A method for preparing a polyurethane anticorrosive coating

Abstract

The application relates to a preparation method of a polyurethane anticorrosive paint and relates to the technical field of polyurethane anticorrosive paint preparation, and comprises the following steps: obtaining a first characteristic vector corresponding to a stirring device; establishing a temperature matrix corresponding to the stirring device; performing feature extraction on the temperature matrix to obtain a second characteristic vector corresponding to the stirring device; fusing the first characteristic vector and the second characteristic vector to obtain a fusion characteristic vector corresponding to the stirring device; inputting the fusion characteristic vector into a preset solvent addition confidence degree prediction model to output the confidence degree of the solvent to be added; if the obtained continuous confidence degrees are all smaller than a preset confidence threshold and present a downward trend, the fusion characteristic vector is input into a preset reaction end time prediction model to obtain a predicted reaction end time; and the polyurethane prepolymer in stirring is detected according to the reaction end time to determine whether the polyurethane prepolymer reaction is completed; and the method can guarantee the quality of the finally prepared polyurethane anticorrosive paint.

Description

Technical Field

[0001] This invention relates to the field of polyurethane anticorrosive coating preparation technology, and specifically to a method for preparing a polyurethane anticorrosive coating. Background Technology

[0002] Polyurethane anti-corrosion coatings, with their excellent chemical resistance, adhesion, and mechanical properties, are widely used in shipbuilding, petrochemicals, steel structures, and other fields. The preparation of polyurethane anti-corrosion coatings begins with the preparation of a polyurethane prepolymer. This prepolymer is generated by the reaction of isocyanate and polyol, requiring gentle and uniform stirring to ensure uniform temperature and concentration, avoiding localized overheating or incomplete reaction. During the prepolymerization reaction, stirring should be slow and prolonged to ensure uniform contact between the isocyanate and polyol, preventing excessively high temperatures that could lead to side reactions. In the polyurethane prepolymer preparation stage, the polyol viscosity is low initially, but increases as the reaction progresses. A small amount of solvent (such as xylene) can be added to reduce viscosity and ensure uniform stirring. Furthermore, accurate determination of the reaction endpoint is crucial to avoid over-reaction, which can affect the quality of the polyurethane anti-corrosion coating. Current technologies typically use real-time monitoring to determine whether solvent needs to be added and whether the reaction endpoint has been reached. However, this method requires continuous monitoring of the polyurethane prepolymer and has a significant time delay. By the time solvent addition is detected or the reaction endpoint is reached, the polyurethane prepolymer may have already over-reacted, affecting the quality of the final polyurethane anti-corrosion coating. Summary of the Invention

[0003] To achieve the objective of this invention, the technical solution adopted is as follows: a method for preparing a polyurethane anti-corrosion coating, the method comprising the following steps: S100: During the stirring reaction of polyurethane prepolymer, the torque, motor output power, and temperature at different locations within the stirring device are obtained at each sampling moment within a preset sliding time window; the end time of the preset sliding time window is the current time. S200, extract time-series features from the acquired torque and motor output power time-series data corresponding to the stirring device to obtain the first feature vector corresponding to the stirring device; S300: Based on the temperature at different locations within the stirring device at each sampling moment within a preset sliding time window, establish a temperature matrix corresponding to the stirring device. The temperature matrix includes several rows and several columns, with each row corresponding to the temperature at different sampling moments at the same location and each column corresponding to the temperature at different locations at the same sampling moment. S400, extract features from the temperature matrix to obtain the second feature vector corresponding to the stirring device; S500, the first feature vector and the second feature vector are fused to obtain the fused feature vector corresponding to the stirring device; S600 inputs the fused feature vector into a preset solvent addition confidence prediction model to output the confidence of the solvent to be added in the preset future time period corresponding to the current sliding time window; S700, if several consecutive confidence scores are all less than the preset confidence threshold and show a downward trend, then the fused feature vector is input into the preset reaction endpoint time prediction model to obtain the predicted reaction endpoint time. S800, based on the reaction endpoint time, detects the polyurethane prepolymer during stirring to determine if the polyurethane prepolymer reaction is complete.

[0004] Compared with the prior art, the beneficial effects of the present invention are: 1. By acquiring the torque, motor output power, and temperature at different positions within a preset sliding time window during the stirring reaction of polyurethane prepolymer in real time, and after feature extraction and fusion, the confidence level of solvent addition and the reaction endpoint time are predicted using a preset model. This allows for the early determination of solvent addition requirements and reaction endpoints, effectively avoiding the delay problem of real-time detection methods in existing technologies, reducing the impact of over-reaction on the quality of polyurethane prepolymer, and thus ensuring the quality of the final polyurethane anticorrosive coating.

[0005] 2. The method in this invention can predict the reaction endpoint time of polyurethane prepolymer. Therefore, when there is a long time before the reaction endpoint, it is not necessary to continuously monitor the polyurethane prepolymer in the reaction. This simplifies the operation process and improves the preparation efficiency while ensuring the quality of the prepared polyurethane anticorrosive coating. Attached Figure Description

[0006] Figure 1 This is a schematic flowchart of the preparation method of the polyurethane anti-corrosion coating provided by the present invention. Detailed Implementation

[0007] 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.

[0008] Example: like Figure 1 As shown, the present invention provides a technical solution: a method for preparing a polyurethane anti-corrosion coating, which may include the following steps: S100: During the stirring reaction of the polyurethane prepolymer, the torque, motor output power, and temperature at different locations within the stirring device are obtained at each sampling moment within a preset sliding time window; the end time of the preset sliding time window is the current time.

[0009] In this embodiment, polyurethane prepolymer is used to prepare polyurethane anti-corrosion coating. The preparation process of polyurethane prepolymer is exothermic. Polyurethane prepolymer is generated by the reaction of isocyanate and polyol. During the reaction, the viscosity of the material changes as the reaction proceeds. The viscosity change affects the stirring difficulty, which is reflected in the torque and motor output power. At the same time, if the reaction is not uniform, temperature differences will occur at different locations. During the stirring reaction of polyurethane prepolymer, the torque of the stirring device, the motor output power, and the temperature at different locations in the stirring device can be obtained at each sampling moment within the preset sliding time window.

[0010] The length of the preset sliding time window can be roughly determined through preliminary experiments: In the initial stage of the reaction, due to slow viscosity changes and small temperature fluctuations, a relatively long sliding time window can be set, such as 30-60 minutes, to stably capture the basic trend of the reaction; as the reaction progresses, when the viscosity of the prepolymer begins to increase significantly and temperature fluctuations intensify, the sliding time window can be shortened to 10-30 minutes to more sensitively capture rapid changes in the reaction state. Simultaneously, the duration can be dynamically adjusted and optimized according to different reaction scales, stirring device characteristics, and requirements for reaction control precision, ensuring that the sliding time window accurately and promptly reflects the reaction state within the preset time period. This provides a reliable data foundation for subsequent feature extraction and model prediction, thereby better enabling accurate judgment of solvent addition and reaction endpoints.

[0011] The torque and motor output power of the stirring device can be obtained through specialized sensing equipment and monitoring systems. For torque, a torque sensor can be installed on the stirring shaft. This sensor can directly detect the torque force experienced by the stirring shaft when it rotates. Its working principle is based on the resistance change of a strain gauge. When the stirring shaft generates torque due to material resistance, the strain gauge deforms, and the resistance changes accordingly. This change is converted into an electrical signal through a circuit, thereby obtaining the specific torque value.

[0012] The motor's output power can be obtained through a motor power monitoring device, which is usually connected to the motor's power supply circuit. This device can monitor parameters such as the motor's voltage, current, and power factor in real time, and then calculate the motor's output power according to the power calculation formula (motor output power = voltage × current × power factor × efficiency). In the stirring reaction of polyurethane prepolymer, the stirring resistance changes with the viscosity of the material, which in turn leads to a change in torque. In order to maintain the stirring operation, the motor's output power will also be adjusted accordingly with the change in torque.

[0013] The advantage of obtaining these two parameters is that they can reflect the viscosity state of the material in real time. In the initial stage of the reaction, the polyol viscosity is low, resulting in lower torque and motor output power. As the reaction progresses, the viscosity of the polyurethane prepolymer increases, leading to increased stirring resistance, and consequently, higher torque and motor output power. Monitoring these two parameters provides fundamental data for subsequent time-series feature extraction, which in turn provides crucial information for determining whether solvent addition is necessary and predicting the reaction endpoint. This helps in timely adjustments to reaction conditions, ensuring the smooth progress of the reaction and the quality of the prepolymer.

[0014] The temperature at different locations within the stirring device can be measured by temperature sensors installed at several different locations within the stirring device. These temperature sensors can be thermocouple temperature sensors.

[0015] S200, the time-series data of torque and motor output power corresponding to the obtained stirring device are used to extract time-series features to obtain the first feature vector corresponding to the stirring device.

[0016] In this embodiment, time-series feature extraction is performed on the acquired torque and motor output power time-series data corresponding to the stirring device to obtain the first feature vector corresponding to the stirring device. In the initial stage of the reaction, the polyol viscosity is low, and the torque and motor output power are small and stable; as the reaction proceeds, the generated polyurethane prepolymer increases the material viscosity, and the torque and motor output power gradually increase. Time-series feature extraction can capture the changing trends and fluctuation frequencies of these parameters.

[0017] This step transforms the original time-series data into a feature vector that reflects the viscosity variation of the material, facilitating precise analysis of the reaction stage and providing a more accurate basis for determining whether solvent needs to be added, thus improving the scientific rigor of reaction state assessment.

[0018] Furthermore, step S200 may include the following steps: S210, preprocess the torque timing data and motor output power timing data within the preset sliding time window. The preprocessing includes outlier handling and missing value filling.

[0019] During the stirring reaction of polyurethane prepolymer, the time-series data of torque and motor output power may exhibit outliers due to momentary equipment failures, signal transmission interference, or missing values ​​due to sampling intervals or brief equipment malfunctions. Outlier handling can identify and correct data deviating from the normal range using statistical methods (such as the Z-score method), while missing value imputation can be achieved using linear interpolation, mean imputation, or other methods to fill in the missing data.

[0020] This step eliminates noise and incomplete information in the original data, ensuring the accuracy and continuity of the time-series data. It provides reliable basic data for subsequent feature extraction, avoids interference from abnormal or missing data on the feature extraction results, and ensures that the extracted features can truly reflect the changes in material viscosity.

[0021] S220 extracts statistical features, trend features, and fluctuation features from the preprocessed torque time series data and motor output power time series data, respectively.

[0022] The feature extraction logic for torque timing data and motor output power timing data is consistent, but they are calculated independently, because torque and power reflect different dimensions of the stirring load.

[0023] Furthermore, the statistical features include: mean, median, variance, standard deviation, range, skewness, kurtosis, maximum value, minimum value, and percentage of peak values; the trend features include: linear trend slope, trend fit goodness of fit, and cumulative change; the fluctuation features include: mean of the first difference, standard deviation of the first difference, root mean square difference of the first difference, and percentage of positive change.

[0024] Statistical feature extraction reflects the overall distribution characteristics. Statistical features are used to describe the overall size, distribution pattern, and extreme values ​​of time series data within a sliding window. The following features are calculated for torque and power series respectively: Central tendency characteristics: mean (reflects average torque / power, such as the mean increases as viscosity increases), median (a central tendency indicator against extreme values).

[0025] Dispersion characteristics: variance (reflects the amplitude of fluctuation, such as increased fluctuation when there is uneven local response), standard deviation (has the same trend as variance, and the unit is consistent with the original data), range (maximum value minus minimum value, reflecting the range of fluctuation).

[0026] Distribution characteristics: skewness (symmetry of data distribution, such as right skewness when viscosity increases suddenly), kurtosis (steepness of data distribution, such as high kurtosis when there are violent fluctuations).

[0027] Extreme value characteristics: maximum value (maximum torque / power within the window), minimum value (minimum torque / power within the window), peak percentage (the percentage of times exceeding 1.5 times the average, reflecting the frequency of high-frequency loads).

[0028] Trend feature extraction reflects the changing trend over time; trend features are used to capture the direction and intensity of torque and power changes as the reaction progresses. Since the viscosity of the prepolymer gradually increases with the reaction, torque and power usually show an upward trend. Calculations are performed separately for the torque and power sequences. Linear trend slope: Linear regression of time series data is performed using the least squares method; specifically, with time as the x-axis and torque / power as the y-axis, the slope k is obtained. k > 0 indicates an upward trend, and the larger the absolute value of k, the more significant the trend.

[0029] Trend fit goodness of fit: R-squared of linear regression 2 R value; this value reflects the degree to which the data fits a linear trend. 2 The closer it is to 1, the more stable the trend.

[0030] Cumulative change: The difference between the data at the last moment and the data at the first moment within the window; the difference reflects the total change within the window.

[0031] Fluctuation features are extracted to reflect short-term dynamic changes. These features describe the drastic changes in time-series data between adjacent time points and reflect the uniformity of stirring. For example, local overheating may cause sudden viscosity changes, leading to increased fluctuations. These features are calculated separately for torque and power sequences. First-order difference statistics: Calculate the difference between adjacent time points, and then calculate the mean (average rate of change), standard deviation (stability of change), and root mean square deviation (RMS, reflecting the energy of fluctuation) of the difference sequence.

[0032] Fluctuation frequency characteristics: Calculate the proportion of "positive change" (Δ>0) in the difference sequence (reflecting the frequency of torque / power increase, with a high proportion when viscosity continues to increase).

[0033] S230, the statistical features, trend features, and fluctuation features corresponding to the torque time series data are sequentially concatenated with the statistical features, trend features, and fluctuation features corresponding to the motor output power time series data to form the first feature vector.

[0034] All features extracted from the torque sequence (such as the 15 features in the above 3 categories) are concatenated in sequence with the same type of features extracted from the motor output power sequence (15 features) to form the first feature vector.

[0035] Through the above steps, the first feature vector can comprehensively capture the overall distribution, trend and dynamic fluctuation of torque and motor power within the sliding window, providing a basis for subsequent prediction of solvent addition requirements and reaction endpoint based on fusion temperature characteristics, because these features are directly related to the viscosity change and stirring uniformity of the prepolymer.

[0036] S300: Based on the temperature at different locations within the stirring device at each sampling time within a preset sliding time window, establish a temperature matrix corresponding to the stirring device; the temperature matrix includes several rows and several columns, with each row corresponding to the temperature at different sampling times at the same location, and each column corresponding to the temperature at different locations at the same sampling time.

[0037] In this embodiment, the temperature at the location where the temperature sensor is installed in the stirring device can be obtained at each sampling time, thereby enabling the establishment of a temperature matrix corresponding to the stirring device.

[0038] This step systematizes and structures the dispersed temperature data, clearly presenting the temperature changes at different locations over time and the temperature differences at different locations at the same moment. This helps to intuitively understand whether there is local overheating in the reaction, which is a problem that needs to be avoided in the prepolymerization reaction. This matrix provides a clear data structure to support subsequent analysis of reaction uniformity.

[0039] S400, extract features from the temperature matrix to obtain the second feature vector corresponding to the stirring device.

[0040] Furthermore, step S400 may include the following steps: S410 constructs a spatial feature extraction layer for the temperature matrix, using a one-dimensional convolution kernel to perform convolution operations on the temperature matrix to extract temperature correlation features between different locations at the same time.

[0041] Based on the temperature matrix generated by S300, spatial feature extraction focuses on the spatial distribution pattern of temperature at different locations at the same time (i.e., the "column dimension" feature of the temperature matrix). A one-dimensional convolution kernel (such as a 3×1 or 5×1 kernel, sliding along the "position dimension") is used to perform convolution operations on each column of the temperature matrix (at the same time): the convolution kernel captures the temperature differences, gradient changes, or cooperative fluctuations between adjacent locations by weighted summation with the temperature values ​​at different locations; for example, the temperature difference between the left and right positions in the stirring device, and the temperature correlation between the local high temperature zone and the surrounding area.

[0042] To enhance feature representation capabilities, nonlinear transformations are introduced after convolution operations using activation functions such as ReLU to generate multiple spatial feature maps (each feature map corresponds to a spatial correlation pattern, such as "local temperature uniformity" or "temperature difference between edge and center").

[0043] This step enables precise extraction of temperature distribution characteristics in the spatial dimension, directly reflecting the temperature uniformity within the stirring device (such as the presence of localized overheating or temperature stratification), thus addressing the core requirement in the background technology of "ensuring temperature uniformity to avoid localized overheating." By leveraging the local sensing capabilities of the convolutional kernel, subtle spatial temperature correlations (such as the synchronous temperature change trends of adjacent sensors) are captured, providing underlying feature support for subsequent judgments on reaction uniformity.

[0044] S420 constructs a time feature extraction layer for the temperature matrix, and uses a bidirectional LSTM network to process the output of the spatial feature extraction layer to capture the temperature change trend features at different times.

[0045] The spatial feature maps output by the spatial feature extraction layer are arranged in temporal order (i.e., spatial features at the same location form a temporal sequence with each sampling time), serving as the input to the bidirectional LSTM network. The bidirectional LSTM network includes a "forward LSTM" and a "backward LSTM": the forward LSTM learns the evolution of temperature spatial features from the beginning of the time window to the current time (e.g., how past temperature spatial distribution affects the current state); the backward LSTM learns backward from the current time to the beginning of the time window (e.g., the dependence of the current state on past features).

[0046] By fusing the hidden layer outputs of the bidirectional LSTM (such as splicing the last hidden state of the forward and backward directions), we obtain temporal features that contain bidirectional time dependencies of "past-present" and "present-future", reflecting the dynamic trend of temperature spatial distribution over time (such as whether temperature uniformity improves or deteriorates with the reaction, and whether the overall temperature shows a continuous upward / fluctuating trend).

[0047] This step allows for the capture of long-term temperature trends over time (such as a continuous temperature rise due to exothermic reaction and reduced temperature fluctuations due to improved stirring uniformity), addressing the need to monitor reaction dynamics to avoid abnormal reactions in the background technology. The LSTM gating mechanism effectively preserves key time node information within a long time window (such as the moment of temperature abrupt change), avoiding the "gradient vanishing" problem of traditional time series models, making it more suitable for extracting time series features of long-term reactions of polyurethane prepolymers.

[0048] S430 constructs a spatiotemporal attention fusion layer, applies a self-attention mechanism to the output of the temporal feature extraction layer, calculates the contextual dependencies between different times and locations, and generates weighted spatiotemporal features.

[0049] The "spatiotemporal sequence features" (each time step corresponds to a temporal feature vector containing spatial information) output by the temporal feature extraction layer are mapped into three matrices: Query, Key, and Value. Query represents the feature to be focused on, Key represents the feature used for matching, and Value represents the feature to be weighted.

[0050] The similarity between the query and the key is calculated (e.g., by dot product operation) to obtain the attention score, which reflects the influence weight of features at different times and positions on the current feature; the score is normalized by the Softmax function to obtain the attention weight of each feature (the higher the weight, the more critical the feature at that time / position).

[0051] The Value matrix is ​​weighted and summed based on attention weights to generate spatiotemporal features that incorporate context-dependent features (e.g., if a local high temperature feature at a certain moment is strongly correlated with the subsequent reaction endpoint, the weight at that moment is amplified; if a temperature fluctuation at a certain location has no effect on the overall reaction, the weight is suppressed).

[0052] By dynamically focusing on key spatiotemporal information through a self-attention mechanism (such as a sudden temperature rise at a certain location in the later stage of the reaction, or a sudden change in temperature uniformity in the middle stage of the reaction), the features that play a decisive role in the reaction process are strengthened, while irrelevant interfering information (such as slight random temperature fluctuations) is weakened. Combined with contextual relationships (such as "whether spatial temperature non-uniformity at a certain moment will lead to subsequent temperature runaway"), the features are made more targeted, providing "high-value" feature support for subsequent solvent addition and reaction endpoint prediction, thus solving the problem of "overreaction due to detection delay" in existing technologies.

[0053] S440 constructs a feature vector generation layer, performs global average pooling on the output of the spatiotemporal attention fusion layer, and maps it to a second feature vector of fixed dimension through a fully connected layer.

[0054] Global average pooling is performed on the feature map output by the spatiotemporal attention fusion layer (which contains spatiotemporally weighted features of multiple channels): the average value of all elements in each channel is taken, and the high-dimensional feature map is compressed into a low-dimensional channel feature vector (reducing the parameter scale and avoiding overfitting).

[0055] The fully connected layer maps the channel feature vectors to fixed-length vectors of a preset dimension (such as 256-dimensional or 512-dimensional), which are the second feature vectors. The fully connected layer further integrates global spatiotemporal features through linear transformation and nonlinear activation (such as ReLU) to ensure the dimensionality of the output features (facilitating subsequent fusion with the first feature vector).

[0056] In this step, global average pooling preserves the global distribution information of spatiotemporal features while reducing computational complexity and avoiding model redundancy caused by excessively high feature dimensions. The fixed-dimensional second feature vector can be directly fused with the first feature vector (torque and power features) extracted by S200, ensuring a unified input format for subsequent prediction models (solvent addition confidence model, reaction endpoint time model), and improving the effectiveness of feature fusion and model inference efficiency.

[0057] The final generated second feature vector fully contains the spatial distribution, temporal trend, and key contextual dependencies of temperature, providing comprehensive temperature feature support for accurately determining the reaction state (such as whether solvent needs to be added or whether it is close to the endpoint).

[0058] Furthermore, the convolution operation of the spatial feature extraction layer includes at least two one-dimensional convolutions, with ReLU activation performed after each convolution.

[0059] First-level convolution: Input: Temperature matrix (dimensions: [batch size, time step, number of sensor locations]).

[0060] Convolution operation: Use multiple one-dimensional convolution kernels (such as 16 convolution kernels of size 3) to convolve the temperature vector (i.e., each row of the matrix) at each time step.

[0061] Mathematical expression: Conv1(x)=ReLU(W1∗x+b1); where x is the input vector, W1 is the convolution kernel weight, b1 is the bias, and ReLU is the activation function.

[0062] Output: Generates 16 feature maps, each capturing a different spatial pattern (such as temperature gradient, local extrema).

[0063] Second-level convolution: Input: The output of the first-level convolution (16 feature maps).

[0064] Convolution operation: Use larger convolution kernels (e.g., 5) or more (e.g., 32) to further extract advanced features.

[0065] Mathematical expression: Conv2(x) = ReLU(W2∗Conv1(x)+b2); W2 is the convolution kernel weight, and b2 is the bias.

[0066] Output: 32 more abstract feature maps that incorporate combination patterns of low-level features (such as "hotspot diffusion" or "temperature stratification").

[0067] By progressively extracting spatial features from local to global through multi-level convolution (such as the first level capturing temperature differences between adjacent locations and the second level capturing coordinated temperature changes between regions), it conforms to the human cognitive logic of temperature fields; the ReLU activation function introduces nonlinearity, enabling the model to learn complex temperature-response relationships (such as nonlinear heat conduction effects), avoiding the expressive limitations of linear models; the second-level convolution can reuse the basic features of the first level (such as temperature gradients) and combine them into higher-level patterns, while reducing redundant information (by increasing the number of convolution kernels rather than simply stacking layers).

[0068] The bidirectional LSTM network of the time feature extraction layer concatenates the hidden layer states as output.

[0069] Bidirectional LSTM structure: The forward LSTM processes the sequential data from t=1 to t=T, capturing the "past → future" dependency. The backward LSTM processes the sequential data from t=T to t=1 in reverse order, capturing the "future → past" dependency.

[0070] The output at each time step t is the concatenation of the forward hidden state ht1 and the backward hidden state ht2: ht = [ht1; ht2]; where the dimension of ht is twice the dimension of the unidirectional LSTM hidden layer.

[0071] In temperature feature extraction, bidirectional LSTM can simultaneously focus on: the impact of past temperature changes on the current state (e.g., continuous temperature increases predict accelerated reactions) and the feedback of future temperature trends on current decisions (e.g., predicting upcoming temperature peaks).

[0072] Compared to unidirectional LSTM, bidirectional structures can capture a longer range of temporal dependencies (such as temperature changes in the later stages of a reaction that may be related to the uniformity of stirring in the earlier stages), solving the problem of "accurately determining the reaction endpoint" in the background technology; by fusing bidirectional information, the model's sensitivity to temperature abrupt changes (such as a sudden drop in temperature after solvent addition) is improved, reducing prediction latency (meeting the requirement of "real-time detection"); the spliced ​​hidden states contain bidirectional temporal features, providing more comprehensive contextual information for subsequent attention mechanisms (such as simultaneously considering "temperature rising trend" and "the temperature threshold to be reached").

[0073] The spatiotemporal attention fusion layer calculates the attention score matrix of Query, Key and Value, and applies the Softmax function to generate attention weights, and performs weighted aggregation on Value.

[0074] Query, Key, Value Mapping: The hidden state ht output by the bidirectional LSTM is transformed into three matrices through a linear transformation: Q = htW q ,K=htW k V=htW v Among them, W q W k W v This is a learnable weight matrix.

[0075] Attention score calculation: The similarity between the query and the key is calculated using the dot product: Attention(Q,K,V) = Softmax(QK) T / (d k ) 1 / 2 V; where d k The dimension of the key is used for scaling to prevent gradient vanishing; Softmax ensures that the sum of the weights is 1.

[0076] Weighted aggregation: Attention weight matrix Softmax(QK) T / (d k ) 1 / 2 Weight the Values ​​to highlight key time steps and locations: Output = ∑ y x=1Weight x ×V x Among them, weight x The x-th time step / position represents the importance of the x-th time step / position, and y represents the number of time steps / positions.

[0077] The attention mechanism automatically identifies the most critical temperature features for the reaction state (such as temperature near the reaction center and temperature fluctuations at the moment of viscosity change) and suppresses irrelevant information (such as random noise at the edge). Through score matrix calculation, the model can establish long-distance spatiotemporal dependencies (such as the correlation between "temperature change at a certain location 30 minutes ago" and "current solvent addition requirements"), solving the problem of long-sequence dependency decay in traditional LSTM. Compared with fixed-weight feature fusion, the attention mechanism enables the model to adaptively adjust its focus at different reaction stages (such as the initial, middle, and late stages) (such as focusing on the heating rate in the initial stage and on temperature stability in the later stage), which is consistent with the dynamic characteristics of polyurethane prepolymer reaction.

[0078] A hierarchical structure for extracting spatial features is achieved through multi-level convolution and ReLU activation; bidirectional LSTM concatenation captures complete temporal dependencies; and attention mechanism weighted aggregation focuses on key contexts. These three elements work together to achieve the following: Spatial sensitivity: accurately identifies areas of uneven temperature within the mixing device (such as localized overheating).

[0079] Predictive timing: Predicting temperature change trends in advance (such as the temperature inflection point when the reaction is about to reach its end).

[0080] Dynamic adaptability: Adaptively adjusts the focus on different spatiotemporal regions, improves the accuracy of solvent addition and reaction endpoint prediction, and ultimately solves the core problem of "overreaction caused by real-time detection delay" in the background technology.

[0081] S500, the first feature vector and the second feature vector are fused to obtain the fused feature vector corresponding to the stirring device.

[0082] Furthermore, step S500 may include the following steps: S510, Obtain the first feature vector A = (A1, A2, ..., A... i A n ), i=1,2,…,n; and the second eigenvector B=(B1,B2,…,B j B m ), j=1,2,…,m; where, A i This refers to the i-th feature value obtained by performing time-series feature extraction on the torque and motor output power time-series data corresponding to the acquired stirring device, where n is the number of feature values ​​obtained; B j Let be the j-th eigenvalue obtained by feature extraction from the temperature matrix, and m be the number of eigenvalues ​​obtained.

[0083] S520, based on A and B, obtain the fused feature vector C = (A1, A2, ..., A...). i A n B1, B2, ..., B j B m ).

[0084] Through the above steps, the first and second feature vectors are concatenated to obtain a fused feature vector, achieving complementary advantages of multi-dimensional physical parameters and significantly improving the ability to characterize the reaction state. This method organically combines mechanical load characteristics with thermodynamic characteristics, capturing both the temporal evolution of the reaction process (such as viscosity change trends) and the spatial distribution and interaction of temperature (such as local hotspot diffusion), fully covering the key physicochemical processes in the preparation of polyurethane prepolymers. The fused feature vector retains all the information of the original features, avoiding information loss due to feature selection, while enhancing anti-interference capabilities through a multi-feature voting mechanism and reducing the impact of single sensor anomalies on prediction results. In addition, the fixed-length vector format provides a unified input for downstream prediction models, facilitating the design of efficient solvent addition confidence models and reaction endpoint time prediction models. This fusion strategy directly solves the problem of "over-reaction caused by real-time detection delay" in the background technology; by capturing the correlation signals between temperature distribution changes and mechanical parameter fluctuations in advance (such as the periodic change of torque accompanied by an increase in temperature gradient), the model can make predictions before significant changes in the reaction state, gaining a time window for process adjustment, and ultimately improving the preparation quality and stability of polyurethane anti-corrosion coatings.

[0085] S600 inputs the fused feature vector into a preset solvent addition confidence prediction model to output the confidence of the solvent to be added within a preset future time period corresponding to the current sliding time window.

[0086] The fused feature vector (containing both temporal features of torque and power and spatiotemporal features of temperature) is input into the solvent addition confidence prediction model, which outputs a confidence value between 0 and 1, representing the probability that solvent needs to be added within the future time period corresponding to the current sliding time window (e.g., in the next 5 minutes). This confidence value reflects the model's judgment on whether the current reaction state is close to requiring solvent dilution to maintain uniform stirring. The core objective is to predict solvent addition needs in advance to avoid local overheating or uneven reaction due to excessive viscosity.

[0087] Training method for solvent addition confidence prediction model: 1. Data preparation stage: Training data collection: During the preparation of polyurethane prepolymer, the following data were collected at fixed time intervals (e.g., every minute): the torque of the stirring device and the output power of the motor (used to generate the first feature vector).

[0088] Temperature sensor data at different locations (used to generate the second feature vector).

[0089] Real label: Record the time and amount of solvent added each time it is manually added, and convert it into a binary label (e.g., 1 for the first 10 minutes of addition and 0 for the rest of the time) or a continuous value (e.g., countdown to the next addition).

[0090] Feature engineering: Based on the S100-S500 method, the raw sensor data is converted into a fused feature vector. The sliding time window length needs to be adjusted according to the response characteristics (e.g., 20 minutes) to ensure that sufficient historical information is included.

[0091] A hybrid architecture combining a time-series prediction model (such as LSTM / GRU) and an attention mechanism can be adopted. LSTM / GRU: captures the time-series dependencies in the fused feature vector (such as the viscosity increase trend and temperature change cycle); Attention mechanism: automatically focuses on key features (such as torque mutation points and temperature gradients in local high-temperature areas); Sigmoid activation: maps the output to 0-1 confidence values, which conforms to the probabilistic interpretation.

[0092] 2. Training process: Objective function: Uses binary cross-entropy loss.

[0093] Optimization strategy: Data augmentation: Randomly shift the sliding time window to increase the diversity of training samples.

[0094] Early stopping mechanism: Monitor the F1 score or AUC-ROC on the validation set to prevent overfitting.

[0095] Learning rate scheduling: Use ReduceLROnPlateau to dynamically reduce the learning rate and improve convergence stability.

[0096] 3. Model Evaluation and Deployment Evaluation indicators: AUC-ROC: Measures the classification performance of a model at different confidence thresholds.

[0097] Precision-Recall Curve: Focus on high recall (situations where solvent needs to be added to avoid missed detections).

[0098] False alarm rate: Keep it at a low level to avoid increasing costs due to frequent solvent additions.

[0099] Deployment optimization: Online learning: Regularly fine-tune the model using new data to adapt to changes in the production environment.

[0100] Ensemble learning: Integrating the prediction results of multiple models (such as random forest and XGBoost) to improve robustness.

[0101] This model has the following advantages: Early prediction capability: By capturing the combined trends of torque, power and temperature changes, the model can predict solvent demand before viscosity increases significantly, solving the problem of "overreaction due to detection delay" in the background technology.

[0102] Multi-feature synergy: Integrating mechanical load and temperature distribution characteristics to avoid misjudgment by a single sensor (e.g., a temperature rise may be a normal exothermic reaction or local overheating, which can be distinguished by combining torque changes).

[0103] Interpretability: The attention mechanism visualizes the key features that the model focuses on (such as the temperature gradient at a certain location), making it easier for engineers to understand the basis of the prediction and optimize process parameters.

[0104] This model allows the timing of solvent addition during the preparation of polyurethane prepolymers to be transformed from "experience-based judgment" to "data-driven precise prediction," thereby improving product quality stability.

[0105] Furthermore, the step size of the current sliding time window is determined through the following steps: S610, according to the chronological order, sequentially obtain the confidence lists Z = (Z1, Z2, ..., Zn) corresponding to the first q sliding time windows adjacent to the current sliding time window. p , ..., Z q ), p=1,2,…,q; where Z p This represents the confidence level corresponding to the p-th sliding time window among the q sliding time windows adjacent to the current sliding time window.

[0106] S620, if the confidence level in Z shows a decreasing trend, then determine the step size QT corresponding to the current sliding time window. now =α×QT1; where α is the preset adjustment coefficient for the first future time period, α>1; QT1 is the step size of the previous sliding time window adjacent to the current sliding time window.

[0107] During the stable reaction phase, high-frequency monitoring is unnecessary. Increasing the step size reduces computational resource consumption while maintaining trend tracking. In the later stages of the reaction (when solvent demand decreases), increasing the step size reduces redundant computation while ensuring the window still covers key trend changes (e.g., increasing the step size from 5 minutes to 6 minutes still captures the continuous decline in confidence).

[0108] S630, if the confidence level in Z shows an upward trend, then determine the step size QT corresponding to the current sliding time window. now =β×QT1; where β is the preset adjustment coefficient for the second future time period, and β<1.

[0109] During the active reaction period, the monitoring frequency needs to be increased, and subtle changes should be captured by reducing the step size to avoid missing the opportunity to add solvent.

[0110] During the active reaction period (when solvent demand increases), reducing the step size can improve time resolution (e.g., from a 5-minute step size to a 4-minute step size), allowing for more accurate tracking of viscosity changes and avoiding delays in solvent addition due to excessively large step sizes.

[0111] By following the steps above, the step size can be increased when the reaction is stable, thereby reducing the frequency of data acquisition and model inference (e.g., reducing the amount of computation by 20% per day) and saving system resources.

[0112] Focusing on key stages: Reduce the step size when the reaction fluctuates to ensure sensitive capture of solvent addition requirements (e.g., high-frequency monitoring can provide early warning 3-5 minutes in advance when viscosity changes abruptly), thus resolving the contradiction of "redundancy in the stable period and insufficient in the fluctuating period" when the step size is fixed.

[0113] S700, if several consecutive confidence scores are less than the preset confidence threshold and show a downward trend, the fused feature vector is input into the preset reaction endpoint time prediction model to obtain the predicted reaction endpoint time.

[0114] In this embodiment, several recent consecutive confidence levels can be obtained. Each confidence level is compared with a preset confidence threshold (e.g., 0.3, set according to historical data and process requirements). If all confidence levels are less than the threshold, it indicates that no solvent needs to be added in the current and near future. The direction of change of the consecutive confidence levels is analyzed by linear fitting or difference method. If the confidence level of the next window is less than that of the previous window, it is determined to be a downward trend, indicating that the solvent demand is continuously decreasing and the reaction is stabilizing.

[0115] When the above conditions are met, the fusion feature vector (containing the temporal features of torque and power and the spatiotemporal features of temperature) corresponding to the current sliding time window is input into the preset reaction endpoint time prediction model; the model outputs a specific future time point (such as "current time + 40 minutes") or time period (such as "within the next 30-50 minutes"), which is the predicted reaction endpoint time (the moment when the reaction is completed).

[0116] The training method for the reaction endpoint time prediction model is as follows: I. Training Data Preparation 1. Data collection and annotation: Input feature data: Collect the fused feature vector of each complete reaction cycle in historical production, extract it according to the sliding time window (e.g., one window every 5 minutes), and each window corresponds to one feature sample.

[0117] Label data (reaction endpoint time): Record the actual reaction endpoint time in each reaction cycle (labeled as "duration from the end of the current sliding time window to the completion of the reaction". For example, if the current window ends 100 minutes after the start of the reaction and the actual endpoint is 150 minutes, then the label is 50 minutes).

[0118] Data screening: Only samples from the stage where "solvent addition confidence continuously decreases and falls below the threshold" are retained (consistent with the triggering condition of S700) to ensure that the training data matches the model application scenario (only samples from the later stages of the reaction are used for training).

[0119] 2. Data preprocessing: Outlier handling: Samples with abnormal labels (such as abnormal fluctuations in endpoint time due to equipment failure) are removed by IQR (interquartile range) method, and outliers in the feature vector are replaced by the mean of adjacent windows.

[0120] Normalization: Min-Max normalization (mapping to the 0-1 interval) is performed on the features of each dimension of the fused feature vector (such as the mean torque, temperature gradient, etc.) to eliminate the impact of dimensional differences on model training.

[0121] Dataset partitioning: The dataset is divided into a training set (for model parameter learning), a validation set (for hyperparameter tuning), and a test set (for final performance evaluation) in a 7:2:1 ratio.

[0122] II. Model Architecture Design The time-series regression model, combined with the dynamic characteristics and long-term dependencies of the reaction process, is structured as follows: Base network: Bidirectional LSTM is selected as the backbone network because it can capture the temporal evolution pattern in the fused feature vector (such as the decay trend of torque change rate over time and the stabilization process of temperature field).

[0123] Attention Mechanism: A temporal attention mechanism is introduced into the LSTM output layer, enabling the model to automatically focus on feature windows that are key to the prediction of the endpoint (such as the stage where torque fluctuations narrow in the later stages).

[0124] Output layer: The predicted endpoint time (regression task) is output through a fully connected layer and a linear activation function, ensuring that the output is a continuous time value (unit: minutes).

[0125] Architectural advantages: Bidirectional LSTM solves the long-term dependence of reaction features (such as the impact of early mixing uniformity on the later endpoint), and the attention mechanism strengthens the weight of key stage features, improving prediction accuracy.

[0126] III. Training Process 1. Loss Function and Optimizer: Loss function: Mean squared error (MSE) loss is used.

[0127] Optimizer: The Adam optimizer is used with an initial learning rate of 0.001. The convergence stability is improved by using a learning rate decay strategy (e.g., decaying to 0.5 every 10 epochs).

[0128] 2. Training Strategies Batch training: Input the training set into the model in batches (e.g., 32 samples per batch) and update the network parameters through backpropagation.

[0129] Early stopping mechanism: Monitor the MSE of the validation set, and stop training when the loss on the validation set no longer decreases after 5 consecutive epochs to avoid overfitting (overfitting will cause the model to increase the prediction error on new data).

[0130] Data augmentation: Temporal perturbation of training samples (e.g., randomly shortening or extending the sliding window time interval by ±1 minute) increases sample diversity and improves the model's generalization ability.

[0131] IV. Model Evaluation and Optimization 1. Evaluation indicators: Key metrics: The prediction accuracy is evaluated using mean absolute error (MAE) and root mean square error (RMSE). MAE reflects the average prediction bias (unit: minutes), while RMSE is more sensitive to large errors, ensuring the model's prediction stability for extreme cases.

[0132] Auxiliary indicator: Calculate the Pearson correlation coefficient between the predicted value and the actual value to verify the consistency between the model's predicted trend and the actual response process (the closer the correlation coefficient is to 1, the higher the trend consistency).

[0133] 2. Optimization direction: Hyperparameter tuning: Optimize the hidden layer dimension (e.g., 64, 128), number of attention heads, learning rate, etc. of LSTM through grid search to reduce validation set loss.

[0134] Feature importance analysis: Identify the features that contribute most to the endpoint prediction (such as "torque fluctuation variance" and "temperature field uniformity index") by visualizing SHAP values ​​or weights, and enhance the acquisition accuracy of key features (such as increasing the sampling frequency of the corresponding sensors).

[0135] V. Model Deployment and Iteration Online fine-tuning: After the model is deployed, new production data (including actual end time) is periodically added to the training set, and the model parameters are updated through incremental learning to adapt to slow changes in process conditions (such as raw material batch differences).

[0136] Threshold calibration: Adjust the confidence interval of the model output according to the actual prediction effect (e.g., the prediction endpoint time ±5 minutes is a confidence range) to provide a flexible reference for the detection operation of S800.

[0137] Furthermore, after step S700 and before step S800, the method further includes the following steps: S710, if the confidence level corresponding to the current sliding time window is greater than or equal to the preset confidence threshold, then add solvent within the preset future time period corresponding to the current sliding time window; proceed to S100.

[0138] In this embodiment, if the confidence level corresponding to the current sliding time window is greater than or equal to the preset confidence threshold, it means that solvent needs to be added within the preset future time period corresponding to the current sliding time window. Solvent can be added at the corresponding time, and then proceed to S100.

[0139] S800, based on the reaction endpoint time, detects the polyurethane prepolymer during stirring to determine if the polyurethane prepolymer reaction is complete.

[0140] Furthermore, step S800 includes the following steps: S810, if the current time is less than or equal to the time until the reaction end point, the polyurethane prepolymer being stirred is detected.

[0141] S820, if the test results meet the preset conditions, it is determined that the polyurethane prepolymer reaction is complete.

[0142] Based on the predicted reaction endpoint time output by the S700, set the detection start window in advance (e.g., 5-10 minutes before the predicted endpoint time) to ensure that the detection starts when the reaction is close to the endpoint but before the reaction has ended. For example, if the predicted endpoint time is 180 minutes after the reaction starts, start the first detection at 170-175 minutes to avoid detection too early (before the reaction has reached the endpoint) or detection too late (the reaction may have ended).

[0143] If the predicted endpoint time fluctuates (e.g., 175-185 minutes), continuous detection (e.g., once every 2 minutes) is initiated at the beginning of that range until the reaction is confirmed to be complete, covering the possible endpoint range.

[0144] By predicting the endpoint time to guide the detection, compared with the traditional "random detection" or "experience judgment", the reaction can be terminated in time when the NCO content and viscosity meet the standards, avoiding excessive consumption of NCO (leading to insufficient crosslinking density of the coating) or abnormal increase in viscosity (affecting subsequent coating formulation).

[0145] To ensure the accuracy of the test, combined with the verification of temperature field uniformity, we can eliminate "false endpoints" caused by incomplete local reactions (such as local NCO content meeting the standard but the overall reaction not being complete), and ensure that the test results reflect the overall state of the prepolymer, thus ensuring that the quality of the prepared polyurethane anti-corrosion coating meets the requirements.

[0146] By combining the strategy of "predicting the endpoint + continuous detection", the reaction can be confirmed to be complete in as little as 10 minutes, which significantly shortens the process time compared to the traditional "fixed interval detection" (which may take more than 30 minutes).

[0147] In this embodiment, by acquiring the torque, motor output power, and temperature at different positions within a preset sliding time window in real time during the stirring reaction of the polyurethane prepolymer, and after feature extraction and fusion, the confidence level of solvent addition and the reaction endpoint time are predicted using a preset model. This allows for the early determination of solvent addition requirements and the reaction endpoint, effectively avoiding the delay problem of real-time detection methods in the prior art, reducing the impact of over-reaction on the quality of the polyurethane prepolymer, and thus ensuring the quality of the final polyurethane anticorrosive coating.

[0148] The method of this invention can predict the reaction endpoint time of polyurethane prepolymer. Therefore, when there is a long time before the reaction endpoint, it is not necessary to continuously monitor the polyurethane prepolymer in the reaction. This simplifies the operation process and improves the preparation efficiency while ensuring the quality of the prepared polyurethane anticorrosive coating.

[0149] The embodiments disclosed herein are preferred embodiments, but are not limited thereto. Those skilled in the art can readily grasp the spirit of the present invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A method for preparing a polyurethane anti-corrosion coating, characterized in that, The method includes the following steps: S100: During the stirring reaction of polyurethane prepolymer, the torque, motor output power, and temperature at different locations within the stirring device are obtained at each sampling moment within a preset sliding time window; the end time of the preset sliding time window is the current time. S200, extract time-series features from the acquired torque and motor output power time-series data corresponding to the stirring device to obtain the first feature vector corresponding to the stirring device; S300: Based on the temperature at different locations within the stirring device at each sampling moment within a preset sliding time window, establish a temperature matrix corresponding to the stirring device. The temperature matrix includes several rows and several columns, with each row corresponding to the temperature at different sampling moments at the same location and each column corresponding to the temperature at different locations at the same sampling moment. S400, extract features from the temperature matrix to obtain the second feature vector corresponding to the stirring device; S500, the first feature vector and the second feature vector are fused to obtain the fused feature vector corresponding to the stirring device; S600 inputs the fused feature vector into a preset solvent addition confidence prediction model to output the confidence of the solvent to be added in the preset future time period corresponding to the current sliding time window; S700, if several consecutive confidence scores are all less than the preset confidence threshold and show a downward trend, then the fused feature vector is input into the preset reaction endpoint time prediction model to obtain the predicted reaction endpoint time. S710, if the confidence level corresponding to the current sliding time window is greater than or equal to the preset confidence threshold, then add solvent within the preset future time period corresponding to the current sliding time window; proceed to S100; S800, based on the reaction endpoint time, detects the polyurethane prepolymer during stirring to determine if the polyurethane prepolymer reaction is complete.

2. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, Step S200 includes the following steps: S210, preprocess the torque timing data and motor output power timing data within the preset sliding time window, the preprocessing including outlier handling and missing value filling; S220 extracts statistical features, trend features, and fluctuation features from the preprocessed torque time series data and motor output power time series data, respectively. S230, the statistical features, trend features, and fluctuation features corresponding to the torque time series data are sequentially concatenated with the statistical features, trend features, and fluctuation features corresponding to the motor output power time series data to form the first feature vector.

3. The method for preparing the polyurethane anti-corrosion coating according to claim 2, characterized in that, The statistical features include: mean, median, variance, standard deviation, range, skewness, kurtosis, maximum value, minimum value, and percentage of peak value; the trend features include: linear trend slope, trend fit goodness of fit, and cumulative change; the fluctuation features include: mean of first-order differences, standard deviation of first-order differences, root mean square difference of first-order differences, and percentage of positive change.

4. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, Step S400 includes the following steps: S410, constructs a spatial feature extraction layer for the temperature matrix, uses a one-dimensional convolution kernel to perform convolution operation on the temperature matrix, and extracts the temperature correlation features between different locations at the same time. S420 constructs a time feature extraction layer for the temperature matrix, and uses a bidirectional LSTM network to process the output of the spatial feature extraction layer to capture the temperature change trend features at different times. S430, construct a spatiotemporal attention fusion layer, apply a self-attention mechanism to the output of the temporal feature extraction layer, calculate the contextual dependencies between different times and locations, and generate weighted spatiotemporal features; S440 constructs a feature vector generation layer, performs global average pooling on the output of the spatiotemporal attention fusion layer, and maps it to a second feature vector of fixed dimension through a fully connected layer.

5. The method for preparing the polyurethane anti-corrosion coating according to claim 4, characterized in that, The convolution operation of the spatial feature extraction layer includes at least two one-dimensional convolutions, with ReLU activation performed after each convolution. The bidirectional LSTM network of the time feature extraction layer concatenates the hidden layer states as output. The spatiotemporal attention fusion layer calculates the attention score matrix of Query, Key and Value, and applies the Softmax function to generate attention weights, and performs weighted aggregation on Value.

6. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, Step S500 includes the following steps: S510, Obtain the first feature vector A = (A1, A2, ..., A... i A n ), i=1,2,…,n; and the second eigenvector B=(B1,B2,…,B j B m ), j=1,2,…,m; where, A i This refers to the i-th feature value obtained by performing time-series feature extraction on the torque and motor output power time-series data corresponding to the acquired stirring device, where n is the number of feature values ​​obtained; B j Let m be the j-th eigenvalue obtained by feature extraction from the temperature matrix, and m be the number of eigenvalues ​​obtained. S520, based on A and B, obtain the fused feature vector C = (A1, A2, ..., A...). i A n B1, B2, ..., B j B m ).

7. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, The step size of the current sliding time window is determined by the following steps: S610, according to the chronological order, sequentially obtain the confidence lists Z = (Z1, Z2, ..., Zn) corresponding to the first q sliding time windows adjacent to the current sliding time window. p , ..., Z q ), p=1,2,…,q; where Z p The confidence level is the value of the p-th sliding time window among the q sliding time windows adjacent to the current sliding time window. S620, if the confidence level in Z shows a decreasing trend, then determine the step size QT corresponding to the current sliding time window. now =α×QT1; where α is the preset adjustment coefficient for the first future time period, α>1; QT1 is the step size of the previous sliding time window adjacent to the current sliding time window; S630, if the confidence level in Z shows an upward trend, then determine the step size QT corresponding to the current sliding time window. now =β×QT1; where β is the preset adjustment coefficient for the second future time period, and β<1.

8. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, Step S800 includes the following steps: S810, if the current time is less than or equal to the time until the reaction end point, then the polyurethane prepolymer in the stirring is detected; S820, if the test results meet the preset conditions, it is determined that the polyurethane prepolymer reaction is complete.

9. The method for preparing the polyurethane anti-corrosion coating according to claim 1, characterized in that, The temperature at different locations within the stirring device is obtained through temperature sensors installed at corresponding locations.

Images