Intelligent temperature control system and method for control box electric heat tracing band

Abstract

The application relates to the technical field of electric heat tracing intelligent temperature control, and discloses an intelligent temperature control system and method for an electric heat tracing belt of a control box. The method comprises the following steps: collecting data of a plurality of temperature sensors in the control box, constructing a temperature time sequence, and forming a time sequence graph structure reflecting the correlation strength between the sensors by calculating spatial proximity and time correlation. The time sequence graph is input into a state space model, and the state transition matrix of the model is dynamically adjusted according to the edge weight value of the graph, so as to accurately simulate the dynamic process of heat transfer in the box. The output state of the model is separated into a periodic component and a trend component, the periodic component is matched with a reference mode library to identify an abnormal fluctuation phase, and the trend component is subjected to multi-scale decomposition and slope curvature analysis to quantify the change trend. The method realizes dynamic and accurate modeling of a complex heat environment, can more accurately control the temperature and give an early warning of an abnormality.

Description

Technical Field

[0001] This invention relates to the field of intelligent temperature control technology for electric heat tracing, specifically to an intelligent temperature control system and method for electric heat tracing cables in a control box. Background Technology

[0002] In outdoor industrial settings such as power and chemical industries, electric heat tracing insulation of control boxes is a crucial measure to prevent internal components from failing due to condensation or freezing at low temperatures. Existing intelligent temperature control methods mostly rely on multiple temperature sensors arranged inside the box to collect real-time temperature data at various points and adjust the heat tracing power based on preset thresholds or simple physical models. These methods treat each sensor as an independent monitoring point, primarily focusing on whether its instantaneous or short-term historical readings exceed safe limits.

[0003] Conventional technical solutions have limitations when dealing with objects like control boxes, which have limited space but complex thermal environments. Due to factors such as the internal structure of the control box, the layout of components, and external environmental wind speed, the generation, diffusion, and dissipation of heat is a spatially uneven and temporally dynamic process. Existing methods lack in-depth exploration of the inherent correlations between sensor readings and cannot characterize the heat transfer paths and mutual influences between different locations within the box. Furthermore, using state models or control strategies with fixed parameters is difficult to adapt to this dynamically changing spatiotemporal thermal process, leading to distortion in temperature field simulations and potentially causing localized overheating or underprotected areas. The accuracy and energy efficiency of temperature control need to be improved.

[0004] The purpose of this invention is to address how to more comprehensively and accurately perceive and simulate the complex thermal dynamics inside a control box, thereby achieving more refined temperature management. This requires overcoming the limitations of traditional discrete-point monitoring and static modeling, reconstructing the understanding of the temperature field from a spatiotemporal correlation perspective, and establishing a dynamic model capable of adapting to changes in this correlation. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent temperature control system and method for electric heating cables in control boxes, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent temperature control method for an electric heating cable in a control box, the method comprising:

[0007] The readings of multiple temperature sensors arranged inside the control box are collected and analyzed to construct a temperature time series with time dimension, spatial dimension and event tag dimension;

[0008] The spatial proximity calculation and temporal correlation calculation are performed on the temperature time series to construct a time series graph structure that reflects the spatiotemporal correlation strength between different sensor nodes.

[0009] The time sequence diagram structure is input into a state space model, and the state space model dynamically adjusts its state transition matrix according to the edge weights of the time sequence diagram structure to simulate and derive the potential dynamic process of heat transfer inside the control box.

[0010] From the state vector output by the state space model, components representing periodic temperature fluctuation patterns and components representing trend-based temperature change patterns are separated.

[0011] The periodic temperature fluctuation pattern component is matched and aligned point by point with a preset reference temperature fluctuation pattern library to identify and mark the abnormal fluctuation phase in the periodic temperature fluctuation pattern component.

[0012] Multiscale decomposition is performed on the trend temperature change pattern components to obtain trend sub-components at different time scales. Slope and curvature analyses are then performed on each trend sub-component to quantify its rate of change and direction.

[0013] Preferably, constructing a temperature time series sequence with time dimension, spatial dimension, and event tag dimension includes:

[0014] The readings from multiple temperature sensors are timestamped and aligned, and each reading is assigned a spatial location code according to a preset spatial coordinate mapping table.

[0015] In the temperature time series, mark the times of electric heating cable start-up and shutdown events and control box door opening and closing events, and attach the event type and duration as event marker dimensions to the data points of the corresponding time period;

[0016] The temperature time series after the marked event is filled with missing values ​​and outliers are removed to form a regular temperature-space-time matrix as the input basis for subsequent calculations.

[0017] Preferably, the step of constructing a time-series graph structure reflecting the spatiotemporal correlation strength between different sensor nodes includes:

[0018] Using each temperature sensor as a node, calculate the correlation coefficient of the temperature reading sequences of any two sensor nodes within the sliding time window, and use it as the initial correlation strength.

[0019] Based on the spatial location codes recorded in the temperature spatiotemporal matrix, the physical distance between sensor nodes is calculated, and the initial correlation strength is corrected using a distance decay function to obtain the spatial constraint correlation strength.

[0020] The spatial constraint correlation strength is fused with the temporal graph structure to form a weighted undirected graph with sensors as nodes and the fused spatiotemporal correlation strength as edge weights. The edge weights of the temporal graph structure are updated with the sliding time window.

[0021] Preferably, the step of simulating and deriving the potential dynamic process of heat transfer inside the control box includes:

[0022] The state-space model adaptively constructs the adjacency weight part of the state transition matrix based on the edge weights of the current time sequence graph structure.

[0023] The predicted state vector at the current moment is calculated by multiplying the state vector of the previous moment with the state transition matrix and combining it with an external input driven by historical ambient temperature data.

[0024] The predicted state vector is compared with the temperature spatiotemporal matrix fragment obtained from actual observations, the observation residual is calculated, and the internal parameters of the state-space model are corrected through a feedback mechanism using the observation residual to complete a one-step deduction of the potential dynamic process.

[0025] Preferably, separating the component characterizing the periodic temperature fluctuation pattern and the component characterizing the trend temperature change pattern includes:

[0026] Fourier spectrum analysis is performed on the state vector output by the state space model to extract frequency components with energy exceeding a preset threshold, and the signal corresponding to the frequency components is reconstructed into periodic temperature fluctuation mode components.

[0027] The periodic temperature fluctuation pattern component is subtracted from the original state vector to obtain a residual sequence. A local weighted regression method is applied to the residual sequence to fit a smooth trend line as the trend temperature change pattern component.

[0028] Preferably, the step of identifying and marking anomalous fluctuation phases in the periodic temperature fluctuation pattern components includes:

[0029] From the preset reference temperature fluctuation pattern library, call up a historical normal fluctuation pattern that matches the current season and time as a reference;

[0030] The difference between the periodic temperature fluctuation mode component and the reference mode at each sampling point is calculated. When the difference between multiple consecutive sampling points exceeds the dynamic threshold, the abnormal time period composed of the multiple consecutive abnormal sampling points is determined to be an abnormal fluctuation phase, and its start time, end time and average difference intensity are recorded.

[0031] Preferably, the slope and curvature analysis performed on each trend sub-component includes:

[0032] For each trend sub-component at each time scale, its first-order difference sequence is calculated as the instantaneous slope sequence, and its second-order difference sequence is calculated as the instantaneous curvature sequence.

[0033] Statistical features are extracted from the instantaneous slope sequence and instantaneous curvature sequence to obtain the slope and variance and curvature variance, which characterize the overall drastic change of each trend sub-component.

[0034] The slope and curvature features corresponding to the trend sub-components at different time scales are concatenated from coarse to fine time scales to form a multi-scale feature vector describing trend changes.

[0035] Preferably, it includes: inputting the labeling information of the abnormal fluctuation phase, the multi-scale feature vector, and the latest edge weights of the time series graph structure into a fusion decision network;

[0036] The fusion decision network includes a temporal attention subnetwork for processing labeled information, a fully connected subnetwork for processing multi-scale feature vectors, and a graph convolution subnetwork for processing temporal graph structures.

[0037] The fusion decision network performs feature-level fusion of the outputs of the three sub-networks and calculates the probability distribution of the control intensity that the electric tracing heat cable needs to be applied at the current moment through a probability output layer.

[0038] Preferably, the method includes: selecting the highest probability modulation intensity level according to the modulation intensity probability distribution, and querying a preset modulation intensity-duty cycle mapping table to obtain the target pulse width modulation duty cycle;

[0039] A corresponding pulse width modulation signal is generated based on the target pulse width modulation duty cycle, and the pulse width modulation signal is sent to the power drive unit of the electric heating tape to adjust the heating power of the electric heating tape.

[0040] After applying new heating power, the process returns to the step of collecting and analyzing the readings of multiple temperature sensors arranged inside the control box, thus starting the next temperature control cycle.

[0041] Preferably, the present invention also includes an intelligent temperature control system for the control box electric heating cable, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent temperature control method for the control box electric heating cable described above.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] After collecting data from multiple sensors, instead of treating them as independent sequences, spatial proximity and temporal correlation are calculated to construct a time-series graph structure with sensors as nodes and spatiotemporal correlation strength as edge weights. This aggregates discrete temperature monitoring points into a network model with topological and dependency relationships, enabling the system to explicitly capture the spatial distribution characteristics and temporal co-change patterns of heat transfer. This provides a structured and relational data foundation for understanding the overall thermal behavior inside the enclosure, surpassing the isolated judgment of local temperature values ​​used in traditional methods.

[0044] The aforementioned time-series graph structure is input into the state-space model, and the model's state transition matrix is ​​dynamically adjusted based on the edge weights in the graph. This allows the mathematical model to no longer rely on a preset, universal heat transfer coefficient, but rather to adaptively adjust the rules of internal state evolution based on the real-time calculated correlation strength between sensors. Essentially, this synchronizes the model's learning and reasoning process with the actual heat transfer dynamics in the physical world, thereby improving the simulation accuracy and prediction capability for the potential dynamics of complex, nonlinear thermodynamic processes, and overcoming the insufficient adaptability of fixed-parameter models under varying operating conditions.

[0045] After separating the periodic and trend components from the model output, the periodic components are matched and aligned point-by-point with the benchmark model library to identify abnormal fluctuation phases. The trend components are decomposed into multiple scales, and the slope and curvature of each sub-component are analyzed. This achieves refined analysis and quantitative diagnosis of temperature change patterns. It can not only detect periodic anomalies that deviate from the normal rhythm, but also quantitatively assess the rate, acceleration, and inflection point of temperature change at different time scales, thereby identifying slowly developing trend risks in advance. It provides fault warning and status assessment capabilities with deep insights beyond simple threshold alarms. Attached Figure Description

[0046] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent temperature control method for the electric heating cable of the control box described in this invention.

[0047] Figure 2 A flowchart for constructing the temperature-space-time matrix;

[0048] Figure 3 A flowchart illustrating the dynamic process of heat transfer;

[0049] Figure 4 A visualization of the temperature timing and component separation in the control box;

[0050] Figure 5 The spatiotemporal correlation intensity thermogram of the six temperature sensors inside the control box. Detailed Implementation

[0051] 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, and 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.

[0052] Please see Figure 1 This invention provides an intelligent temperature control method for an electric heating cable in a control box. The method includes: acquiring and analyzing readings from multiple temperature sensors arranged inside the control box to construct a temperature time series sequence with time, space, and event marker dimensions. Proximity calculation in the spatial dimension and correlation calculation in the temporal dimension are performed on the temperature time series sequence to construct a time series graph structure reflecting the spatiotemporal correlation strength between different sensor nodes. The time series graph structure is input into a state-space model, which dynamically adjusts its state transition matrix according to the edge weights of the time series graph structure to simulate and derive the potential dynamic process of heat transfer inside the control box. From the state vector output by the state-space model, components representing periodic temperature fluctuation patterns and components representing trend temperature change patterns are separated. The periodic temperature fluctuation pattern components are matched and aligned point-by-point with a preset reference temperature fluctuation pattern library to identify and mark abnormal fluctuation phases in the periodic temperature fluctuation pattern components. A multi-scale decomposition operation is performed on the trend temperature change pattern components to obtain trend sub-components at different time scales, and slope and curvature analysis is performed on each trend sub-component to quantify its rate of change and direction.

[0053] Example 1: See Figure 2 The readings from multiple temperature sensors are timestamped and aligned, and each reading is assigned a spatial location code according to a preset spatial coordinate mapping table. The start / stop events of the electric heating cable and the opening / closing events of the control box door are marked in the temperature time series, and the event type and duration are appended to the data points of the corresponding time periods as event marker dimensions. Missing values ​​are filled and outliers are removed from the temperature time series after the events are marked, forming a regular temperature spatiotemporal matrix as the input basis for subsequent calculations. Using each temperature sensor as a node, the correlation coefficient of the temperature reading sequences of any two sensor nodes within the sliding time window is calculated as the initial correlation strength. Based on the spatial location codes recorded in the temperature spatiotemporal matrix, the physical distance between sensor nodes is calculated, and the initial correlation strength is corrected using a distance decay function to obtain the spatial constraint correlation strength. The spatial constraint correlation strength is fused with the time series graph structure to form a weighted undirected graph with sensors as nodes and the fused spatiotemporal correlation strength as edge weights. The edge weights of the time series graph structure are updated with the sliding time window.

[0054] In practice, four temperature sensors, numbered S1, S2, S3, and S4, are deployed inside the control box. The data acquisition module synchronously records the temperature readings of all sensors once per minute. The temperature reading data packet contains the sensor number, temperature value, and timestamp information. After receiving the data, the data processing unit performs a timestamp alignment operation, normalizing the readings of all sensors at the same sampling time to the same time index. A spatial position code is assigned to each sensor reading according to a preset spatial coordinate mapping table. The spatial coordinate mapping table defines the correspondence between the sensor number and its three-dimensional coordinates within the control box. For example, the coordinate code for sensor S1 is (0.2, 0.5, 0.8), in meters. This coordinate is used for subsequent spatial proximity calculations. During the construction of the temperature time series, the event monitoring module synchronously records the start / stop events of the electric heating cable and the opening / closing events of the control box door. When the event monitoring module detects a signal change in the power drive unit of the electric heating cable or a change in the state of the magnetic control switch of the control box door, it generates an event marker with a precise timestamp and event type. The data processing unit associates these event markers with temperature data points of the corresponding time period and integrates the event type and duration as additional dimensions into the temperature time series. The temperature time series after the events are marked undergoes a data cleaning operation. The cleaning operation includes linear interpolation to fill in missing values ​​caused by brief communication interruptions and removing outliers exceeding three standard deviations based on the statistical distribution within a sliding window. Finally, a regular temperature spatiotemporal matrix is ​​formed, where the rows of the temperature spatiotemporal matrix correspond to time points, and the columns correspond to the temperature readings of each sensor and their additional event marker dimensions.

[0055] In some embodiments, a time series graph structure reflecting the spatiotemporal correlation strength between different sensor nodes is constructed based on a temperature spatiotemporal matrix. In a specific implementation, the length of the sliding time window is set to 30 minutes. Taking each temperature sensor as a node, the Pearson correlation coefficient of the temperature reading sequences of any two sensor nodes within the latest sliding time window is calculated. The correlation coefficient calculation result is used as the initial correlation strength. For example, the correlation coefficient of the temperature sequences of sensors S1 and S2 within a certain window is 0.92. Based on the spatial location codes recorded in the temperature spatiotemporal matrix, the physical distance between sensor nodes is calculated using the Euclidean distance formula. For example, the coordinates of sensors S1 and S2 are (0.2, 0.5, 0.8) and (0.2, 0.5, 0.3) respectively, and the calculated physical distance is 0.5 meters. A distance decay function is used to correct the initial correlation strength to obtain the spatially constrained correlation strength. The distance decay function is used to weaken the excessively high initial correlation strength between sensor pairs that are physically far apart. A specific form of this function is that the weighting coefficient is equal to the initial correlation strength multiplied by an exponential decay term. In a specific implementation, the expression for the distance decay function is:

[0056]

[0057] in: The edge weight represents the spatiotemporal correlation strength between sensor nodes i and j after fusion. This represents the initial correlation strength between the temperature reading sequences of sensor nodes i and j within the sliding time window. This represents the preset distance attenuation coefficient. This represents the physical distance between sensor nodes i and j. The spatial constraint correlation strength is fused with the temporal graph structure to form a weighted undirected graph with four sensors as nodes and six fused spatiotemporal correlation strength values ​​as edge weights. The edge weights of this temporal graph structure are updated every minute as the sliding time window moves. It can be understood that the length of the sliding time window, the type of correlation coefficient, and the specific form of the distance decay function can be adjusted according to the actual number of sensors deployed and the thermal characteristics of the control box.

[0058] Example 2: See Figure 3 The state-space model adaptively constructs the adjacency weight part of the state transition matrix based on the edge weights of the current time-series graph structure. The state vector from the previous time step is multiplied by the state transition matrix, and combined with an external input driven by historical ambient temperature data, the predicted state vector for the current time step is calculated. The predicted state vector is compared with a fragment of the observed temperature spatiotemporal matrix to calculate the observation residual. This residual is then used to correct the internal parameters of the state-space model through a feedback mechanism, thus completing a one-step deduction of the potential dynamic process.

[0059] In practical implementation, the state-space model is used to simulate and derive the potential dynamic process of heat transfer inside the control box. The input to the state-space model is a time-series graph structure that updates over time, containing the edge weight information between each sensor node at the current moment. Based on the edge weights of the time-series graph structure at the current moment, the state-space model adaptively constructs the adjacency weight part of the state transition matrix. The construction of the state transition matrix relies on a function that maps edge weights to matrix element weights. In practice, this function takes the edge weights between sensor nodes in the time-series graph structure as input, and generates the adjacency weight part of the state transition matrix, representing the mutual influence relationships between nodes, by weighted summation of the basic matrix and application of scaling and normalization functions. The scaling and normalization functions ensure that the generated weight values ​​are within the stable numerical range required for model operation, enabling the state transition matrix to accurately reflect the spatial correlation characteristics of the current heat transfer. Using the constructed state transition matrix, combined with the state vector from the previous moment and input terms driven by historical ambient temperature data, the state-space model performs matrix multiplication to deduce the system state at the current moment. This process dynamically simulates the propagation path and intensity of heat within the sensor node network inside the control box. The adjacency weights are adjusted in real-time based on the update of the edge weights in the time-series graph, achieving synchronous adaptation between model parameters and the spatiotemporal relationship of the physical field. In specific implementation, the adjacency weights of the state transition matrix are constructed as follows: a square matrix is ​​built based on the number of sensor nodes. The element value in the i-th row and j-th column of the matrix is ​​determined by the edge weights between sensor nodes i and j using a scaling function. This scaling function ensures that the weight values ​​are within the range required by the model. The formula for constructing the state transition matrix is ​​expressed as:

[0060]

[0061] in: This represents the adjacency weight part of the state transition matrix constructed at time t. This represents the edge weights between sensor nodes i and j at time t in the time series graph structure. Let represent a basic matrix where only the positions (i,j) and (j,i) are 1, and the remaining positions are 0. This matrix is ​​used to place edge weights in their corresponding positions. This represents a preset scaling and normalization function. Let represent a diagonal matrix indicating the heat retention characteristics of a node. The predicted state vector for the current moment is calculated by multiplying the state vector from the previous moment by the state transition matrix and incorporating an external input term driven by historical ambient temperature data. The external input term is a vector whose elements are obtained by linearly transforming readings from ambient temperature sensors deployed outside the control box over a fixed time window, reflecting the known driving influence of the external environment on the thermal processes inside the box. The calculation of the predicted state vector is the core derivation step of the state-space model, and its output is a predicted temperature field distribution vector.

[0062] In some embodiments, the predicted state vector is compared with a segment of the observed temperature-space-time matrix to calculate the observation residual. The segment of the observed temperature-space-time matrix refers to the vector composed of the latest temperature readings of all sensors extracted from the regularized temperature-space-time matrix at the latest moment. The observation residual is the difference vector between the predicted state vector and the actual observed vector, with each element in the difference vector corresponding to the prediction error of a sensor. The internal parameters of the state-space model are corrected using a feedback mechanism based on the observation residual. This feedback mechanism employs a recursive least squares algorithm, which dynamically adjusts the coefficients of the internal parameter matrices in the state-space model, excluding the adjacency weights of the state transition matrix, according to the magnitude and direction of the observation residual. It can be understood that the adjacency weights of the state transition matrix are directly determined by the edge weights of the time-series graph structure and updated in real time, while the coefficients of other internal parameter matrices are slowly corrected and learned through the observation residual. After the correction process is complete, the state-space model, using all the updated internal parameters and the current state vector, can perform a one-step deduction of the potential dynamic process of heat transfer inside the control box and prepare for the prediction calculation at the next moment.

[0063] Example 3: Fourier spectrum analysis is performed on the state vector output by the state-space model to extract frequency components with energy exceeding a preset threshold. The signals corresponding to these frequency components are then reconstructed into periodic temperature fluctuation mode components. The periodic temperature fluctuation mode components are subtracted from the original state vector to obtain a residual sequence. A local weighted regression method is applied to the residual sequence to fit a smooth trend line as the trend temperature change mode component. From a preset reference temperature fluctuation mode library, a historical normal fluctuation mode matching the current season and time is called as a reference. The difference between the periodic temperature fluctuation mode component and the reference mode at each sampling point is calculated. When the difference between multiple consecutive sampling points exceeds a dynamic threshold, the abnormal time period consisting of multiple consecutive abnormal sampling points is determined to be an abnormal fluctuation phase, and its start time, end time, and average difference intensity are recorded.

[0064] In practical implementation, the periodic temperature fluctuation mode component and the trend temperature change mode component are separated from the state vector output by the state-space model. The state vector is a multi-dimensional vector output by the state-space model at each derivation step, and its dimension is related to the number of temperature sensors. Fourier spectrum analysis is performed on the state vector output by the state-space model. A fast Fourier transform is independently performed on each component sequence in the state vector to obtain its frequency spectrum. Frequency components with energy exceeding a preset threshold are identified in the frequency spectrum. The preset threshold is set as a certain proportion of the total energy of the spectrum. Frequency components with energy exceeding the preset threshold are considered to be frequency components reflecting the main periodic fluctuation mode. All frequency components with energy exceeding the preset threshold are extracted, and the corresponding sine and cosine signals are superimposed and reconstructed. The reconstructed signal is the component representing the periodic temperature fluctuation mode. The periodic temperature fluctuation mode component is subtracted from the original state vector to obtain the residual sequence, which is the state vector sequence after removing the main periodic fluctuation. The local weighted regression method is applied to the residual sequence. The local weighted regression method fits a smooth trend line by assigning different weights to the data points in the neighborhood of each point to be fitted. The fitted smooth trend line is used as the trend temperature change pattern component, which reflects the long-term trend or slow change of temperature after removing periodic fluctuations.

[0065] In some embodiments, the identification and labeling of abnormal fluctuation phases in the periodic temperature fluctuation pattern components are based on a preset reference temperature fluctuation pattern library. This library stores historical normal periodic temperature fluctuation patterns of the control box under different seasons and time periods. From the preset reference temperature fluctuation pattern library, a historical normal fluctuation pattern matching the current season and time is retrieved as a reference. The matching process uses the date and time information provided by the system clock to search for historical pattern data with the same season and time period labels in the reference temperature fluctuation pattern library. The difference between the periodic temperature fluctuation pattern component and the reference pattern at each sampling point is calculated. The difference is a scalar value that measures the degree of deviation between the two signals at corresponding points. In a specific implementation, the formula for calculating the difference is:

[0066]

[0067] in: This represents the degree of difference at sampling point k. This represents the amplitude of the periodic temperature fluctuation mode component at sampling point k. This represents the amplitude of the reference model retrieved from the reference temperature fluctuation model library at sampling point k. This represents the standard deviation of the baseline pattern over a complete cycle. When the difference between multiple consecutive sampling points exceeds the dynamic threshold, the abnormal time period consisting of these consecutive abnormal sampling points is determined to be an abnormal fluctuation phase. The dynamic threshold is dynamically calculated based on the standard deviation of the recent historical difference sequence, and is usually set to a number of standard deviations multiples of the recent average difference. The start time, end time, and average difference intensity of the abnormal fluctuation phase are recorded. The average difference intensity is the arithmetic mean of the differences of all sampling points within the abnormal fluctuation phase. It is understood that the construction of the baseline temperature fluctuation pattern library requires the prior collection of normal operating condition data for a sufficiently long period of time and pattern extraction. Optionally, the difference can also be calculated using root mean square error or other distance metrics. In some embodiments, the minimum number of consecutive sampling points used to determine anomalies is a configurable parameter. Optionally, the calculation window length of the dynamic threshold can be related to the sliding time window length of the time sequence diagram structure. It is understood that the matching accuracy of the baseline pattern, the calculation method of the difference, and the setting rules of the dynamic threshold can be adjusted according to the stability requirements of the control box application.

[0068] See Figure 4 This is a visualization of the temperature time-series and component separation in the control box. The original temperature is composed of a superposition of "trend component" (long-term variation) and "periodic component" (regular fluctuation). This diagram intuitively illustrates the decomposition logic of the temperature signal in the temperature control system. Events marked by orange dashed lines are related to changes in the temperature trend (the trend component fluctuations are more pronounced after the event), reflecting the impact of external operations on the temperature. This helps technicians identify the driving factors (trend / period / event) of temperature changes, providing a data foundation for subsequent identification of abnormal fluctuations and optimization of the electric heating cable temperature control strategy.

[0069] Example 4: For each trend sub-component at each time scale, its first-order difference sequence is calculated as the instantaneous slope sequence, and its second-order difference sequence is calculated as the instantaneous curvature sequence. Statistical feature extraction is performed on the instantaneous slope and instantaneous curvature sequences to obtain the slope and variance, and curvature variance, which characterize the overall drastic change of each trend sub-component. The slope and curvature features corresponding to the trend sub-components at different time scales are concatenated from coarse to fine according to the time scale to form a multi-scale feature vector describing the trend change.

[0070] In practice, a multi-scale decomposition operation is performed on the trend temperature change pattern components. This multi-scale decomposition employs the Empirical Mode Decomposition (EMD) method, which adaptively decomposes the trend temperature change pattern components into a series of intrinsic mode functions ranging from high to low frequencies. Specifically, after decomposition, three trend sub-components at specific time scales are retained. These three time scales correspond to short-term fluctuations, medium-term changes, and long-term trends, respectively. The time scale for the short-term fluctuation sub-component is several minutes to tens of minutes, for the medium-term change sub-component it is several hours, and for the long-term trend sub-component it is several hours to tens of hours. After obtaining the trend sub-components at different time scales, slope and curvature analyses are performed for each sub-component at each time scale. For the short-term fluctuation sub-component, its first-order difference sequence is calculated as the instantaneous slope sequence, reflecting the rate of change of amplitude between adjacent sampling points. Simultaneously, the second-order difference sequence of the short-term fluctuation sub-component is calculated as the instantaneous curvature sequence, reflecting the rate of change of the first-order difference sequence, i.e., an approximate representation of the acceleration of temperature change. The same calculation process is performed on the intermediate-term change sub-component and the long-term trend sub-component to obtain their respective instantaneous slope sequences and instantaneous curvature sequences. Statistical feature extraction is then performed on the instantaneous slope sequences and instantaneous curvature sequences. This extraction includes calculating the mean, variance, maximum, and minimum values ​​for each sequence. In practice, the mean and variance of the instantaneous slope sequence and the mean and variance of the instantaneous curvature sequence are extracted to obtain the slope mean and slope variance, and the curvature mean and curvature variance, which characterize the overall drastic change of each trend sub-component.

[0071] In some embodiments, the slope and curvature features corresponding to trend sub-components at different time scales are concatenated to form a multi-scale feature vector describing trend changes. The concatenation is performed in order of coarse to fine time scale, i.e., features of long-term trend sub-components first, features of medium-term change sub-components in the middle, and features of short-term fluctuation sub-components last. The multi-scale feature vector is a numerical vector, as shown in Table 1.

[0072] Table 1: Multi-scale Feature Vector Construction Table

[0073]

[0074] In practical implementation, the formula for calculating the curvature variance is:

[0075]

[0076] in: This represents the curvature variance calculated at a certain time scale. This represents the total number of data points representing curvature values ​​in this trend sub-component. This represents the curvature value at the nth point in the instantaneous curvature sequence of the trend subcomponent. This represents the arithmetic mean of all curvature values ​​in the instantaneous curvature sequence of the trend subcomponent. It is understood that the empirical mode decomposition method can be replaced by wavelet decomposition or other time-frequency analysis methods. Optionally, higher-order moment features such as skewness and kurtosis can be added to the extracted statistical features. In some embodiments, the number and division of time scales can be adjusted according to the granularity requirements of trend analysis in the actual application. Optionally, the concatenation order of the multi-scale feature vectors can also adopt a fine-to-coarse or other custom order.

[0077] Example 5: The labeling information of the abnormal fluctuation phase, the multi-scale feature vector, and the latest edge weights of the time-series graph structure are input into a fusion decision network. The fusion decision network includes a temporal attention sub-network for processing the labeling information, a fully connected sub-network for processing the multi-scale feature vector, and a graph convolution sub-network for processing the time-series graph structure. The fusion decision network performs feature-level fusion of the outputs of the three sub-networks and calculates the probability distribution of the modulation intensity that the electric heating cable needs to be applied at the current moment through a probability output layer. According to the modulation intensity probability distribution, the modulation intensity level with the highest probability is selected, and the target pulse width modulation duty cycle is obtained by querying the preset modulation intensity-duty cycle mapping table. The corresponding pulse width modulation signal is generated according to the target pulse width modulation duty cycle and sent to the power drive unit of the electric heating cable to adjust the heating power of the electric heating cable. After applying the new heating power, the process returns to the step of collecting and parsing the readings of multiple temperature sensors arranged inside the control box, and the next round of temperature control cycle is started.

[0078] In practical implementation, the labeling information of anomalous fluctuation phases, multi-scale feature vectors, and the latest edge weights of the temporal graph structure are jointly input into a fusion decision network. The labeling information of anomalous fluctuation phases includes the start time, end time, and average difference intensity of the anomalous time period. The multi-scale feature vectors are numerical vectors containing statistical features of slope and curvature at different time scales. The latest edge weights of the temporal graph structure are matrices reflecting the spatiotemporal correlation strength between sensor nodes at the current moment. The fusion decision network includes a temporal attention sub-network for processing labeling information, a fully connected sub-network for processing multi-scale feature vectors, and a graph convolution sub-network for processing the temporal graph structure. The temporal attention sub-network receives the labeling information of anomalous fluctuation phases, first encoding the start time and end time into positional encoding vectors relative to the current moment, and using the average difference intensity as a numerical feature. These features are concatenated and input into a sequence model based on a self-attention mechanism. The sequence model can capture the temporal dependencies and importance differences of different anomalous phases and output a fixed-length feature vector. The fully connected subnetwork receives multi-scale feature vectors, which are one-dimensional real-valued vectors. The fully connected subnetwork consists of multiple alternating linear layers and non-linear activation functions, used for non-linear transformation and higher-order feature extraction of the multi-scale feature vectors, outputting a feature vector whose dimension matches that of the temporal attention subnetwork. The graph convolutional subnetwork receives the latest edge weights of the temporal graph structure, represented as an adjacency matrix. The graph convolutional subnetwork performs feature propagation and aggregation operations on the nodes of the graph. Node features are initialized to the temperature readings of the corresponding sensors. After multiple layers of graph convolution operations, global pooling is performed on all updated node features. This pooling operation generates a global graph feature vector whose dimension matches the output dimensions of the other subnetworks.

[0079] In some embodiments, the fusion decision network performs feature-level fusion of the outputs of the three sub-networks. Feature-level fusion uses vector concatenation, connecting the feature vectors output by the temporal attention sub-network, the fully connected sub-network, and the graph convolution sub-network along their feature dimensions to form a fused feature vector. A probability output layer calculates the probability distribution of the required control intensity of the electric heating cable at the current moment. This probability output layer consists of a linear transformation layer and a Softmax activation function. The linear transformation layer maps the fused feature vector into a logistic vector with K elements, where K represents the total number of preset control intensity levels. The calculation formula for the probability output layer is:

[0080]

[0081] in: The calculated probability distribution of regulatory intensity is represented by a K-dimensional probability vector, where each element represents the probability of the corresponding regulatory intensity level being selected. This represents the fused feature vector composed of the outputs of the three sub-networks. and Let represent the weight matrix and bias vector of the linear transformation of the probability output layer, respectively. The Softmax function is used to normalize a logic value vector into a probability distribution. Based on the probability distribution of the control intensity, the control intensity level with the highest probability is selected. This selection operation uses a maximum value index function, which returns the level number corresponding to the highest probability value in the probability vector. The target pulse width modulation (PWM) duty cycle is obtained by querying a preset control intensity-duty cycle mapping table, which stores the mapping relationship from each control intensity level to the specific PWM signal duty cycle value. A corresponding PWM signal is generated based on the target PWM duty cycle. The frequency of the PWM signal is fixed, and the duration of the high level is determined by the target duty cycle and the signal period. The PWM signal is sent to the power drive unit of the electric heating tape. The power drive unit adjusts its average output power to the electric heating tape based on the received PWM signal, thereby adjusting the heating power. After applying a new heating power, the control system returns to the step of collecting and parsing the readings of multiple temperature sensors arranged inside the control box, starting the next temperature control cycle.

[0082] It is understood that the specific architecture of the temporal attention subnetwork, the number of layers in the fully connected subnetwork, the number of layers in the graph convolutional subnetwork, and the total number of modulation intensity levels K are all configurable parameters. Optionally, feature-level fusion can also employ weighted summation or a weighted fusion method based on an attention mechanism. In some embodiments, the content of the modulation intensity-duty cycle mapping table can be adjusted through offline calibration or online learning. Optionally, the power drive unit can be an insulated gate bipolar transistor driver or a solid-state relay.

[0083] See Figure 5 This is a heatmap showing the spatiotemporal correlation strength of six temperature sensors within the control box. The yellow areas on the diagonal (e.g., S1-S1, S2-S2) represent complete correlation within the sensors themselves; the yellow / green areas off-diagonal (e.g., S1-S2, S2-S3) indicate strong positive correlations between the corresponding sensors, reflecting the synchronicity of temperature changes. This diagram corresponds to the time-series structure of the temperature control system: by analyzing the correlation strength between sensors, tightly coupled regions of temperature transfer can be identified, providing a topological basis for constructing a state-space model and simulating the heat transfer process. It helps technicians locate sensor groups with high temperature change synchronicity, assisting in optimizing sensor layout or the regional targeting of temperature control strategies.

[0084] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent temperature control of an electric heating cable in a control box, characterized in that, The method includes: The readings of multiple temperature sensors arranged inside the control box are collected and analyzed to construct a temperature time series with time dimension, spatial dimension and event tag dimension; The spatial proximity calculation and temporal correlation calculation are performed on the temperature time series to construct a time series graph structure that reflects the spatiotemporal correlation strength between different sensor nodes. The time sequence diagram structure is input into a state space model, and the state space model dynamically adjusts its state transition matrix according to the edge weights of the time sequence diagram structure to simulate and derive the potential dynamic process of heat transfer inside the control box. From the state vector output by the state space model, components representing periodic temperature fluctuation patterns and components representing trend-based temperature change patterns are separated. The periodic temperature fluctuation pattern component is matched and aligned point by point with a preset reference temperature fluctuation pattern library to identify and mark the abnormal fluctuation phase in the periodic temperature fluctuation pattern component. Multiscale decomposition is performed on the trend temperature change pattern components to obtain trend sub-components at different time scales, and slope and curvature analysis is performed on each trend sub-component to quantify its rate of change and direction. The process of simulating and deriving the potential dynamic process of heat transfer inside the control box includes: The state-space model adaptively constructs the adjacency weight part of the state transition matrix based on the edge weights of the current time sequence graph structure. The predicted state vector at the current moment is calculated by multiplying the state vector of the previous moment with the state transition matrix and combining it with an external input driven by historical ambient temperature data. The predicted state vector is compared with the temperature spatiotemporal matrix fragment obtained from actual observation, the observation residual is calculated, and the internal parameters of the state space model are corrected through a feedback mechanism using the observation residual to complete a one-step deduction of the potential dynamic process. The separation of components characterizing periodic temperature fluctuation patterns and components characterizing trend-based temperature change patterns includes: Fourier spectrum analysis is performed on the state vector output by the state space model to extract frequency components with energy exceeding a preset threshold, and the signal corresponding to the frequency components is reconstructed into periodic temperature fluctuation mode components. The periodic temperature fluctuation pattern component is subtracted from the original state vector to obtain a residual sequence. A local weighted regression method is applied to the residual sequence to fit a smooth trend line as the trend temperature change pattern component.

2. The intelligent temperature control method for the electric heating cable of the control box according to claim 1, characterized in that, The construction of a temperature time series with time, space, and event marker dimensions includes: The readings from multiple temperature sensors are timestamped and aligned, and each reading is assigned a spatial location code according to a preset spatial coordinate mapping table. In the temperature time series, mark the times of electric heating cable start-up and shutdown events and control box door opening and closing events, and attach the event type and duration as event marker dimensions to the data points of the corresponding time period; The temperature time series after the marked event is filled with missing values ​​and outliers are removed to form a regular temperature-space-time matrix as the input basis for subsequent calculations.

3. The intelligent temperature control method for the control box electric heating cable according to claim 2, characterized in that, The method for constructing a time series graph structure that reflects the spatiotemporal correlation strength between different sensor nodes includes: Using each temperature sensor as a node, calculate the correlation coefficient of the temperature reading sequences of any two sensor nodes within the sliding time window, and use it as the initial correlation strength. Based on the spatial location codes recorded in the temperature spatiotemporal matrix, the physical distance between sensor nodes is calculated, and the initial correlation strength is corrected using a distance decay function to obtain the spatial constraint correlation strength. The spatial constraint correlation strength is fused with the temporal graph structure to form a weighted undirected graph with sensors as nodes and the fused spatiotemporal correlation strength as edge weights. The edge weights of the temporal graph structure are updated with the sliding time window.

4. The intelligent temperature control method for the electric heating cable of the control box according to claim 3, characterized in that, The step of identifying and marking anomalous fluctuation phases in the periodic temperature fluctuation pattern components includes: From the preset reference temperature fluctuation pattern library, call up a historical normal fluctuation pattern that matches the current season and time as a reference; The difference between the periodic temperature fluctuation mode component and the reference mode at each sampling point is calculated. When the difference between multiple consecutive sampling points exceeds the dynamic threshold, the abnormal time period composed of the multiple consecutive abnormal sampling points is determined to be an abnormal fluctuation phase, and its start time, end time and average difference intensity are recorded.

5. The intelligent temperature control method for the electric heating cable of the control box according to claim 1, characterized in that, The slope and curvature analysis performed on each trend sub-component includes: For each trend sub-component at each time scale, its first-order difference sequence is calculated as the instantaneous slope sequence, and its second-order difference sequence is calculated as the instantaneous curvature sequence. Statistical features are extracted from the instantaneous slope sequence and instantaneous curvature sequence to obtain the slope and variance and curvature variance, which characterize the overall drastic change of each trend sub-component. The slope and curvature features corresponding to the trend sub-components at different time scales are concatenated from coarse to fine time scales to form a multi-scale feature vector describing trend changes.

6. The intelligent temperature control method for the electric heating cable of the control box according to claim 5, characterized in that, Includes: inputting the labeling information of the abnormal fluctuation phase, the multi-scale feature vector, and the latest edge weights of the time series graph structure into a fusion decision network; The fusion decision network includes a temporal attention subnetwork for processing labeled information, a fully connected subnetwork for processing multi-scale feature vectors, and a graph convolution subnetwork for processing temporal graph structures. The fusion decision network performs feature-level fusion of the outputs of the three sub-networks and calculates the probability distribution of the control intensity that the electric tracing heat cable needs to be applied at the current moment through a probability output layer.

7. The intelligent temperature control method for the electric heating cable of the control box according to claim 6, characterized in that, Includes: selecting the highest probability modulation intensity level based on the modulation intensity probability distribution, and querying a preset modulation intensity-duty cycle mapping table to obtain the target pulse width modulation duty cycle; A corresponding pulse width modulation signal is generated based on the target pulse width modulation duty cycle, and the pulse width modulation signal is sent to the power drive unit of the electric heating tape to adjust the heating power of the electric heating tape. After applying new heating power, the process returns to the step of collecting and analyzing the readings of multiple temperature sensors arranged inside the control box, thus starting the next temperature control cycle.

8. A control box-type 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 the electric heating cable of the control box as described in any one of claims 1 to 7.

Images