A double-layer time series pair regression cement f-calcium oxide prediction method based on TCN-GRU

Abstract

This invention discloses a two-layer temporal contrastive regression method for predicting f-CaO content in cement based on TCN-GRU. The method includes collecting multi-source data from the cement clinker calcination process, using labeled samples as anchors, and reconstructing unlabeled data into temporal data blocks using a sliding window. Local temporal features are extracted using a temporal convolutional network and obtained through pooling to obtain local feature vectors. Weights are dynamically assigned based on sample time intervals, and noise is suppressed and feature alignment is achieved through temporal weighted contrastive learning. Long-term global features are extracted using gated recurrent units. Local and global features are concatenated after linear projection and dynamically fused through a gating mechanism to generate fused features for f-CaO content prediction. Training is driven collaboratively by prediction regression loss and temporal weighted contrastive loss, introducing time decay constraints in the low-level features to suppress noise and improve feature quality, ultimately achieving highly robust prediction. This method has high data utilization and strong anti-interference capabilities, achieving a determination coefficient R² of 0.7970 on real datasets, significantly outperforming existing algorithms.

Description

Technical Field

[0001] This invention relates to the field of industrial soft measurement technology, specifically to a two-layer time-series comparative regression method for predicting cement f-Ca0 based on TCN-GRU. Background Technology

[0002] Cement, as a core material for modern infrastructure, is indispensable in national defense, water conservancy, and industrial construction. Clinker calcination is a crucial step in cement production, and the content of free calcium oxide (f-CaO) is the most critical indicator for measuring clinker quality and production energy consumption. Excessive f-CaO content can lead to concrete cracking and reduced durability; conversely, insufficient f-CaO content indicates over-burning of the clinker, which not only significantly increases energy consumption but also reduces the early strength of the cement. Therefore, accurate prediction of f-CaO content is key to achieving closed-loop control of cement production.

[0003] Currently, the detection of f-CaO content at cement production sites still mainly relies on traditional methods combining manual sampling with offline physicochemical analysis. This method has a detection cycle of 1-2 hours, and data feedback lags significantly behind production status. As a result, by the time an anomaly is detected, a large amount of substandard clinker has already been generated, failing to meet the requirements of real-time closed-loop control.

[0004] To address the time lag challenge in offline detection, soft measurement methods have become a current research hotspot. Existing soft measurement methods mainly include traditional machine learning methods (such as SVM and Random Forest) and deep learning methods (such as CNN and LSTM). However, in complex cement industry scenarios, existing methods have significant limitations:

[0005] Label sparsity and temporal discontinuity issues: Industrial distributed control systems (DCS) can generate a large amount of operating data in real time, but the true value of f-CaO is sparse. Existing multivariate time series forecasting methods mostly adopt the strategy of truncating the time stream into independent data blocks. This discretization process severs the long-term temporal correlation across data blocks in cement calcination, a process with strong thermal inertia, leading to a decrease in forecasting performance.

[0006] Environmental noise and dynamic changes under complex operating conditions: Sensors are susceptible to high-frequency random noise caused by dust and high temperatures. Meanwhile, the influence weights of different process parameters on f-CaO exhibit significant time-varying characteristics as production conditions change, making it difficult for traditional static models to adaptively adjust.

[0007] In summary, existing soft-sensor models struggle to balance local temporal dynamics and global long-range thermal inertia under sparse labeling, making them susceptible to noise interference and resulting in poor prediction stability during long-term operation. Therefore, designing an f-CaO content prediction method that can effectively establish long-range operational condition dependencies, suppress environmental noise, and adaptively fuse features based on operational condition dynamics has become a pressing technical problem in this field. Summary of the Invention

[0008] The purpose of this invention is to design a two-layer time-series contrastive regression method for predicting cement f-CaO content based on TCN-GRU. This method can improve the prediction accuracy and long-term robustness of cement clinker f-CaO content by effectively fusing local dynamic features and global evolution features of the data, overcoming label sparsity and high-frequency noise interference.

[0009] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution: a two-layer time-series contrastive regression method for predicting cement f-Ca0 based on TCN-GRU, characterized by comprising the following steps:

[0010] S1. Multi-source data acquisition and matrix reconstruction: Sensor data of cement clinker calcination process is acquired through distributed control system and sparse test labels are obtained simultaneously. Labeled samples are used as anchor points and sliding windows are used to reconstruct unlabeled data into two-dimensional time-series data blocks.

[0011] S2. Local Temporal Feature Extraction: A Temporal Convolutional Network (TCN) is used to extract local temporal features from data blocks, and these features are then mapped to a local feature vector hl using global average pooling to unify the dimensionality. t ;

[0012] S3, Temporal Weighted Feature Alignment: Utilizing a temporal weighted contrastive learning mechanism, consistency weights are dynamically allocated based on sample time intervals to achieve progressive smoothing and noise suppression across sample feature spaces;

[0013] S4. Long-span feature extraction: A gated recurrent unit (GRU) is used to recursively evolve and model the local feature vectors of each data block, extracting the global feature vector hg representing long-term temporal dependencies across data blocks. t ;

[0014] S5, Gated Fusion: The local feature vector h extracted by TCN is... l,t Compared with the global feature vector hg extracted by GRU, t Each feature vector is mapped to the same dimensional space via a linear projection layer, and then concatenated along the channel dimension to obtain the joint feature vector u. t Subsequently, through a linear mapping layer and a Sigmoid activation function, from u t The fusion weight G is calculated in the middle. t The expression is as follows:

[0015]

[0016] In the formula, W G With b G G represents the weight matrix and bias vector to be learned in the gated layer, respectively; t This characterizes the proportion of global dynamic dependencies in feature fusion, while 1−Gt This corresponds to the proportion of local details;

[0017] To achieve soft selection of feature information, a complementary weighting strategy is used to perform a weighted summation of the two types of features, as shown in the following expression:

[0018]

[0019] In the formula, h′ g,t is the global eigenvector after linear mapping, h′ l,t is the local eigenvector after linear mapping, and ⊙ represents the element-wise multiplication of vectors;

[0020] Under this mapping mechanism, the resulting fused feature vector hf, t It can dynamically adjust the ratio of local feature sensitivity to global trend stability based on real-time input status; ultimately, hf, t It is fed into the regression layer to calculate the predicted value of f-CaO content.

[0021] S6. Collaborative Loss Regression Prediction: The training and optimization of the model are driven collaboratively by two complementary loss terms. The prediction regression loss acts on the multi-scale joint features after gated fusion, outputting the predicted f-CaO concentration value through the regression decoding layer and constraining its deviation from the actual test value. The time-weighted contrast loss acts on the local feature layer extracted by TCN. By introducing an exponentially decaying weight constraint based on time distance in the bottom feature space, it forces the local feature representation of temporally adjacent samples to maintain consistency, thereby suppressing high-frequency noise interference from the sensor at the source of feature extraction and improving the feature quality of the input to the subsequent GRU and gated fusion modules. The two losses are weighted and summed through the hyperparameter λ to form the total loss function, realizing end-to-end collaborative optimization and outputting a highly robust final predicted value of cement clinker f-CaO concentration.

[0022] Furthermore, in S1, the distributed control system collects sensor data of the cement clinker calcination process and simultaneously acquires sparse test labels. Specifically, it acquires key process sensor data such as preheater outlet temperature, decomposition furnace pressure, rotary kiln speed, and pulverized coal flow rate recorded by the cement kiln system, and at the same time acquires the true value of f-CaO concentration fed back by the offline laboratory as a sparse label.

[0023] Furthermore, in S2, the TCN employs a two-layer structure: First, a 1×1 convolution layer maps the original 10 variable channels to 48 feature channels. Then, a first-layer causal dilated convolution and residual connections extract local temporal patterns, maintaining the output channel count at 48. The second layer further compresses the channel count to 24, expanding the receptive field while reducing the feature dimension. With appropriate causal padding, each convolution layer does not change the time step length. Therefore, the final feature map output by the TCN contains C=24 channels, each corresponding to a feature sequence of length T′=W=20, expressed as follows:

[0024]

[0025] In the formula, θ represents the set of all learnable weights and biases within the network; the feature map H l,t It is a two-dimensional array with dimensions T′ rows and C columns, where T′ is the time step length after convolution and pooling, C is the number of channels, and T′=20, C=24; This feature map characterizes the dynamic evolution mode of process parameters within the window during the calcination of cement clinker.

[0026] Subsequently, to obtain local feature vectors, global average pooling is performed along the time dimension, as shown in the following expression:

[0027]

[0028] The pooled vector h l,t It is a C-dimensional vector that retains the main information of the feature map while eliminating redundant dimensions; through this continuous mapping, the original two-dimensional data block sequence is reconstructed into a set of feature vectors H={h} with temporal logic. l,1 ,h l,2 ,...,h l,T}

[0029] Furthermore, in S3, the time-series weighted feature alignment is achieved by using an exponential decay function to calculate the temporal continuity similarity weight based on the sample time interval span, mapping and transforming it to a low-dimensional embedding through a nonlinear projection module to control the feature components, and achieving smooth alignment of the continuous state condition feature space through the coordinated action of the clustering loss of neighboring positive samples in the numerator and the pull-repulsion loss of global batch samples in the denominator.

[0030] Furthermore, in S4, the gated recurrent unit (GRU) is used to recursively evolve and model the local feature vectors of each data block to extract global long-range dependency features across data blocks. Specifically, based on the local feature sequence extracted by the TCN, the GRU is used to further model the local feature sequence extracted by the TCN to learn the temporal dependency relationship across data blocks. Here, the GRU simplifies the structure of the traditional recurrent neural network by integrating the forget gate and the input gate into a single update gate, thereby alleviating the gradient vanishing problem in deep networks while reducing the number of model parameters and improving convergence performance.

[0031] In the specific recursive update process, the model first updates the global hidden state hg from the previous time step. t-1 The local feature vector h input at the current time l,t The update gate z is calculated. t With Reset Gate r t The gate value is expressed as follows:

[0032]

[0033]

[0034] Subsequently, using the reset gate r t Filter historical state information and combine it with the input features at the current time to calculate candidate hidden states h̃ g,t :

[0035]

[0036] Finally, by updating gate z t The historical background information and the current candidate state are weighted and fused to output the final global feature vector hg at that moment. t :

[0037]

[0038] In the formula, σ and tanh(⋅) are activation functions, ⊙ represents element-wise multiplication, and W and b are the weights and biases to be learned; under this recursive update mechanism, the hidden state hg at the current time step, t Based on the global information hg from the previous time step, t-1 With the local features h at the current time l,t To be determined jointly;

[0039] Through the recursive update mechanism described above, GRU maps the local feature sequence H extracted by TCN to a global feature vector hg containing long-term temporal dependencies. t .

[0040] Furthermore, in S6, the model's total loss function is composed of the prediction regression loss L. reg Compared with time-weighted loss L con Composition, through hyperparameters Adjust the contribution weights of the two; where L reg The prediction accuracy of the joint features after constraint-gated fusion, output by the regression layer, L con Constrained TCN extracts local feature vector h l,t The temporal consistency distribution in the feature space; two losses act on different levels of the model: the regression loss backpropagates from the output to update all network parameters to ensure prediction accuracy, and the contrastive loss directly optimizes the representational ability of the TCN from the bottom feature level to ensure feature quality; the total loss expression is as follows:

[0041]

[0042] In the formula, L total Let L be the total loss function. reg To predict regression loss, L con The loss is a time-weighted comparison loss, where λ is a hyperparameter that adjusts the contribution weights of the two factors.

[0043] Among them, regression loss L reg Mean squared error is selected to calculate the deviation between the predicted and actual f-CaO concentration values. This loss term constrains the prediction accuracy at the output, allowing the model to continuously reduce the prediction error during training. The calculation formula is as follows:

[0044]

[0045] In the formula, N is the number of samples in the training batch, y i With ŷ i The measured data of f-CaO and the network prediction results correspond to the i-th sample, respectively;

[0046] Time-weighted contrast loss L con The local feature vector h extracted by TCN and subjected to global average pooling is applied. l,t Because each h l,t Each corresponds to a compressed representation of a complete data block. This loss adjusts the feature distribution by introducing a soft constraint based on time distance in the feature space.

[0047] Then, by employing a soft-weighted comparison mechanism, positive and negative samples are not explicitly specified, but rather based on the time interval |t between samples. i -t j | Assign a continuous weight w to each pair of samples. ijThe weights decay exponentially with increasing intervals, making it easier to bring temporally adjacent samples closer together in the feature space, while samples with larger intervals have only a weak impact due to their small weights. This design incorporates the temporal smoothness of the cement calcination process as prior knowledge into feature learning, suppressing high-frequency random noise while maintaining responsiveness to changes in real-world conditions. During calculation, the local feature vectors are first mapped to a low-dimensional embedding vector z through a nonlinear projection head. The temporal weighted contrastive loss is then defined as:

[0048]

[0049] In the formula, z i z is the feature vector of anchor point sample i after being mapped by the projection head; j and z k These represent the feature vectors of samples other than i; τ is a temperature parameter used to adjust the smoothness of the similarity distribution.

[0050] Among them, the weighting factor w ij Determined by the time interval between samples, using an exponential decay function:

[0051]

[0052] In the formula, |t i -t j | represents the time interval between sample pairs, and β is the decay coefficient that controls the rate at which the weights decrease with time difference;

[0053] The weighted summation in the numerator enhances the consistency of feature representation of samples with similar time proximity, while the global summation in the denominator plays a normalization and comparison role, preventing the model from mapping all sample features to the same point and losing its ability to distinguish between different working conditions. Through this construction, the model can effectively suppress the interference of high-frequency random noise on feature representation while capturing the continuous evolution characteristics of working conditions.

[0054] The beneficial effects of this invention are:

[0055] The f-CaO soft measurement method for cement clinker provided by this invention effectively solves the problem of time series interruption caused by data slicing in traditional methods through the TCN-GRU dual-layer time series architecture. At the same time, by introducing time series aligned contrastive learning and adaptive gating fusion technology, it effectively alleviates the problem of high-frequency random noise from sensors and endows the model with the ability to cope with complex and variable working conditions.

[0056] While meeting extremely high real-time requirements, this invention achieved a coefficient of determination (R²) of 0.7970 on a real cement production dataset, significantly improving the model's prediction accuracy and usability in industrial settings. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a system structure diagram of a two-layer time-series contrastive regression method for predicting cement f-CaO based on TCN-GRU as described in this invention;

[0059] Figure 2 This is a flowchart illustrating the steps of a two-layer time-series contrastive regression method for predicting cement f-CaO based on TCN-GRU as described in this invention.

[0060] Figure 3 This is a schematic diagram comparing the accuracy of f-CaO content prediction between the TCN-GRU-based two-layer time-series comparative regression method for predicting cement f-CaO and existing comparative models.

[0061] Figure 4 This is a prediction curve of the two-layer time-series contrastive regression method for predicting cement f-CaO based on TCN-GRU according to the present invention.

[0062] Figure 5 This is a data variable table for the cement clinker production process described in this invention. Detailed Implementation

[0063] 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.

[0064] Example 1

[0065] like Figure 1As shown, existing cement f-CaO prediction methods mostly rely on simple shallow regression or single deep learning models, which have obvious limitations: they cannot simultaneously model short-term process fluctuations and long-term thermal inertia changes, and they are prone to overfitting when faced with high-frequency noise and sparse sampling, resulting in significant lag and instability in prediction. This invention proposes a two-layer time-series contrastive regression cement f-CaO prediction method based on TCN-GRU. The system structure of this method includes a data acquisition and reconstruction unit, a local time-series feature extraction unit, a time-series weighted feature alignment unit, a global long-range dependency extraction unit, an adaptive gating fusion unit, and an f-CaO content prediction unit.

[0066] Specifically, the collaborative processing flow of each module in the above system architecture is as follows:

[0067] The data acquisition and reconstruction unit is mainly used to acquire on-site sensor sequences and sparse test labels online and synchronously, and to reconstruct these discrete data into a two-dimensional time series matrix containing local topological information using a sliding window mechanism.

[0068] The local temporal feature extraction unit uses the TCN network to perform in-depth analysis of the dynamic change patterns within the above two-dimensional data block, and then completes the compression of the feature dimension through global average pooling operation.

[0069] To address the high-noise environment in industrial settings, the temporal weighted feature alignment unit introduces a contrastive learning mechanism. By calculating the time span between samples, it assigns exponentially decaying weights, promoting a smooth transition in the feature space and effectively suppressing interference from high-frequency sensor clutter and environmental noise.

[0070] To overcome the information discontinuity caused by discrete slicing, the global long-range dependency extraction unit not only receives the above features, but also performs recursive calculations on the change trajectory of the features through GRU nodes, thereby capturing the long-period thermal inertia and large time delay system laws across data blocks.

[0071] Based on this, the adaptive gating fusion unit dynamically calculates the fusion weight according to the current real-time operating conditions, dynamically allocates between local transient sensitive information and global macro steady-state trend, and outputs multi-level complementary time-series representations.

[0072] The model's training and optimization are driven collaboratively by two complementary loss terms. The prediction regression loss acts on the multi-scale joint features after gated fusion, outputting the predicted f-CaO concentration through the regression decoding layer and constraining its deviation from the actual test value. The time-weighted contrast loss acts on the local feature layer extracted by TCN, forcing the consistency of local feature representations of temporally adjacent samples by introducing time-distance-based exponential decay weight constraints in the bottom feature space. This suppresses high-frequency noise interference from the sensor at the source of feature extraction and improves the feature quality of the input to the subsequent GRU and gated fusion modules. The two losses are weighted and summed through the hyperparameter λ to form the total loss function, achieving end-to-end collaborative optimization and outputting a highly robust final predicted value of cement clinker f-CaO concentration.

[0073] Example 2

[0074] like Figure 2 As shown, a two-layer time-series contrastive regression method for predicting cement f-CaO based on TCN-GRU includes the following steps:

[0075] Step 1: Collect cement production process data and sparse tags through a distributed control system:

[0076] Simultaneously acquire 10 key process sensor data (such as preheater outlet temperature, decomposer pressure, rotary kiln speed, and pulverized coal flow rate) recorded by the cement kiln system. Figure 5 As shown in the figure, the true value of f-CaO concentration fed back by the offline laboratory is obtained as a sparse label.

[0077] Step 2: Structured preprocessing of multi-source sample matrix.

[0078] To address the discontinuous sampling of the quality index f-CaO and the strong thermal inertia of the process during cement clinker calcination, this paper establishes a matrix-structured mapping mechanism based on sliding window constraints to reconstruct features from multi-source process data. Mathematically, let the process variable vector at time t be x. t It is a 10-dimensional vector composed of 10 selected key process parameters. Due to the quality label y... t The acquisition of data relies on offline testing, and its distribution on the time axis is much sparser than that of process variables. To compensate for the discontinuity of time-series information caused by discrete sampling, this paper uses each labeled time t as an anchor point and, by defining a sliding window of length W, stacks all observation vectors within the interval [t-W+1,t] to construct a two-dimensional feature matrix X. t =[x t-W+1 ,…,x t-1 ,x t ], and the corresponding test value y at that time. tThis serves as its label. The essence of this matrix representation is to map scattered test points into time-series blocks with local topological characteristics, using the originally free process monitoring data to explicitly describe the dynamic changes in the operating conditions.

[0079] Step 3: Use Temporal Convolutional Network (TCN) to extract temporal features of local dynamic characteristics.

[0080] In the feature extraction stage, this paper utilizes a temporal convolutional network to extract features from the reconstructed two-dimensional data block X. t Local dynamic modeling is performed. Following a matrix strategy, each data block is composed of stacked process variables within a sliding window, with dimensions of 10 rows and W columns. The TCN employs a two-layer structure: first, a 1×1 convolution maps the original 10 variable channels to 48 feature channels; then, a first-layer causal dilated convolution and residual connections extract local temporal patterns, maintaining the output channel count at 48; the second layer further compresses the channel count to 24, expanding the receptive field while reducing the feature dimension. With appropriate causal padding, each convolution layer does not change the time step length; therefore, the final feature map output by the TCN contains C=24 channels, each corresponding to a feature sequence of length T′=W=20, i.e.:

[0081]

[0082] In the formula, θ represents the set of all learnable weights and bias parameters within the network. Feature map H l,t It is a two-dimensional array with dimensions T′ rows and C columns, where T′ is the time step length after convolution and pooling, and C is the number of channels; in this paper, T′=20 and C=24. This feature map characterizes the dynamic evolution pattern of process parameters within a window during cement clinker calcination. Subsequently, to obtain local feature vectors, global average pooling is performed along the time dimension:

[0083]

[0084] The pooled vector h l,t It is a C-dimensional vector that retains the main information of the feature map while eliminating redundant dimensions. Through this continuous mapping, the original two-dimensional data block sequence is reconstructed into a set of feature vectors H={h} with temporal logic. l,1 ,h l,2 ,...,h l,T}

[0085] Step 4: Use Gated Recurrent Unit (GRU) to perform long-range dependency modeling of global time-varying patterns.

[0086] To model the long-range time delay effect and strong thermal inertia in cement calcination, this paper utilizes a gated recurrent unit (GRU) to further model the local feature sequences extracted by the traditional recurrent neural network (TCN) to learn the temporal dependencies across data blocks. GRU simplifies the structure of the traditional recurrent neural network by integrating the forget gate and input gate into a single update gate, thereby alleviating the gradient vanishing problem in deep networks while reducing the number of model parameters and improving convergence performance. In the specific recursive update process, the model first updates the global hidden state hg from the previous time step... t-1 The local feature vector h input at the current time l,t The update gate z is calculated. t With Reset Gate r t Gating value:

[0087]

[0088]

[0089] Subsequently, using the reset gate r t Filter historical state information and combine it with the input features at the current time to calculate candidate hidden states h̃ g,t :

[0090]

[0091] Finally, by updating gate z t The historical background information and the current candidate state are weighted and fused to output the final global feature vector hg at that moment. t :

[0092]

[0093] In the formula, σ and tanh(⋅) are activation functions, ⊙ represents element-wise multiplication, and W and b are the weights and biases to be learned; under this recursive update mechanism, the hidden state hg at the current time step, t Based on the global information hg from the previous time step, t-1 With the local features h at the current time l,t Together, the GRU maps the local feature sequence H extracted by the TCN to a global feature vector hg containing long-term temporal dependencies through the aforementioned recursive update mechanism. t .

[0094] Step 5: Adaptive dynamic fusion of local details and global trends.

[0095] In the calcination process of cement clinker, the influence weights of different process parameters on f-CaO content exhibit time-varying characteristics with changes in production conditions. Typically, when production is in a steady state, model predictions need to shift towards the long-term trend to maintain stability; however, when sudden disturbances occur, the impact of instantaneous fluctuations on the prediction results needs to be increased accordingly. Traditional hard-joining or static average allocation methods are difficult to reflect this dynamic weight change. Therefore, this paper designs an adaptive gating fusion module, which adjusts the ratio of local features to global information in real time by constructing a learnable weight generation network. In this module, the local feature vector hl extracted by the TCN is first... t Compared with the global feature vector hg extracted by GRU, t Each feature vector is mapped to the same dimensional space via a linear projection layer, and then concatenated along the channel dimension to obtain the joint feature vector u. t .

[0096] Subsequently, through a linear mapping layer and a Sigmoid activation function, from u t The fusion weight G is calculated in the middle. t :

[0097]

[0098] In the formula, W G With b G G represents the weight matrix and bias vector to be learned in the gated layer, respectively. t This characterizes the proportion of global dynamic dependencies in feature fusion, while 1−G t This corresponds to the proportion of local details.

[0099] To achieve soft selection of feature information, this paper adopts a complementary weighting strategy to perform weighted summation on the two types of features:

[0100]

[0101] In the formula, h′g,t is the global feature vector after linear mapping, h′l,t is the local feature vector after linear mapping, and ⊙ denotes element-wise multiplication of vectors. Under this mapping mechanism, the resulting fused feature vector hf, t It can dynamically adjust the ratio of local feature sensitivity to global trend stability based on real-time input status; ultimately, hf, t It is fed into the regression layer to calculate the predicted value of f-CaO content.

[0102] Step 6: Validation and model prediction of comparative loss and regression loss based on collaborative alignment mechanism.

[0103] The model's total loss function consists of the prediction regression loss L reg Compared with time-weighted loss L conThe composition is determined by adjusting the contribution weights of the two components through the hyperparameter λ.

[0104]

[0105] (1) Regression loss L reg Mean squared error is selected to calculate the deviation between the predicted and actual f-CaO concentration values. This loss term constrains the prediction accuracy at the output, allowing the model to continuously reduce the prediction error during training. The calculation formula is as follows:

[0106]

[0107] In the formula, N is the number of samples in the training batch, y i With ŷ i The measured data of f-CaO and the network prediction results correspond to the i-th sample, respectively.

[0108] (2) Time-weighted contrast loss L con The local feature vector h extracted by TCN and subjected to global average pooling is applied. l,t Because each h l,t Each corresponds to a compressed representation of a complete data block. This loss adjusts the feature distribution by introducing a soft constraint based on time distance in the feature space. Specifically, for any anchor sample i, its positive sample set is naturally composed of samples adjacent to its sampling time. However, to avoid information loss caused by rigidly dividing positive and negative samples, this paper adopts a soft-weighted comparison mechanism, which does not explicitly specify positive and negative samples, but rather uses the time interval |t between samples. i -t j | Assign a continuous weight w to each pair of samples. ij The weights decay exponentially with increasing intervals, making it easier to bring temporally adjacent samples closer together in the feature space, while samples with larger intervals have only a weak impact due to their small weights. This design incorporates the temporal smoothness of the cement calcination process as prior knowledge into feature learning, suppressing high-frequency random noise while maintaining responsiveness to changes in real-world conditions. During computation, local feature vectors are first mapped to low-dimensional embedding vectors z through a nonlinear projection head. The temporal weighted contrastive loss is then defined as:

[0109]

[0110] In the formula, z i z is the feature vector of anchor point sample i after being mapped by the projection head; j and z k Let represent the feature vectors of samples other than i; τ is a temperature parameter used to adjust the smoothness of the similarity distribution. Weighting factor w ij Determined by the time interval between samples, using an exponential decay function:

[0111]

[0112] In the formula, |t i -t j | represents the time interval between sample pairs, and β is the decay coefficient controlling the rate of decrease of the weights with time difference. The weighted summation in the numerator strengthens the consistency of feature representations of samples with similar time intervals, while the global summation in the denominator serves as a normalization comparison, preventing the model from mapping all sample features to the same point and losing its ability to distinguish between different operating conditions. Through this construction, the model can effectively suppress the interference of high-frequency random noise on feature representation while capturing the continuous evolution characteristics of operating conditions.

[0113] Example 3

[0114] like Figure 3 The experimental data table shows a comparison of the prediction performance of the method of this invention (AGTC-TCN-GRU) with several existing mainstream methods (such as SVM, KNN, BO-LightGBM, MVTS-CNN, SIBSS, etc.) on real cement production datasets.

[0115] The dataset in this example comes from 14,950 operational records from an actual production line of a cement plant, of which 2,640 are labeled samples. The data is divided into training, validation, and test sets in an 8:1:1 chronological order.

[0116] The evaluation metrics used are: coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).

[0117] This paper uses actual industrial production data to evaluate and show that traditional models such as SVM have an R² of only 0.4404, making it difficult to fit time-series data; the advanced ensemble learning BO-LightGBM improves the index to 0.5805, but still cannot overcome the time latency. In contrast, the AGTC-TCN-GRU method proposed in this study achieves a determination coefficient of 0.7970, reduces RMSE to 0.1480, and lowers MAE to 0.1153.

[0118] like Figure 4 As shown in the predicted curve, the f-CaO content curve predicted by the method described in this paper maintains a high degree of overlap and fit with the true value curve obtained from actual testing. Even during peak and trough periods of sudden changes in production conditions and drastic fluctuations in f-CaO content, the model of the method described in this paper can still accurately track the changing trend without exhibiting the obvious lag and off-target phenomena common in traditional models. This demonstrates that the temporal contrastive learning and adaptive gating fusion mechanism in the method described in this paper effectively balance local fluctuations and global trends, greatly enhancing the model's robustness in dealing with complex noise and sudden operating conditions in industrial settings.

[0119] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A TCN-GRU-based double-layer time-series versus regression cement f-CaO prediction method, characterized in that, Includes the following steps: S1. Multi-source data acquisition and matrix reconstruction: Sensor data of cement clinker calcination process is acquired through distributed control system and sparse test labels are obtained simultaneously. Labeled samples are used as anchor points and sliding windows are used to reconstruct unlabeled data into two-dimensional time-series data blocks. S2. Local Temporal Feature Extraction: Local temporal features of data blocks are extracted using a Temporal Convolutional Network (TCN), and then mapped to a local feature vector h using global average pooling to unify the dimension. l,t ; S3, Temporal Weighted Feature Alignment: Utilizing a temporal weighted contrastive learning mechanism, consistency weights are dynamically allocated based on sample time intervals to achieve progressive smoothing and noise suppression across sample feature spaces; S4, long-span feature extraction: the local feature vectors of each data block are recursively evolved and modeled by using a gated recurrent unit (GRU), and a global feature vector hg of long-time sequence dependency across data blocks is extracted, t ; S5, Gated Fusion: The local feature vector h extracted by TCN is... l,t Compared with the global feature vector hg extracted by GRU, t Each feature vector is mapped to the same dimensional space via a linear projection layer, and then concatenated along the channel dimension to obtain the joint feature vector u. t Subsequently, through a linear mapping layer and a Sigmoid activation function, from u t The fusion weight G is calculated in the middle. t The expression is as follows: In the formula, W G With b G G represents the weight matrix and bias vector to be learned in the gated layer, respectively; t This characterizes the proportion of global dynamic dependencies in feature fusion, while 1−G t This corresponds to the proportion of local details; To achieve soft selection of feature information, a complementary weighting strategy is used to perform a weighted summation of the two types of features, as shown in the following expression: In the formula, h′ g,t is the global eigenvector after linear mapping, h′ l,t is the local eigenvector after linear mapping, and ⊙ represents the element-wise multiplication of vectors; Under this mapping mechanism, the resulting fused feature vector hf, t It can dynamically adjust the ratio of local feature sensitivity to global trend stability based on real-time input status; ultimately, hf, t It is fed into the regression layer to calculate the predicted value of f-CaO content; S6. Collaborative Loss Regression Prediction: The training and optimization of the model are driven collaboratively by two complementary loss terms. The prediction regression loss acts on the multi-scale joint features after gated fusion, outputting the predicted f-CaO concentration value through the regression decoding layer and constraining its deviation from the actual test value. The time-weighted contrast loss acts on the local feature layer extracted by TCN. By introducing an exponentially decaying weight constraint based on time distance in the bottom feature space, it forces the local feature representation of temporally adjacent samples to maintain consistency, thereby suppressing high-frequency noise interference from the sensor at the source of feature extraction and improving the feature quality of the input to the subsequent GRU and gated fusion modules. The two losses are weighted and summed through the hyperparameter λ to form the total loss function, realizing end-to-end collaborative optimization and outputting a highly robust final predicted value of cement clinker f-CaO concentration.

2. The TCN-GRU-based double-layer time-series versus regression cement f-CaO prediction method according to claim 1, characterized in that, In S1, the distributed control system collects sensor data of the cement clinker calcination process and simultaneously obtains sparse test labels. Specifically, it acquires key process sensor data such as preheater outlet temperature, decomposition furnace pressure, rotary kiln speed, and pulverized coal flow rate recorded by the cement kiln system, and at the same time acquires the true value of f-CaO concentration fed back by the offline laboratory as a sparse label.

3. The TCN-GRU-based double-layer time-series versus regression cement f-CaO prediction method according to claim 1, characterized in that, In S2, the TCN adopts a two-layer structure: First, a 1×1 convolution layer maps the original 10 variable channels to 48 feature channels. Then, the first causal dilated convolution layer and residual connection are used to extract local temporal patterns, and the number of output channels remains at 48. The second layer further compresses the number of channels to 24, which expands the receptive field while reducing the feature dimension. With appropriate causal padding, the time step length remains unchanged across all convolutional layers. Therefore, the final feature map output by TCN contains C=24 channels, each corresponding to a feature sequence of length T′=W=20, as shown in the following expression: In the formula, θ represents the set of all learnable weights and biases within the network; the feature map H l,t It is a two-dimensional array with dimensions T′ rows and C columns, where T′ is the time step length after convolution and pooling, C is the number of channels, and T′=20, C=24; this feature map characterizes the dynamic evolution mode of process parameters within the window during the calcination of cement clinker; Subsequently, to obtain local feature vectors, global average pooling is performed along the time dimension, as shown in the following expression: The pooled vector h l,t It is a C-dimensional vector that retains the main information of the feature map while eliminating redundant dimensions; through this continuous mapping, the original two-dimensional data block sequence is reconstructed into a set of feature vectors H={h} with temporal logic. l,1 ,h l,2 ,...,h l,T } 4. The TCN-GRU-based double-layer time-series versus regression cement f-CaO prediction method according to claim 1, characterized in that, In S3, the time-weighted feature alignment is achieved by using an exponential decay function to calculate the temporal continuity similarity weight based on the sample time interval span, mapping and transforming it to a low-dimensional embedding through a nonlinear projection module to control the feature components, and achieving smooth alignment of the continuous state condition feature space through the coordinated action of the clustering loss of neighboring positive samples in the numerator and the pull-repulsion loss of global batch samples in the denominator.

5. The method for predicting cement f-CaO based on a two-layer time-series contrastive regression according to claim 1, characterized in that, In S4, the gated recurrent unit (GRU) is used to recursively evolve and model the local feature vectors of each data block to extract global long-range dependency features across data blocks. Specifically, based on the local feature sequence extracted by the TCN, the gated recurrent unit (GRU) is used to further model the local feature sequence extracted by the TCN in order to learn the temporal dependency relationship across data blocks. GRU simplifies the structure of traditional recurrent neural networks by integrating the forget gate and input gate into a single update gate, thereby alleviating the gradient vanishing problem in deep networks, reducing the number of model parameters, and improving convergence performance. In the specific recursive update process, the model first updates the global hidden state hg from the previous time step. t-1 The local feature vector h input at the current time l,t The update gate z is calculated. t With Reset Gate r t The gate value is expressed as follows: Subsequently, the reset gate r t The historical state information is filtered, and the candidate hidden state h̃ g,t is calculated in combination with the input features at the current time Finally, the gate z is updated t The historical background information and the current candidate state are weightedly fused to output the final global feature vector hg of this moment, t : In the formula, σ and tanh(⋅) are activation functions, ⊙ represents element-wise multiplication, and W and b are the weights and biases to be learned; under this recursive update mechanism, the hidden state hg at the current time step, t Based on the global information hg from the previous time step, t-1 With the local features h at the current time l,t To be determined jointly; Through the above recursive update mechanism, the GRU maps the local feature sequence H extracted by the TCN into a global feature vector hgcontaining long sequential dependencies, t .

6. The TCN-GRU-based double-layer time-series versus regression cement f-CaO prediction method according to claim 1, characterized in that, In S6, the total loss function of the model is composed of the prediction regression loss L. reg Compared with time-weighted loss L con Composition, through hyperparameters Adjust the contribution weights of the two; where L reg The prediction accuracy of the joint features after constraint-gated fusion, output by the regression layer, L con Constrained TCN extracts local feature vector h l,t The temporal consistency distribution in the feature space; two losses act on different levels of the model: the regression loss backpropagates from the output to update all network parameters to ensure prediction accuracy, and the contrastive loss directly optimizes the representational ability of the TCN from the bottom feature level to ensure feature quality; the total loss expression is as follows: where L total is the total loss function, L reg is the prediction regression loss, L con is the temporal weighted contrastive loss, and λ is a hyperparameter that adjusts the contribution weight of the two. Among them, regression loss L reg Mean squared error is selected to calculate the deviation between the predicted and actual f-CaO concentration values. This loss term constrains the prediction accuracy at the output, allowing the model to continuously reduce the prediction error during training. The calculation formula is as follows: In the formula, N is the number of samples in the training batch, y i With ŷ i The measured data of f-CaO and the network prediction results correspond to the i-th sample, respectively; Temporal weighted contrastive loss L con the local feature vector h acting on the TCN extraction and global average pooling l,t Since each h l,t corresponds to the compressed representation of a complete data block, this loss adjusts the feature distribution by introducing a soft constraint based on the temporal distance in the feature space; Then, by employing a soft-weighted comparison mechanism, positive and negative samples are not explicitly specified, but rather based on the time interval |t between samples. i -t j | Assign a continuous weight w to each pair of samples. ij The weights decay exponentially with increasing intervals, making it easier to bring temporally adjacent samples closer together in the feature space, while samples with larger intervals have only a weak impact due to their small weights. This design incorporates the temporal smoothness of the cement calcination process as prior knowledge into feature learning, suppressing high-frequency random noise while maintaining responsiveness to changes in real-world conditions. During calculation, the local feature vectors are first mapped to a low-dimensional embedding vector z through a nonlinear projection head. The temporal weighted contrastive loss is then defined as: In the formula, z i z is the feature vector of anchor point sample i after being mapped by the projection head; j and z k Let represent the feature vectors of samples other than i; τ is a temperature parameter used to adjust the smoothness of the similarity distribution; where the weighting factor w ij Determined by the time interval between samples, using an exponential decay function: In the formula, |t i -t j | is the time interval between sample pairs, and β is the decay coefficient that controls the rate of decline of the weight with time difference. The weighted summation in the numerator enhances the consistency of feature representation of samples with similar time proximity, while the global summation in the denominator plays a normalization and comparison role, preventing the model from mapping all sample features to the same point and losing its ability to distinguish between different working conditions. Through this construction, the model can effectively suppress the interference of high-frequency random noise on feature representation while capturing the continuous evolution characteristics of working conditions.

Images