length of the current iteration, constructing a periodic subsequence according to the period length, and performing local weighted regression smoothing on each periodic subsequence to obtain the estimated value of the periodic component of the current iteration; subtracting the estimated value of the periodic component of the current iteration from the original temperature time series to obtain a de-periodic sequence, and performing local weighted regression smoothing on the de-periodic sequence to obtain the estimated value of the temperature trend component of the current iteration; subtracting the estimated value of the temperature trend component and the estimated value of the periodic component of the current iteration from the original temperature time series to obtain the estimated value of the residual component of the current iteration, and calculating the root mean square error of the estimated value of the residual component between two adjacent iterations, stopping the iteration when the root mean square error is less than a preset convergence threshold.

[0009] This decomposition scheme adaptively identifies the implicit periodic fluctuation patterns in temperature sequences without requiring pre-setting fixed period parameters. Through iterative stripping, it decouples complex temperature signals into trend components, periodic components, and residual components with clear physical meanings. The trend component reflects the overall temperature trend of the electric heating tape caused by changes in heating power or slow environmental changes; the periodic component captures regular fluctuations caused by factors such as diurnal temperature differences or process cycles; and the residual component condenses random disturbances and local anomalies, allowing subsequent models to utilize the characteristics of different components in a divide-and-conquer manner.

[0010] As a preferred embodiment of the present invention, when the residual components are input into a pre-trained conditional variational autoencoder (CDAE), the residual components of all temperature monitoring points at the current sampling time are arranged into a residual vector according to the temperature monitoring point numbering order, and this residual vector is used as the input condition of the CDAE. The CDAE consists of an encoder network and a decoder network. The encoder network maps the residual vector to the mean vector and log-variance vector of a Gaussian distribution according to the input condition, and samples the latent feature vector from this Gaussian distribution. The decoder network concatenates the latent feature vector with the input condition to reconstruct the estimated value of the residual vector. During training, the relative entropy between the reconstruction error and the prior distribution of the latent feature vector is minimized. This process compresses and maps the high-dimensional and noisy residual vector to a low-dimensional continuous latent space. The latent feature vector removes redundant noise and captures the complex nonlinear correlations and intrinsic driving factors of abnormal patterns contained in the residual components. At the same time, by regularizing the latent space to approximate a standard normal distribution, good continuity and interpretability of the latent features are ensured, providing compact and information-rich auxiliary features for temperature prediction.

[0011] As a further optimization of the above technical solution, when fusing the latent feature vector with the temperature trend component, the latent feature vector of each temperature monitoring point is concatenated element-wise with the value of the temperature trend component of that temperature monitoring point at the current sampling time to form the fused feature vector of that temperature monitoring point. The fused feature vectors of all temperature monitoring points are stacked in order of temperature monitoring point number to form a fused feature matrix, and the fused feature matrix is ​​input into a time-convolutional network based on gated recurrent units. This time-convolutional network performs dilated convolution operations on the fused feature matrix along the time dimension to capture long-range dependencies between multiple sampling periods. The hidden state sequence output by the time-convolutional network is used as the internal state of a dynamic temperature prediction model. This dynamic temperature prediction model takes the fused feature matrix at the current sampling time as input and outputs the predicted temperature value of each temperature monitoring point at multiple future sampling times. The time-convolutional network based on gated recurrent units includes multiple dilated convolutional layers, and the dilation coefficient of each dilated convolutional layer increases exponentially with the number of network layers.

[0012] This fusion prediction architecture can uniformly encode the trend component describing the direction of macroscopic temperature changes and the potential feature vector reflecting microscopic uncertainties at the feature level, forming a fusion feature that combines deterministic and stochastic information. The temporal convolutional network based on gated recurrent units, with its exponentially increasing dilated convolutional structure, achieves a large receptive field while maintaining computational efficiency. It can capture spatiotemporal dependencies across multiple sampling periods from long sequences of the fusion feature matrix, giving the model a stronger ability to learn about the inertia, delay effects, and multi-point spatial coupling relationships of temperature changes, thereby significantly improving the accuracy and robustness of temperature predictions for multiple future moments.

[0013] As a specific implementation of this invention, when calculating the predicted temperature change rate for each temperature monitoring point based on the output of the dynamic temperature prediction model, the predicted temperature value at the next sampling time and the real-time temperature value at the current sampling time are extracted from the predicted temperature values ​​of each temperature monitoring point at multiple future sampling times output by the dynamic temperature prediction model. The predicted temperature value at the next sampling time is subtracted from the real-time temperature value at the current sampling time to obtain the temperature difference for that temperature monitoring point. The temperature difference is then divided by the duration of the sampling period to obtain the predicted temperature change rate for that temperature monitoring point. This calculation method directly uses the difference between the model's predicted value at the most recent future time and the current actual value to obtain the instantaneous rate of change, which can sensitively reflect the acceleration characteristics of temperature changes and provide timely and intuitive quantitative basis for power adjustment decisions.

[0014] When generating adaptive heating power adjustment commands, the predicted temperature change rate at each temperature monitoring point is compared with preset upper and lower thresholds for the rate of change. If the predicted temperature change rate at a certain temperature monitoring point is greater than the upper threshold, a power reduction command is generated for the area where that temperature monitoring point is located, carrying the target power reduction percentage. If the predicted temperature change rate at a certain temperature monitoring point is less than the lower threshold, a power increase command is generated for the area where that temperature monitoring point is located, carrying the target power increase percentage. All generated power adjustment commands are encapsulated into a command frame according to the temperature monitoring point number and sent to the solid-state relay module in the corresponding area in the form of a pulse width modulation signal through the output module of the supervisory control and data acquisition system. The solid-state relay module adjusts the heating power of the electric heating tape according to the duty cycle of the pulse width modulation signal. Preferably, the upper and lower thresholds for the rate of change are dynamically adjusted according to the ambient temperature change rate of the environment where the electric heating tape is located.

[0015] This control strategy proactively considers temperature change trends. Based on the direction and degree of deviation of the predicted rate of change from the threshold, it generates a power adjustment amount with a clear proportional relationship in advance. This is then achieved by directly controlling a solid-state relay via a PWM signal to realize stepless or near-stepless power adjustment, replacing traditional on / off or simple step-by-step control. The adjustment process is smooth and the response is rapid. The design of dynamically adjusting the threshold according to the rate of change of ambient temperature allows the control system to automatically adapt to the dynamic response differences of the heated object under different climates or process conditions, avoiding adjustment lag or over-adjustment caused by a fixed threshold when the environment changes drastically.

[0016] As a preferred embodiment of the present invention, the training process of the conditional variational autoencoder includes: extracting the original temperature time series of all temperature monitoring points at multiple historical sampling times from the historical database of the supervisory control and data acquisition system; performing time series decomposition operation on the original temperature time series at each historical sampling time to obtain the residual component vector corresponding to each historical sampling time; forming a training dataset from the residual component vectors of all historical sampling times; using each residual component vector in the training dataset as the input condition of the conditional variational autoencoder in sequence; the encoder network outputs the mean vector and log-variance vector corresponding to the input condition; sampling a random noise vector from the standard normal distribution; multiplying the random noise vector by the standard deviation obtained by the exponentiation of the log-variance vector and adding the mean vector to obtain the latent feature vector; the decoder network concatenates the latent feature vector with the input condition and maps it to a reconstructed residual vector through multiple fully connected layers; calculating the mean square error between the reconstructed residual vector and the input condition; calculating the relative entropy between the latent feature vector and the standard normal distribution; using the weighted sum of the mean square error and the relative entropy as the loss function; and using an adaptive moment estimation optimizer to iteratively update the weight parameters of the encoder network and the decoder network until the loss function converges. By using large-scale historical data to train the conditional variational autoencoder offline, the encoder network learns to extract the most expressive latent features from the residual vector, while the decoder network has the ability to recover residual details from the latent features and conditions. After training convergence, the model can generalize to various unknown residual patterns applicable to the online inference stage.

[0017] This invention also provides a data analysis-based intelligent temperature control system for electric heating cables, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned data analysis-based intelligent temperature control method for electric heating cables. The system integrates data acquisition, time-series decomposition, variational autoencoder inference, dynamic model prediction, rate of change calculation and instruction generation, and power regulation into a unified automatic control process. This enables distributed, refined temperature control across all areas along the electric heating cable, maintaining high uniformity and stability of the temperature field while reducing energy consumption.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0019] The residual components obtained from the temperature time series decomposition are input into a conditional variational autoencoder. Using the residual vectors of all temperature monitoring points at the current time as conditions, the encoder network maps these residuals to Gaussian distribution parameters and samples them to generate latent feature vectors. The decoder then reconstructs the residuals by concatenating the latent feature vectors with the conditions. During training, by minimizing the weighted sum of reconstruction error and relative entropy, the latent feature vectors are forced to approximate a standard normal distribution while preserving the multi-point perturbation correlation patterns implicit in the residuals. This latent feature vector is then concatenated element-wise with the temperature trend components to form a fused feature vector, which is input into a time convolutional network based on gated recurrent units. Dilated convolutions capture long-range dependencies between multiple sampling periods and output predicted temperature values ​​for multiple future times. This approach no longer discards the residuals as useless noise but extracts the latent structural information reflecting the influence of common environmental perturbations on each monitoring point. This enables the prediction model to infer random perturbations based on trend extrapolation, maintaining stable prediction accuracy even under non-stationary conditions such as gusts and rainfall. After calculating the predicted temperature change rate for each temperature monitoring point, it is compared with preset upper and lower thresholds. When the change rate exceeds the upper limit, a power reduction command is generated; when it falls below the lower limit, a power increase command is generated. All commands are encapsulated into command frames according to the monitoring point number and sent to the corresponding solid-state relay module in the form of a pulse width modulation signal through the output module. The upper and lower thresholds are dynamically adjusted according to the ambient temperature change rate, making the sensitivity of triggering power adjustment adaptively change with the amplitude of environmental fluctuations. Power adjustment is directly based on the predicted temperature change rate, advancing the triggering time of the control response from after the temperature deviation occurs to the beginning of the temperature change trend. By matching the temperature change rate under different environmental conditions with dynamic thresholds, it avoids slow response or over-adjustment during stable phases when the fixed threshold is used, reducing temperature overshoot and oscillation amplitude. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a flowchart of an intelligent temperature control method for electric heating cables based on data analysis.

[0022] Figure 2 This is a flowchart of an intelligent temperature control method for electric heating cables based on data analysis.

[0023] Figure 3 This is a schematic diagram of a temperature prediction model based on a conditional variational autoencoder and a gated recurrent unit temporal convolutional network.

[0024] Figure 4 This is a flowchart of the synchronous latching and sequential reading of temperature data from electric heating tape;

[0025] Figure 5 This is a flowchart of the iterative decomposition of temperature time series;

[0026] Figure 6 This is a comparison chart of the residual vector and its reconstruction result;

[0027] Figure 7 This is a distribution map of the predicted temperature change rate and upper and lower limit thresholds of the temperature monitoring points of the electric heating cable;

[0028] Figure 8 This is a schematic diagram illustrating the peak and trough identification of the original temperature time series. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] See Figure 1 This invention provides a data analysis-based intelligent temperature control method for electric heating cables. The method acquires real-time and historical temperature data from multiple temperature monitoring points along the electric heating cable using a supervisory control and data acquisition system. It performs time-series decomposition on the acquired real-time and historical temperature data, extracting temperature trend components, periodic components, and residual components from the decomposition results. The residual components are input into a pre-trained conditional variational autoencoder (CVA), which generates a latent feature vector based on the input residual components. The generated latent feature vector is then fused with the temperature trend components, and a dynamic temperature prediction model is constructed based on the fused features. This dynamic temperature prediction model is used to predict the future temperature state of each temperature monitoring point along the electric heating cable. Based on the output of the dynamic temperature prediction model, the predicted temperature change rate for each temperature monitoring point per unit time is calculated. When the predicted temperature change rate for any temperature monitoring point exceeds a pre-set threshold, an adaptive heating power adjustment command for the area where that temperature monitoring point is located is generated and sent to the solid-state relay module of the electric heating cable to adjust the heating power of the corresponding area.

[0031] Example 1:

[0032] In specific implementation, please refer to Figure 2The data analysis-based intelligent temperature control method for electric heating cables includes acquiring real-time and historical temperature data from multiple temperature monitoring points along the cable. Multiple thermocouple sensors are evenly spaced on the surface of the heating cable, each serving as a temperature monitoring point. K-type thermocouples are used. The signal output of each thermocouple sensor is connected to the data acquisition server of the monitoring and control system via an analog input module. A cold junction compensation circuit is connected in series between the thermocouple sensor and the analog input module to compensate for the cold junction temperature of the thermocouple sensor. The monitoring and control system polls the temperature values ​​reported by each thermocouple sensor according to a preset sampling period. During the polling process, the data acquisition server sends a broadcast synchronization pulse to all analog input modules. Each analog input module, upon receiving the broadcast synchronization pulse, synchronously latches the current temperature value of its connected thermocouple sensor. The data acquisition server reads the temperature values ​​latched by each analog input module in ascending order of temperature monitoring point numbers. After all temperature values ​​from all analog input modules have been read, the data acquisition server sends a broadcast synchronization pulse for the next sampling cycle to all analog input modules. The data acquisition server writes the mapping relationship between all the temperature values ​​read each time, the sampling cycle number of the current sampling cycle, and the numbers of each temperature monitoring point into the real-time database of the monitoring and control and data acquisition system.

[0033] When acquiring historical temperature data, all historical temperature data for each temperature monitoring point within multiple consecutive sampling periods prior to the current sampling time are extracted from the historical database of the monitoring and control and data acquisition system. Linear interpolation is used to fill in missing values ​​in the extracted historical temperature data. When a temperature value is missing at a certain sampling time, the two sampling times with the closest valid temperature values ​​before and after the missing time are determined. A linear interpolation is calculated based on the temperature values ​​of these two sampling times and the time interval, and the result of the linear interpolation is used as the temperature filler value for the missing sampling time. Outliers in the historical temperature data are removed based on the Laida criterion. The arithmetic mean and standard deviation of all historical temperature data are calculated. Historical temperature data with values ​​exceeding the range of the arithmetic mean plus or minus three times the standard deviation are identified as outliers and removed, thus obtaining the cleaned historical temperature data.

[0034] After obtaining real-time temperature data and cleaned historical temperature data, a time-series decomposition operation is performed for each temperature monitoring point. For each temperature monitoring point, the real-time temperature data at the current sampling time and the corresponding cleaned historical temperature data are concatenated in chronological order of sampling time to form the original temperature time series of the monitoring point. The original temperature time series is then iteratively decomposed using a locally weighted regression scatter smoothing method. In each iteration, adjacent peaks and troughs in the original temperature time series are identified through extreme value detection, the time interval between peaks and troughs is determined, and the median of all time intervals is calculated. This median is used as the period length for this iteration. Based on the determined period length, the original temperature time series is divided into multiple periodic subsequences, each with a length equal to the determined period length. Locally weighted regression smoothing is performed on each periodic subsequence, and the smoothed periodic subsequences are concatenated in chronological order within the original temperature time series to obtain the estimated periodic component sequence for this iteration. Subtracting the estimated periodic component from the original temperature time series yields the deperiodic series. Local weighted regression smoothing is then applied to the deperiodic series to obtain the estimated temperature trend component for this iteration. Subtracting both the estimated temperature trend component and the estimated periodic component from the original temperature time series yields the estimated residual component for this iteration.

[0035] After each iteration, the root mean square error (RMSE) is calculated between the residual component estimates obtained in the current iteration and those obtained in the previous iteration. In the first iteration, the residual component estimates from the previous iteration are set to a sequence with all values ​​equal to zero. The formula for calculating the RMSE is:

[0036]

[0037] in, This represents the total number of sampling times contained in the original temperature time series. Represents the first time series of the original temperature. Each sampling time, Indicates the first Sampling time in the next iteration The corresponding residual component estimates, Indicates the first Sampling time in the next iteration The corresponding residual component estimates, when hour When the calculation is obtained When the error falls below the preset convergence threshold of 0.001, the iteration stops, and the final time-series decomposition result of the original temperature time series is taken from the estimated values ​​of the temperature trend component, periodic component, and residual component output from the last iteration. The preset convergence threshold of 0.001 is determined to avoid excessive iterative calculations while ensuring the stability of the decomposed components. Decomposition experiments were conducted on multiple typical temperature time series, and the shape of the trend and periodic components no longer changed significantly after the root mean square error dropped below 0.001.

[0038] Example 2:

[0039] In specific implementation, please refer to Figure 3 The residual components of all temperature monitoring points at the current sampling time are arranged into a residual vector according to the temperature monitoring point numbers. This residual vector is then used as the input condition for a pre-trained conditional variational autoencoder (CVA). The CVA consists of an encoder network and a decoder network. The encoder network receives the input residual vector as conditions and maps it to a Gaussian distribution with a mean vector and a log-variance vector. The encoder network contains an input layer, three fully connected hidden layers, and two parallel output layers. The number of neurons in the input layer is equal to the dimension of the residual vector, which is equal to the total number of temperature monitoring points. The first fully connected hidden layer contains 128 neurons and uses a linear rectified activation function (LRC); the second fully connected hidden layer contains 64 neurons and uses an LRC; and the third fully connected hidden layer contains 32 neurons and uses an LRC. The output of the third fully connected hidden layer is connected to the mean output layer and the log-variance output layer, respectively. The number of neurons in both the mean and log-variance output layers is equal to the preset dimension of the latent feature vector, which is set to 16 dimensions. The mean output layer outputs the mean vector, the log-variance output layer outputs the log-variance vector, and the log-variance output layer has no activation function set, as does the mean output layer.

[0040] The latent feature vector is obtained by sampling from a Gaussian distribution defined by the mean vector and the log-variance vector. The sampling process involves sampling a random noise vector from a standard normal distribution, multiplying the random noise vector element-wise by the standard deviation obtained after exponential operation on the log-variance vector, and then adding the product element-wise to the mean vector to obtain the latent feature vector. The decoder network concatenates the latent feature vector with the residual vector used as input conditions. The concatenated vector is then mapped by the decoder network to reconstruct the residual vector. The decoder network consists of an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the input layer equals the sum of the dimensions of the latent feature vector and the residual vector. The first fully connected hidden layer contains 32 neurons and uses a linear rectified activation function; the second fully connected hidden layer contains 64 neurons and uses a linear rectified activation function; the third fully connected hidden layer contains 128 neurons and uses a linear rectified activation function. The number of neurons in the output layer equals the dimension of the residual vector, and the output layer uses a linear activation function to output the reconstructed residual vector.

[0041] The training process of the conditional variational autoencoder is as follows: The original temperature time series of all temperature monitoring points at multiple historical sampling times are extracted from the historical database of the supervised control and data acquisition system. A time-series decomposition operation is performed on the original temperature time series at each historical sampling time to obtain the residual component vector corresponding to each historical sampling time. The residual component vectors of all historical sampling times are then combined to form the training dataset. The total number of residual component vectors in the training dataset is denoted as . Each residual component vector in the training dataset is used sequentially as the input condition of the conditional variational autoencoder, and the encoder network outputs the mean vector corresponding to that input condition. Sum of logarithmic variance vector ,in, For training sample index, Sampling a random noise vector from a standard normal distribution. , The dimension of is the same as the dimension of the latent feature vector. Each component independently follows a normal distribution with a mean of 0 and a variance of 1. Calculate the latent eigenvector. :

[0042]

[0043] in, Represented by natural constant Exponential operations with base 0. This represents element-wise multiplication. This represents the log-variance vector. The vector is obtained by dividing each component by 2. The decoder network then divides the latent feature vectors... After being concatenated with the input conditions, i.e., the residual component vector, it is mapped to the reconstructed residual vector through a fully connected layer. Calculate the reconstructed residual vector. With input conditions Mean square error between :

[0044]

[0045] in, The dimension of the residual vector represents the total number of temperature monitoring points. Represents the reconstructed residual vector The One component; Indicates input conditions The Each component.

[0046] Calculate latent feature vectors Relative entropy with the standard normal distribution :

[0047]

[0048] in, The dimension of the latent feature vector. ; Represents the log-variance vector The One component; Represents the mean vector The Each component. Loss function This is a weighted sum of mean squared error and relative entropy:

[0049]

[0050] in, As a weighting factor, The value of is set to 0.002, which is determined by selecting a value from the candidate set {0.001, 0.002, 0.005, 0.01} using grid search on the validation set that balances the reconstruction error of the validation set with the generation quality. An adaptive moment estimation optimizer is used to iteratively update the weight parameters of the encoder and decoder networks. The initial learning rate is set to 0.001, the batch size is set to 64, and the iteration continues until the loss function is realized. The decrease over 10 consecutive training rounds was less than Training should be stopped at this time.

[0051] After obtaining the latent feature vectors, for each temperature monitoring point, the latent feature vector of the temperature monitoring point is concatenated element-wise with the temperature trend component value of the temperature monitoring point at the current sampling time to form a fused feature vector of the temperature monitoring point. The fused feature vectors of all temperature monitoring points are stacked in order of temperature monitoring point number to form a fused feature matrix. The number of rows in the fused feature matrix is ​​equal to the number of temperature monitoring points, and the number of columns is equal to the sum of the dimension of the latent feature vector and 1.

[0052] The fused feature matrix is ​​input into a temporal convolutional network based on gated recurrent units (ROUs). This network comprises four dilated convolutional layers with dilation coefficients of 1, 2, 4, and 8, increasing exponentially with the number of layers. Each dilated convolutional layer uses 64 kernels of size 3 with a stride of 1, employing causal convolution and padding the input sequence with zeros to maintain the temporal length. The output of each dilated convolutional layer is processed by a gated recurrent unit (ROU), whose hidden state dimension is set to 64. The four dilated convolutional layers and the RROs are stacked alternately, with the hidden state sequence output from the previous RRO serving as the input to the next dilated convolutional layer. The hidden state sequence output from the last RRO serves as the internal state of the dynamic temperature prediction model. The dynamic temperature prediction model takes the fused feature matrix at the current sampling time as input and outputs the predicted temperature value for each temperature monitoring point at multiple future sampling times. The output of the dynamic temperature prediction model is achieved through a fully connected layer, which maps the hidden state sequence of the last gated recurrent unit to the temperature prediction values ​​at multiple future sampling times. The number of output neurons is equal to the product of the number of temperature monitoring points and the number of future prediction sampling times.

[0053] See Figure 6 In the figure, the horizontal axis represents the temperature monitoring point number, covering multiple temperature monitoring points deployed along the electric heating cable, with numbers increasing from 1 to approximately 500 points. The vertical axis represents the residual value of the corresponding temperature monitoring point, in degrees Celsius. The solid curve in the figure represents the residual vector input to the conditional variational autoencoder, while the dashed curve with square markings represents the residual vector reconstructed based on the conditional variational autoencoder. The two curves highly overlap in overall trend, with residual values ​​mainly distributed in the range of -1.0℃ to 1.0℃, indicating a small error between the input residual vector and the reconstructed residual vector, and a good reconstruction effect. The fluctuation of the residual values ​​shows no obvious trend, exhibiting a relatively random distribution characteristic, which conforms to the definition and processing requirements of the residual components in Example 2. This figure reflects the effective encoding and decoding capability of the conditional variational autoencoder for the input residual vector, providing an accurate feature expression basis for the subsequent generation and fusion of potential feature vectors, and ensuring the input data quality of the dynamic temperature prediction model.

[0054] Example 3:

[0055] In practical implementation, the predicted temperature value for each temperature monitoring point at the next sampling time is extracted from the predicted temperature values ​​of each temperature monitoring point at multiple future sampling times output by the dynamic temperature prediction model. Simultaneously, the real-time temperature value of the same temperature monitoring point at the current sampling time is obtained. The temperature difference between the temperature monitoring points is obtained by subtracting the real-time temperature value at the current sampling time from the predicted temperature value at the next sampling time. The quotient obtained by dividing the temperature difference by the sampling period is the predicted temperature change rate of the temperature monitoring point. The formula for calculating the predicted temperature change rate is:

[0056]

[0057] in, Indicates the number is The predicted temperature change rate at the temperature monitoring points, expressed in degrees Celsius per second; Indicates the number is The predicted temperature value of the temperature monitoring point at the next sampling time, in degrees Celsius; Indicates the number is The real-time temperature value of the temperature monitoring point at the current sampling time, in degrees Celsius; This indicates the duration of the sampling period, in seconds. The sampling period is determined by the preset sampling period of the monitoring, control and data acquisition system, which is set to 5 seconds.

[0058] The predicted temperature change rate for each temperature monitoring point is calculated and compared with preset upper and lower thresholds for the rate of change. These thresholds are dynamically adjusted based on the ambient temperature change rate of the environment where the electric heating cable is located. The ambient temperature change rate is collected by an ambient temperature sensor deployed in the environment where the electric heating cable is located. This sensor is connected to the data acquisition server of the monitoring, control, and data acquisition system via an analog input module. The monitoring, control, and data acquisition system synchronously collects the temperature values ​​from the ambient temperature sensor in each sampling period and calculates the absolute value of the difference between the current and previous sampling times, divided by the length of the sampling period, to obtain the ambient temperature change rate. The upper threshold is obtained by multiplying the ambient temperature change rate by an upper limit scaling factor, set to 1.5, meaning the upper threshold is 1.5 times the ambient temperature change rate. The lower threshold is obtained by multiplying the ambient temperature change rate by a lower limit scaling factor, set to 0.5, meaning the lower threshold is 0.5 times the ambient temperature change rate. The upper limit scaling factor of 1.5 and the lower limit scaling factor of 0.5 are determined based on the fact that during the actual operation of the electric heating tape, the rate of temperature change in the heating area is allowed to fluctuate within the range of 0.5 to 1.5 times the rate of temperature change in the ambient temperature. Exceeding this range indicates that there is a tendency for insufficient heating or overheating.

[0059] During the comparison process, if the predicted temperature change rate at a certain temperature monitoring point exceeds the upper limit threshold, a power reduction instruction is generated for the area where the temperature monitoring point is located. This instruction carries the target power reduction percentage. The target power reduction percentage is determined linearly based on the magnitude by which the predicted temperature change rate exceeds the upper limit threshold. Specifically, the difference between the predicted temperature change rate and the upper limit threshold is divided by the upper limit threshold, and then multiplied by a reduction ratio coefficient set to 10. The resulting product is the target power reduction percentage, with an upper limit truncation of 50%. The reduction ratio coefficient of 10 is set because when the predicted temperature change rate exceeds the upper limit threshold by 10% of the threshold value, the target power reduction percentage reaches 100%. To avoid over-adjustment, the upper limit truncation limits the single power reduction to within 50%.

[0060] If the predicted temperature change rate at a certain temperature monitoring point is less than the lower limit threshold, a power increase command is generated for the area where the temperature monitoring point is located. This command carries the target power increase percentage. The target power increase percentage is determined linearly based on the extent to which the predicted temperature change rate is lower than the lower limit threshold. Specifically, the difference between the lower limit threshold and the predicted temperature change rate is divided by the lower limit threshold, and then multiplied by the increase ratio coefficient. The increase ratio coefficient is set to 10. The resulting product is the target power increase percentage, with an upper limit truncation of 50%. The basis for setting the increase ratio coefficient of 10 is the same as that for setting the decrease ratio coefficient of 10, limiting the single power increase to within 50% through the upper limit truncation.

[0061] All power adjustment commands generated based on each temperature monitoring point are encapsulated into a command frame in the order of the temperature monitoring point numbers. The command frame contains a frame header, the number of temperature monitoring points, a command entry sequence, and a checksum. The frame header is a fixed byte sequence 0xA50x5A, used to identify the start of the command frame. The number of temperature monitoring points occupies one byte, indicating the total number of power adjustment command entries contained in the command frame. Each command entry in the command entry sequence corresponds to a temperature monitoring point that generated the power adjustment command. The command entry contains the temperature monitoring point number, an adjustment direction flag, and a target power adjustment percentage. The temperature monitoring point number occupies two bytes, the adjustment direction flag occupies one bit (0x00 indicates power reduction, 0x01 indicates power increase), and the target power adjustment percentage occupies one byte, representing an integer value indicating a percentage decrease or increase in target power. The checksum is the byte-by-byte XOR result of all bytes from the frame header to the end of the command entry sequence, occupying one byte.

[0062] The encapsulated instruction frame is sent to the corresponding solid-state relay module in the form of a pulse-width modulation (PWM) signal through the output module of the supervisory control and data acquisition system. The output module converts each byte in the instruction frame into 8 bits of serial data, with each bit corresponding to one PWM cycle. When a bit is 1, the duty cycle of the PWM signal is 80%; when a bit is 0, the duty cycle is 20%. The frequency of the PWM cycle is set to 1 kHz. The solid-state relay module reconstructs the instruction frame by parsing the received PWM signal. Based on the instruction entry in the instruction frame corresponding to the temperature monitoring point number in its controlled area, it adjusts the heating power of the electric heating tape in the corresponding area according to the target power adjustment percentage. The solid-state relay module adjusts the heating power by changing the AC conduction angle. The decrease in conduction angle corresponding to the percentage decrease in target power is linearly related to the percentage decrease in target power, and the increase in conduction angle corresponding to the percentage increase in target power is linearly related to the percentage increase in target power.

[0063] See Figure 7 In the graph, the vertical axis represents the predicted rate of temperature change in degrees Celsius per second (°C / s), and the horizontal axis represents the temperature monitoring point number, ranging from 1 to approximately 500. Each black dot in the scatter plot represents the predicted rate of temperature change for the corresponding monitoring point at the next sampling time. The dashed line represents the upper threshold of the rate of change, approximately 0.3°C / s, and the dotted line represents the lower threshold, approximately 0.1°C / s.

[0064] As shown in the figure, the predicted temperature change rate of most temperature monitoring points falls between the upper and lower thresholds, indicating that the temperature change rate at each monitoring point of the electric heating cable is within the allowable fluctuation range. A small number of temperature monitoring points have predicted temperature change rates exceeding the upper threshold of 0.3℃ / s, suggesting potential overheating in the corresponding area, requiring a power reduction command. Conversely, some temperature monitoring points have predicted temperature change rates below the lower threshold of 0.1℃ / s, indicating insufficient heating, requiring a power increase command. The overall distribution of the scattered points shows no obvious concentration trend and is relatively uniform, reflecting the spatial differences in temperature change rates along the monitoring line. The data in this figure directly corresponds to the temperature change rate calculation and threshold comparison steps in Example 3, providing a basis for adaptively adjusting the heating power of the electric heating cable.

[0065] Example 4:

[0066] In specific implementation, please refer to Figure 4 When the monitoring, control, and data acquisition system acquires temperature data from multiple temperature monitoring points along the electric heating cable, it employs a synchronous latching and sequential reading mechanism. The data acquisition server of the monitoring, control, and data acquisition system generates a broadcast synchronization pulse signal. This broadcast synchronization pulse signal is a negative logic square wave pulse with a high-level duration of 100 microseconds, a low-level duration of 10 microseconds, and a pulse amplitude of 5 volts. The data acquisition server simultaneously sends the broadcast synchronization pulse signal to all analog input modules via a dedicated synchronization bus, which is a shielded twisted-pair cable connected to the synchronization trigger input terminal of each analog input module.

[0067] Upon receiving the falling edge of the broadcast synchronization pulse signal, each analog input module triggers its internal multi-channel sample-and-hold circuit. This circuit simultaneously samples the analog voltage signals output from the cold junction compensation circuit of each thermocouple sensor connected to the module and holds the sampled values ​​in the holding capacitor at the back end, thus achieving synchronous latching of the current temperature values ​​of all thermocouple sensors. The sample-and-hold circuit of the analog input module uses an LF398 sample-and-hold amplifier, and its holding capacitor is a 1000 picofarad polystyrene film capacitor to ensure that the holding voltage drop does not exceed 0.1% during reading.

[0068] After the data acquisition server completes the broadcast synchronization pulse signal transmission and waits for a preset stabilization period of 20 microseconds, it sequentially sends read command frames to each analog input module via the RS-485 communication bus, following the ascending order of temperature monitoring point numbers. The read command frame contains the address number of the target analog input module, the starting channel number, and the number of channels to be read. There is a mapping relationship between the address number of the analog input module and the range of temperature monitoring point numbers it is connected to. Each analog input module connects to 8 thermocouple sensors, with address numbers starting from 0x01 and ascending sequentially. After receiving and verifying the read command frame, the target analog input module uses its internal analog-to-digital converter to sequentially convert the analog voltage values ​​of each channel latched on the holding capacitor into 16-bit digital temperature values. The analog-to-digital converter uses an AD7606 successive approximation type, with a conversion accuracy of 16 bits and a conversion time of no more than 4 microseconds per channel. The analog input module encapsulates the converted digital temperature values ​​into a response data frame according to the temperature monitoring point numbering order. The response data frame contains the analog input module address, temperature data sequence, and cyclic redundancy check code, and is sent back to the data acquisition server via the RS-485 communication bus.

[0069] After successfully receiving a response data frame from an analog input module, the data acquisition server sends a read command frame to the next analog input module at the next address after a 10-microsecond interval, until the temperature values ​​latched by all analog input modules have been read. After all analog input modules have been read, the data acquisition server sends a broadcast synchronization pulse signal for the next sampling period to all analog input modules to start the next round of synchronization latching.

[0070] The data acquisition server combines all temperature values ​​read from each analog input module within the current sampling period with the mapping relationship between the current sampling period number and the temperature monitoring point number. The sampling period number is a 64-bit unsigned integer, initially set to 1, and increments by 1 after each complete polling and writing to the real-time database. The temperature monitoring point number is an integer starting from 1 and incrementing, consistent with the physical placement order of the thermocouple sensors on the electric heating tape. The data acquisition server assembles the sampling period number, temperature monitoring point number, and corresponding temperature value into a record row in the format "Sampling Period Number, Temperature Monitoring Point Number, Temperature Value". Multiple record rows are appended to the real-time database of the monitoring and control and data acquisition system. The real-time database uses a memory-mapped file storage engine to achieve a write throughput of over 5000 records per second.

[0071] Example 5:

[0072] In specific implementation, please refer to Figure 5When iteratively decomposing the original temperature time series using the local weighted regression scatter smoothing method, a single iteration is implemented as follows: Identify the time intervals between all adjacent peaks and troughs in the original temperature time series. This is done by calculating the first-order difference sequence of the original temperature time series, iterating through each element in the first-order difference sequence, and marking the sampling point in the original temperature time series corresponding to the element's position as a peak when the value of an element is greater than zero and the value of the next element is less than zero; and marking the sampling point in the original temperature time series corresponding to the element's position as a trough when the value of an element is less than zero and the value of the next element is greater than zero. Record the sampling time indices of the marked peaks and troughs in the original temperature time series, and arrange the peaks and troughs alternately in ascending order of sampling time indices. Calculate the absolute value of the difference between the sampling time indices of each pair of adjacent peaks and troughs. The set of absolute differences obtained is the set of time intervals between adjacent peaks and troughs. Take the median of the time interval set. If the median is not an integer, round it down. Use the rounded value as the period length of the current iteration.

[0073] The original temperature time series is divided into multiple periodic subsequences according to the current iteration period length. During the division, starting from the initial sampling time of the original temperature time series, every period length of sampling points is divided into a periodic subsequence. If the number of remaining sampling points at the end is less than one period length, the remaining sampling points are treated as a separate periodic subsequence.

[0074] Local weighted regression smoothing is performed on each periodic subsequence. During local weighted regression smoothing, for each sampling point in the periodic subsequence to be smoothed, a window width is set with the sampling point's time index as the center. The window width is set to the smaller of half the length of the periodic subsequence and 15. For each neighboring point within the window width, the time index difference between the neighboring point and the center sampling point is calculated, and the time index difference is divided by the window width to obtain the normalized distance. A cubic kernel function is used to calculate the weights of the neighboring points. The form of the cubic kernel function is that when the absolute value of the normalized distance is less than 1, the weight is equal to 1 minus the cube of the absolute value of the normalized distance; when the absolute value of the normalized distance is greater than or equal to 1, the weight is 0. Weighted least squares linear regression is performed on the neighboring points within the window width using the calculated weights. The predicted value of the fitted line at the center sampling point is used as the smoothed value of the center sampling point. This smoothing operation is performed on all sampling points in the periodic subsequence to obtain the smoothed periodic subsequence. All smoothed periodic subsequences are concatenated in chronological order according to their place in the original temperature time series to form the estimated sequence of periodic components for the current iteration.

[0075] The deperiodic sequence is obtained by subtracting the estimated periodic component from the original temperature time series point by point. The deperiodic sequence is then subjected to local weighted regression smoothing, with the smoothing window width set to 0.2 times the length of the deperiodic sequence, and set to 5 when the product is less than 5. Similarly, a cubic kernel function is used to calculate the weights, and weighted least squares linear regression is used to obtain the smoothed value. This smoothed sequence yields the estimated temperature trend component for the current iteration. Finally, the estimated residual component is obtained by subtracting the estimated temperature trend component and the estimated periodic component from the original temperature time series point by point.

[0076] Calculate the root mean square error of the residual component estimates between the current iteration and the previous iteration. Let the total number of sampling times in the original temperature time series be... The sampling time index is , Let the current iteration be denoted as the nth iteration. iteration , No. The sampling time in the residual component estimate sequence obtained in the next iteration The corresponding value is denoted as , No. The sampling time in the residual component estimate sequence obtained in the next iteration The corresponding value is denoted as ,when hour, The value is 0. Root mean square error The calculation formula is:

[0077]

[0078] in, This represents the total number of sampling times contained in the original temperature time series; For sampling time index; For the first The sequence of estimated residual components in the nth iteration The value at each sampling time; For the first The sequence of estimated residual components in the nth iteration The value at each sampling time.

[0079] The calculated The convergence threshold is compared to a preset convergence threshold of 0.001. When... When the value is less than 0.001, the iteration process terminates, and the temperature trend component estimate sequence, periodic component estimate sequence, and residual component estimate sequence obtained in the current iteration are used as the final decomposition result; when When the value is greater than or equal to 0.001, the estimated sequence of residual components obtained in the current iteration is used as the input sequence for the next iteration, and the process of identifying peaks and troughs from the original temperature time series and subsequent steps is returned to start a new iteration.

[0080] A data analysis-based intelligent temperature control system for electric heating cables includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements all the steps included in the aforementioned data analysis-based intelligent temperature control method for electric heating cables.

[0081] See Figure 8 In the figure, the horizontal axis represents the sampling time number, ranging from 0 to approximately 500, and the vertical axis represents the temperature in degrees Celsius, ranging from approximately 22.5℃ to 38.0℃. The solid curve in the figure represents the original temperature time series of a certain temperature monitoring point, and the curve shows obvious periodic fluctuations and an overall upward trend. The periodic temperature values ​​vary between approximately 22.5℃ and 38.0℃, reflecting the periodic characteristics of temperature rising and falling repeatedly over time. The overall temperature trend shows a slow increase from left to right, rising from approximately 25℃ initially to approximately 37℃ at the end. The peaks and troughs in the temperature series are marked with triangles in the figure. Peaks and troughs appear alternately with relatively uniform intervals, meeting the requirement for identifying the period length in time series decomposition.

[0082] The peak and trough markers identified by the first-order difference method clearly demonstrate the distribution of extreme points used in the iterative decomposition of the local weighted regression scatter smoothing method. In this embodiment, the median of the difference between the sampling time indices of adjacent peaks and troughs is used as the period length. The intervals between peaks and troughs in the figure reflect stable periodic characteristics. This period length forms the basis for subsequent segmentation of the original temperature time series into multiple periodic subsequences.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A data analysis-based intelligent temperature control method for electric heating cables, characterized in that, include: Real-time and historical temperature data from multiple temperature monitoring points along the electric heating cable are obtained through a monitoring, control and data acquisition system. The real-time temperature data and historical temperature data are decomposed over time to extract temperature trend components, periodic components and residual components. The residual components are input into a pre-trained conditional variational autoencoder to generate latent feature vectors; The latent feature vectors are fused with temperature trend components to construct a dynamic temperature prediction model. Based on the output of the dynamic temperature prediction model, calculate the predicted temperature change rate for each temperature monitoring point. When the predicted rate of temperature change exceeds the set threshold, an adaptive heating power adjustment command is generated and sent to the solid-state relay module of the electric heating tape.

2. The intelligent temperature control method for electric heating tape based on data analysis according to claim 1, characterized in that, The steps of acquiring real-time and historical temperature data from multiple temperature monitoring points along the electric heating cable through a monitoring, control, and data acquisition system include: Multiple thermocouple sensors are arranged at equal intervals on the surface of the electric heating tape as temperature monitoring points. The signal output terminal of each thermocouple sensor is connected to the data acquisition server of the monitoring control and data acquisition system through an analog input module. The monitoring, control and data acquisition system polls the temperature values ​​reported by each thermocouple sensor according to a preset sampling period, and stores the temperature values ​​collected at each sampling time, along with the corresponding temperature monitoring point number and the timestamp of the sampling time, into the real-time database of the monitoring, control and data acquisition system. The historical temperature data of each temperature monitoring point within multiple consecutive sampling periods before the current sampling time are extracted from the historical database of the monitoring and control and data acquisition system. The historical temperature data is then subjected to linear interpolation to fill in missing values ​​and outliers are removed based on the Laida criterion to obtain cleaned historical temperature data.

3. The intelligent temperature control method for electric heating tape based on data analysis according to claim 2, characterized in that, The thermocouple sensor is a type K thermocouple, and each thermocouple sensor is connected in series with the analog input module via a cold junction compensation circuit.

4. The intelligent temperature control method for electric heating tape based on data analysis according to claim 2, characterized in that, The steps of performing time-series decomposition on the real-time temperature data and historical temperature data to extract temperature trend components, periodic components, and residual components include: The real-time temperature data of each temperature monitoring point at the current sampling time is spliced ​​with the cleaned historical temperature data to form the original temperature time series of that temperature monitoring point; The original temperature time series is iteratively decomposed using the local weighted regression scatter smoothing method. In each iteration, the periodic component is first extracted from the original temperature time series, then the temperature trend component is extracted from the remaining series, and finally the temperature trend component and periodic component are subtracted from the original temperature time series to obtain the residual component. Perform a decomposition operation on each temperature monitoring point to obtain the temperature trend component sequence, periodic component sequence, and residual component sequence corresponding to each temperature monitoring point.

5. The intelligent temperature control method for electric heating tape based on data analysis according to claim 4, characterized in that, The step of inputting the residual components into a pre-trained conditional variational autoencoder to generate latent feature vectors includes: The residual components of all temperature monitoring points at the current sampling time are arranged into a residual vector according to the order of the temperature monitoring point numbers, and this residual vector is used as the input condition of the conditional variational autoencoder. The conditional variational autoencoder consists of an encoder network and a decoder network. The encoder network maps the residual vector to the mean vector and log-variance vector of a Gaussian distribution according to the input conditions, and samples the latent feature vector from the Gaussian distribution. The decoder network reconstructs the estimated residual vector by concatenating the latent feature vector with the input conditions. During training, it minimizes the relative entropy between the reconstruction error and the prior distribution of the latent feature vector.

6. The intelligent temperature control method for electric heating tape based on data analysis according to claim 5, characterized in that, The steps of fusing the latent feature vector with the temperature trend component to construct a dynamic temperature prediction model include: The latent feature vector of each temperature monitoring point is concatenated element by element with the value of the temperature trend component of that temperature monitoring point at the current sampling time to form the fused feature vector of that temperature monitoring point. The fused feature vectors of all temperature monitoring points are stacked into a fused feature matrix in the order of temperature monitoring point numbering. The fused feature matrix is ​​then input into a temporal convolutional network based on gated recurrent units. This temporal convolutional network performs dilation convolution operation on the fused feature matrix along the time dimension to capture long-range dependencies between multiple sampling periods. The hidden state sequence output by the temporal convolutional network is used as the internal state of the dynamic temperature prediction model. This dynamic temperature prediction model takes the fusion feature matrix at the current sampling time as input and outputs the predicted temperature value of each temperature monitoring point at multiple future sampling times.

7. The intelligent temperature control method for electric heating tape based on data analysis according to claim 6, characterized in that, The temporal convolutional network based on gated recurrent units includes multiple dilated convolutional layers, and the dilation coefficient of each dilated convolutional layer increases exponentially with the number of network layers.

8. The intelligent temperature control method for electric heating tape based on data analysis according to claim 6, characterized in that, The step of calculating the predicted temperature change rate for each temperature monitoring point based on the output of the dynamic temperature prediction model includes: From the predicted temperature values ​​of each temperature monitoring point at multiple future sampling times output by the dynamic temperature prediction model, extract the predicted temperature value of each temperature monitoring point at the next sampling time and the real-time temperature value at the current sampling time. The temperature difference of each temperature monitoring point is obtained by subtracting the predicted temperature value at the next sampling time from the real-time temperature value at the current sampling time. Divide the temperature difference by the sampling period to obtain the predicted temperature change rate at the monitoring point.

9. The intelligent temperature control method for electric heating tape based on data analysis according to claim 8, characterized in that, The step of generating an adaptive heating power adjustment command when the predicted temperature change rate exceeds a set threshold, and sending the command to the solid-state relay module of the electric heating tape, includes: The predicted temperature change rate at each temperature monitoring point is compared with the preset upper and lower thresholds for the rate of change. If the predicted temperature change rate at a certain temperature monitoring point is greater than the upper limit threshold of the change rate, a power reduction instruction is generated for the area where the temperature monitoring point is located. The power reduction instruction carries the target power reduction percentage. If the predicted temperature change rate at a certain temperature monitoring point is less than the lower limit threshold of the change rate, a power increase command is generated for the area where the temperature monitoring point is located. The power increase command carries the target power increase percentage. All generated power adjustment commands are encapsulated into a command frame according to the temperature monitoring point number, and sent to the solid-state relay module in the corresponding area in the form of a pulse width modulation signal through the output module of the supervisory control and data acquisition system. The solid-state relay module adjusts the heating power of the electric heating tape according to the duty cycle of the pulse width modulation signal.

10. A data-analysis-based intelligent temperature control system for electric heating cables, 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 steps of the intelligent temperature control method for electric heating cables based on data analysis as described in any one of claims 1 to 9.

Images