A Deep Learning-Based Method and System for Predicting Industrial Smoke Concentration

Abstract

This invention provides a method and system for predicting industrial smoke emission concentration based on deep learning. The method includes collecting multi-source time-series data such as smoke emission concentration and process parameters; performing continuous wavelet transform on auxiliary process parameters based on a historical causal sliding window to extract the fluctuation coefficient magnitude as a multi-scale auxiliary feature; constructing a backbone prediction network and an auxiliary analysis network; setting a hyperbolic tangent polarization gated unit containing a Sigmoid amplitude gate and a Tanh polarization gate in the recurrent unit of the backbone network; mapping the backbone hidden state and auxiliary features to the same latent semantic space; generating non-negative semantic association strength based on cosine similarity; adjusting the hidden state update in combination with the gating mechanism; and finally outputting the predicted smoke emission concentration value through fully connected decoding.

Description

Technical Field

[0001] This application belongs to the field of prediction, and in particular relates to a method and system for predicting industrial smoke concentration based on deep learning. Background Technology

[0002] With the continuous advancement of industrialization, the requirements for energy conservation, emission reduction, and environmental protection in industrial production processes are becoming increasingly stringent. Real-time monitoring and prediction of industrial flue gas concentrations are crucial for achieving compliant industrial waste gas emissions, optimizing production operations, and ensuring the stable operation of industrial systems. In actual industrial scenarios, changes in flue gas concentrations do not occur in isolation but are influenced by various auxiliary process parameters, such as furnace temperature, gas flow pressure, oxygen supply, and fuel supply rate. These auxiliary process parameters typically exist in the form of multi-source time-series data and exhibit nonlinear, non-stationary, and interconnected fluctuation characteristics at different time scales. Therefore, extracting effective features from these auxiliary process parameters across multiple dimensions and time scales is fundamental to improving the accuracy of flue gas concentration prediction.

[0003] Currently, data-driven methods based on deep neural networks have been applied to industrial time series forecasting. However, existing time series forecasting networks typically fuse historical data on target emission concentrations and auxiliary process parameters by splicing or parallel inputs, making it difficult to fully characterize the correspondence between the two types of data in the deep latent space. Because the historical emission concentration sequence and auxiliary process parameters differ in data dimension, change period, and time evolution mechanism, they are prone to feature expression discrepancies, making it difficult for the model to accurately determine the actual contribution of auxiliary process information to the current forecasting task. Under stable operating conditions, auxiliary process parameters may provide a valid reference for changes in emission concentration; however, during sudden changes in operating conditions, sensor disturbances, or enhanced transient interference, some auxiliary information may become redundant or even misleading. Existing methods generally lack dynamic evaluation and adjustment mechanisms for the strength and direction of auxiliary information correlation, making it difficult to selectively utilize auxiliary information according to different operating conditions, thus affecting the accuracy and stability of emission concentration prediction results. Therefore, there is an urgent need for an industrial emission concentration prediction scheme that can extract multi-scale fluctuation features of auxiliary process parameters, align the latent semantic representation of primary and secondary data, and dynamically regulate the transmission effect of auxiliary information. Summary of the Invention

[0004] To address the problems in existing technologies, such as insufficient extraction of multi-scale fluctuation features of auxiliary process parameters, difficulty in semantic alignment of primary and secondary time series data, and difficulty in dynamically adjusting auxiliary information according to the strength of correlation and direction of action, this invention proposes a deep learning-based method and system for predicting industrial smoke emission concentration.

[0005] In a first aspect, this invention proposes a method for predicting industrial smoke concentration based on deep learning, comprising the following steps:

[0006] Acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of target smoke emission concentration and time-series data of auxiliary process parameters. Perform continuous wavelet transform on the time-series data of auxiliary process parameters based on historical causal sliding window to obtain the fluctuation coefficients representing the process at different time scales, and extract the magnitude of the fluctuation coefficients as auxiliary features.

[0007] A backbone prediction network and an auxiliary analysis network are constructed, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate is constructed, wherein the gate includes an amplitude gate activated by the sigmoid function and a polarization gate activated by the hyperbolic tangent function.

[0008] The hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step are projected onto a preset latent semantic space. The cosine similarity between the two projection vectors is calculated, and the semantic association strength is generated based on the cosine similarity. The semantic association strength is multiplied by the output of the amplitude gate to generate a calibration information flux. The output of the polarization gate is used to determine the enhancement or control effect of the calibration information flux on the hidden state update. Based on the updated hidden state of the backbone prediction network at the final time step, the predicted value of smoke concentration at future time steps is generated by decoding through a fully connected layer.

[0009] Optionally, the step of performing continuous wavelet transform on the time-series data of the auxiliary process parameters based on a historical causal sliding window to obtain fluctuation coefficients representing the process at different time scales, and extracting the magnitude of the fluctuation coefficients as auxiliary features, includes:

[0010] Choose complex Morlet wavelet as the mother wavelet basis function;

[0011] A continuously varying set of scaling parameters and a set of translation parameters are defined. The scaling parameter set is used to perform a scaling transformation on the mother wavelet basis function, and the translation parameter set is used to perform a translation transformation to construct a wavelet basis function family.

[0012] The auxiliary process parameter time series data within the historical causal sliding window are respectively multiplied with each basis function in the wavelet basis function cluster to calculate a complex frequency domain feature matrix containing real and imaginary parts.

[0013] Calculate the absolute value of the modulus of each element in the complex frequency domain feature matrix to obtain a real-number two-dimensional time-frequency fluctuation coefficient matrix. Stack the two-dimensional time-frequency fluctuation coefficient matrices corresponding to each channel along the channel dimension to form an auxiliary feature tensor. Input the auxiliary feature tensor as the auxiliary feature into the auxiliary analysis network.

[0014] Optionally, the construction of a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time-series data, and the auxiliary analysis network is used to process the auxiliary features, including:

[0015] A recurrent neural network containing multiple layers of long short-term memory neurons is used as the backbone prediction network to extract the temporal dependencies in the historical time series data and propagate the hidden states forward along the time steps.

[0016] A convolutional neural network containing two-dimensional convolutional layers, batch normalization layers, and max pooling layers along the scale dimension is used as the auxiliary analysis network to extract the local receptive field of the auxiliary feature tensor and reduce its dimensionality along the scale dimension, outputting an auxiliary feature vector sequence that retains the time dimension.

[0017] Optionally, within the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate unit is constructed. The gate unit includes an amplitude gate activated by a sigmoid function and a polarization gate activated by a hyperbolic tangent function, comprising:

[0018] At the current moment, the current input vector of the historical time series data, the hidden state vector of the previous moment, and the output of the auxiliary analysis network at the current moment are concatenated to obtain the joint input vector;

[0019] The joint input vector is input to the amplitude gate, linearly mapped by the first linear transformation matrix and a first bias term is added. After processing by the Sigmoid function, the output value range is restricted to an amplitude-gated vector between 0 and 1.

[0020] The joint input vector is input to the polarization gate, linearly mapped by the second linear transformation matrix and a second bias term is added. After processing by the hyperbolic tangent function, the output value range is limited to the polarization gate vector between -1 and +1.

[0021] Optionally, the step of projecting the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculating the cosine similarity between the two projection vectors, and generating semantic association strength based on the cosine similarity includes:

[0022] The hidden state vector from the previous time step and the auxiliary feature vector from the current time step are multiplied by their respective weight matrices to obtain a first projection vector and a second projection vector of the same dimension.

[0023] Calculate the L2 norm of the first projection vector and the second projection vector respectively, and use them as their respective vector magnitudes;

[0024] Calculate the dot product of the first projection vector and the second projection vector, divide the dot product result by the product of the magnitudes of the two vectors to obtain a cosine similarity scalar with a value range between -1 and +1, and map the cosine similarity scalar to a range between 0 and 1 to obtain the semantic association strength.

[0025] Optionally, the step of generating predicted smoke concentration values ​​for future times by decoding through a fully connected layer based on the updated hidden state of the backbone prediction network at the final time includes:

[0026] Construct a multilayer perceptron without time loop connections as a decoder;

[0027] The final hidden state vector output by the backbone prediction network after processing all historical time series data at all time steps is input into the decoder;

[0028] By mapping the fully connected weight matrix of the multilayer perceptron and performing nonlinear activation calculation, the high-dimensional hidden state vector is reduced to a one-dimensional scalar, and the one-dimensional scalar is processed by a non-negative activation function to serve as the output predicted value of smoke concentration at future time.

[0029] On the other hand, this invention also proposes a deep learning-based industrial smoke concentration prediction system, comprising the following modules:

[0030] An extraction module is used to acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of target smoke emission concentration and time-series data of auxiliary process parameters. Based on a historical causal sliding window, a continuous wavelet transform is performed on the time-series data of the auxiliary process parameters to obtain the fluctuation coefficients representing the process at different time scales, and the magnitude of the fluctuation coefficients is extracted as auxiliary features.

[0031] A construction module is used to construct a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the cyclic computation unit of the backbone prediction network, a hyperbolic tangent polarization gate unit is constructed. The gate unit includes an amplitude gate activated by the Sigmoid function and a polarization gate activated by the hyperbolic tangent function.

[0032] The computation module is used to project the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculate the cosine similarity between the two projection vectors, generate semantic association strength based on the cosine similarity, multiply the semantic association strength with the output of the amplitude gate to generate a calibration information flux, use the output of the polarization gate to determine the enhancement or control effect of the calibration information flux on the hidden state update, and generate the predicted value of smoke concentration for future time steps through decoding by a fully connected layer based on the updated hidden state of the backbone prediction network at the final time step.

[0033] Preferably, the step of performing continuous wavelet transform on the time-series data of the auxiliary process parameters based on a historical causal sliding window to obtain fluctuation coefficients representing the process at different time scales, and extracting the magnitude of the fluctuation coefficients as auxiliary features, includes:

[0034] Choose complex Morlet wavelet as the mother wavelet basis function;

[0035] A continuously varying set of scaling parameters and a set of translation parameters are defined. The scaling parameter set is used to perform a scaling transformation on the mother wavelet basis function, and the translation parameter set is used to perform a translation transformation to construct a wavelet basis function family.

[0036] The auxiliary process parameter time series data within the historical causal sliding window are respectively multiplied with each basis function in the wavelet basis function cluster to calculate a complex frequency domain feature matrix containing real and imaginary parts.

[0037] Calculate the absolute value of the modulus of each element in the complex frequency domain feature matrix to obtain a real-number two-dimensional time-frequency fluctuation coefficient matrix. Stack the two-dimensional time-frequency fluctuation coefficient matrices corresponding to each channel along the channel dimension to form an auxiliary feature tensor. Input the auxiliary feature tensor as the auxiliary feature into the auxiliary analysis network.

[0038] Preferably, the construction of a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time-series data, and the auxiliary analysis network is used to process the auxiliary features, including:

[0039] A recurrent neural network containing multiple layers of long short-term memory neurons is used as the backbone prediction network to extract the temporal dependencies in the historical time series data and propagate the hidden states forward along the time steps.

[0040] A convolutional neural network containing two-dimensional convolutional layers, batch normalization layers, and max pooling layers along the scale dimension is used as the auxiliary analysis network to extract the local receptive field of the auxiliary feature tensor and reduce its dimensionality along the scale dimension, outputting an auxiliary feature vector sequence that retains the time dimension.

[0041] Preferably, within the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate unit is constructed. The gate unit includes an amplitude gate activated by a sigmoid function and a polarization gate activated by a hyperbolic tangent function, comprising:

[0042] At the current moment, the current input vector of the historical time series data, the hidden state vector of the previous moment, and the output of the auxiliary analysis network at the current moment are concatenated to obtain the joint input vector;

[0043] The joint input vector is input to the amplitude gate, linearly mapped by the first linear transformation matrix and a first bias term is added. After processing by the Sigmoid function, the output value range is restricted to an amplitude-gated vector between 0 and 1.

[0044] The joint input vector is input to the polarization gate, linearly mapped by the second linear transformation matrix and a second bias term is added. After processing by the hyperbolic tangent function, the output value range is limited to the polarization gate vector between -1 and +1.

[0045] Preferably, the step of projecting the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculating the cosine similarity between the two projection vectors, and generating semantic association strength based on the cosine similarity includes:

[0046] The hidden state vector from the previous time step and the auxiliary feature vector from the current time step are multiplied by their respective weight matrices to obtain a first projection vector and a second projection vector of the same dimension.

[0047] Calculate the L2 norm of the first projection vector and the second projection vector respectively, and use them as their respective vector magnitudes;

[0048] Calculate the dot product of the first projection vector and the second projection vector, divide the dot product result by the product of the magnitudes of the two vectors to obtain a cosine similarity scalar with a value range between -1 and +1, and map the cosine similarity scalar to a range between 0 and 1 to obtain the semantic association strength.

[0049] Preferably, the step of generating predicted smoke concentration values ​​for future times by decoding through a fully connected layer based on the updated hidden state of the backbone prediction network at the final time includes:

[0050] Construct a multilayer perceptron without time loop connections as a decoder;

[0051] The final hidden state vector output by the backbone prediction network after processing all historical time series data at all time steps is input into the decoder;

[0052] By mapping the fully connected weight matrix of the multilayer perceptron and performing nonlinear activation calculation, the high-dimensional hidden state vector is reduced to a one-dimensional scalar, and the one-dimensional scalar is processed by a non-negative activation function to serve as the output predicted value of smoke concentration at future time.

[0053] This invention improves the feature representation capability and model stability in the industrial smoke emission concentration prediction process. By processing auxiliary process parameters through continuous wavelet transform, temporal evolution features and fluctuation patterns at different time scales are extracted, enabling the auxiliary process parameters to participate in prediction as multi-scale time-frequency features. Through a collaborative architecture designed for the backbone prediction network and the auxiliary analysis network, and by constructing a hyperbolic tangent polarization gating unit within the cyclic computation unit, the hidden state of the backbone and auxiliary features are projected into the latent semantic space. Non-negative semantic association strength is generated based on cosine similarity, and a calibration information flux is generated using an amplitude gate. The polarization gate then adjusts the enhancement or control effect of the calibration information flux on the hidden state update of the backbone network. Therefore, the model can selectively utilize auxiliary information based on the strength and direction of the association between the main and auxiliary features, reducing the impact of irrelevant or perturbing auxiliary information on the prediction results, thereby outputting smoke emission concentration prediction results that better reflect the current operating conditions and trends. Attached Figure Description

[0054] Figure 1 A flowchart of the first embodiment;

[0055] Figure 2 This is a schematic diagram of the output response curves of the amplitude gate and the polarization gate;

[0056] Figure 3 This is a schematic diagram of the distribution of semantic association strength. Detailed Implementation

[0057] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0058] In the first embodiment, the present invention proposes a method for predicting industrial smoke concentration based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0059] S1. Acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of the target smoke emission concentration and time-series data of auxiliary process parameters. Perform continuous wavelet transform on the time-series data of the auxiliary process parameters based on the historical causal sliding window to obtain the fluctuation coefficients representing the process at different time scales, and extract the magnitude of the fluctuation coefficients as auxiliary features.

[0060] The system reads multi-source time-series data in comma-separated value format stored in the industrial distributed control system. It normalizes the historical time-series data of the target smoke emission concentration and the time-series data of auxiliary process parameters, mapping them to the interval 0 to 1. When constructing the historical causal sliding window, the window step size is set to 1, and the historical causal sliding window only includes historical data points up to and including the current window's end time, excluding data points for future target prediction times. For the auxiliary process parameter time-series data segmented by the sliding window, a continuous wavelet transform is performed, using complex Morlet wavelet basis functions, and setting the scale sequence to cover the low-frequency to high-frequency range to calculate the complex wave coefficients. The magnitude of these complex wave coefficients is calculated, constructing an auxiliary feature tensor containing multi-scale time-frequency information.

[0061] In an optional embodiment, the step of performing continuous wavelet transform on the time-series data of the auxiliary process parameters based on a historical causal sliding window to obtain fluctuation coefficients representing the process at different time scales, and extracting the magnitude of the fluctuation coefficients as auxiliary features, includes:

[0062] Choose complex Morlet wavelet as the mother wavelet basis function;

[0063] A continuously varying set of scaling parameters and a set of translation parameters are defined. The scaling parameter set is used to perform a scaling transformation on the mother wavelet basis function, and the translation parameter set is used to perform a translation transformation to construct a wavelet basis function family.

[0064] The auxiliary process parameter time series data within the historical causal sliding window are respectively multiplied with each basis function in the wavelet basis function cluster to calculate a complex frequency domain feature matrix containing real and imaginary parts.

[0065] Calculate the absolute value of the modulus of each element in the complex frequency domain feature matrix to obtain a real-number two-dimensional time-frequency fluctuation coefficient matrix. Stack the two-dimensional time-frequency fluctuation coefficient matrices corresponding to each channel along the channel dimension to form an auxiliary feature tensor. Input the auxiliary feature tensor as the auxiliary feature into the auxiliary analysis network.

[0066] A complex Morlet wavelet with a center frequency and bandwidth parameter is chosen as the mother wavelet basis function for the continuous wavelet transform. The bandwidth parameter is assumed to be set as follows: =1.5, center frequency parameter set to =1.0, where and Used to determine the function shape of complex Morlet wavelets. The normalized center frequency parameter is not used directly as the Hertz value at the actual sampling frequency. The length of the historical causal sliding window is set to W=128, for example, corresponding to a 128-second historical data window at a 1Hz sampling frequency of an industrial sensor. This window is set to only include past and current time data. A set of discrete scale parameters s, equally spaced in logarithmic space, is generated, with values ​​ranging from 1 to 64, resulting in M=32 independent scale values. The set of translation parameters τ is set to correspond to the 128 discrete time steps of the sliding window. Through stretching and translation transformations of the mother wavelet using these parameters, a wavelet basis function cluster containing 32×128 basis functions is constructed, covering multi-timescale evolution characteristics from high-frequency transient noise to low-frequency process trends.

[0067] The time-series data of multivariate auxiliary process parameters collected within the sliding window, assuming the number of sensor acquisition channels is C=8, representing parameters such as temperature, pressure, and coal feed rate, are subjected to discrete convolution inner product operations with the aforementioned wavelet basis function cluster. This step outputs a complex time-frequency feature tensor with a shape of 8×32×128, where 8 represents the number of auxiliary process parameter channels, 32 represents the number of scales, and 128 represents the number of time steps within the window. Each complex element Z=a+bi in the tensor is characterized by the real part a and the imaginary part b, representing the local oscillation response at the corresponding scale and translation position. The magnitude is used to represent the fluctuation intensity at that position, and the phase information can be characterized by the ratio of the real and imaginary parts.

[0068] Moduloization is performed on the complex time-frequency feature tensor, and the square root of the sum of the squares of the real and imaginary parts is calculated element-wise to obtain the absolute value of the modulus of the complex number. This operation can reduce the impact of sampling delay or phase shift on subsequent feature alignment, converting the high-dimensional complex tensor into a dimension-invariant 8×32×128 real-number auxiliary feature tensor. This auxiliary feature tensor is then used as a high-order auxiliary feature with time-frequency domain correlation representation and input into the subsequent auxiliary analysis network for deep mapping.

[0069] S2, construct a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the cyclic computation unit of the backbone prediction network, construct a hyperbolic tangent polarization gate unit, wherein the gate unit includes an amplitude gate activated by the Sigmoid function and a polarization gate activated by the hyperbolic tangent function.

[0070] A backbone prediction network and an auxiliary analysis network are constructed. The auxiliary analysis network is built using a convolutional neural network to extract local scale-time variation patterns in auxiliary features. The channel dimension and scale dimension of the auxiliary feature tensor are merged to form a feature input with the time dimension as the sequence direction, which is then input into a one-dimensional convolutional neural network. Through convolution operations and nonlinear activation, the local fluctuation features of the auxiliary process parameters along the time direction are extracted. The backbone prediction network is built based on a long short-term memory network to extract the temporal dependencies in the historical sequence of target smoke emission concentration. A hyperbolic tangent polarization gating unit is set inside the recurrent computation unit of the backbone prediction network. The hyperbolic tangent polarization gating unit includes an amplitude gate and a polarization gate. The amplitude gate generates an amplitude gate vector with a value range of 0 to 1 through linear transformation and a sigmoid activation function, which is used to control the proportion of auxiliary calibration information injected into the hidden state update process. The polarization gate generates a polarization gate vector with a value range of -1 to 1 through linear transformation and a tanh activation function, which is used to determine whether the auxiliary calibration information has a positive enhancement effect or a negative control effect on the hidden state update.

[0071] In an optional embodiment, the construction of a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time-series data, and the auxiliary analysis network is used to process the auxiliary features, including:

[0072] A recurrent neural network containing multiple layers of long short-term memory neurons is used as the backbone prediction network to extract the temporal dependencies in the historical time series data and propagate the hidden states forward along the time steps.

[0073] A convolutional neural network containing two-dimensional convolutional layers, batch normalization layers, and max pooling layers along the scale dimension is used as the auxiliary analysis network to extract the local receptive field of the auxiliary feature tensor and reduce its dimensionality along the scale dimension, outputting an auxiliary feature vector sequence that retains the time dimension.

[0074] The constructed backbone prediction network is configured as a stacked LSTM architecture with two hidden layers of Long Short-Term Memory neurons, serving as a recurrent neural network to extract temporal features. The input is historical time-series data, and the output is a hidden state vector. The number of hidden units in each LSTM neuron layer is set to [value missing]. =64. At each discrete time step t in the sliding window, where t=1,2,…,128, the backbone prediction network receives single-dimensional historical sequence data of target smoke emission concentration, with a tensor shape of 1×1. Through internal input gates, forget gates, and output gates, it collaboratively computes and outputs a hidden state vector of dimension 64 for the current time step. The long-term nonlinear evolution dependence of the extracted smoke concentration is propagated forward along 128 time steps.

[0075] The auxiliary analysis network is designed as a two-dimensional convolutional neural network architecture for processing multi-channel time-frequency spectra. The input is an auxiliary feature tensor of shape 8×32×128, and the output is an auxiliary feature matrix. The first processing module of the auxiliary analysis network includes a two-dimensional convolutional layer with a kernel size of 3×3, a stride of 1×1, and uniform padding, mapping the number of feature channels from 8 to 16. A batch normalization layer with a cascaded momentum parameter of 0.1 is used to improve the stability of intermediate feature distribution during training, and the ReLU activation function is used to extract nonlinear local receptive field features. The network is then connected to an asymmetric two-dimensional max-pooling layer with a kernel size of 4×1 and a stride of 4×1. This pooling layer performs unidirectional four-fold downsampling along the scale dimension (M=32) while keeping the time dimension (W=128) unchanged, compressing the scale dimension to 8.

[0076] After processing through a second-level convolutional layer with a similar kernel size of 3×3, the number of channels is expanded to [number missing]. =64. A second unidirectional pooling layer with a kernel size of 8×1 and a stride of 8×1 is applied to compress the remaining scale dimension to 1. The auxiliary analysis network outputs a high-dimensional tensor of 64×1×128. After compression and dimensionality reduction to remove the scale dimension of size 1, the output tensor becomes a 64×128 auxiliary feature matrix. This allows a sequence of auxiliary feature vectors of length 64 to be extracted at each discrete time step t, achieving temporal and feature space alignment with the hidden state dimension of the backbone prediction network.

[0077] In an optional embodiment, within the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate is constructed. The gate includes an amplitude gate activated by a sigmoid function and a polarization gate activated by a hyperbolic tangent function, comprising:

[0078] At the current moment, the current input vector of the historical time series data, the hidden state vector of the previous moment, and the output of the auxiliary analysis network at the current moment are concatenated to obtain the joint input vector;

[0079] The joint input vector is input to the amplitude gate, linearly mapped by the first linear transformation matrix and a first bias term is added. After processing by the Sigmoid function, the output value range is restricted to an amplitude-gated vector between 0 and 1.

[0080] The joint input vector is input to the polarization gate, linearly mapped by the second linear transformation matrix and a second bias term is added. After processing by the hyperbolic tangent function, the output value range is limited to the polarization gate vector between -1 and +1.

[0081] At any time t within the sliding time window, the hyperbolic tangent polarization gating unit performs joint detection of multimodal temporal features. The scalar input vector of the target smoke concentration at the current time is used. This vector has a dimension of 1 and is the hidden state vector passed by the backbone prediction network from the previous time t-1. The vector has a dimension of 64, and the auxiliary feature vector is extracted and output by the auxiliary analysis network at the current time. The vector has a dimension of 64, and a concatenation operation is performed along the feature dimensions to construct a high-dimensional joint input vector with dimensions 1 + 64 + 64 = 129. This allows contextual historical associations and current multi-source process features to participate in gated computation within the same joint input space.

[0082] Joint input vector Parallel distribution enters two branch channels: the amplitude gate and the polarization gate. Within the amplitude gate calculation channel, With the first linear transformation matrix initialized by Xavier The shape parameter is 64×129. A matrix multiplication mapping is performed, and a first bias vector of shape 64 is accumulated. Generate intermediate log-odds values. Element-wise mapping is performed using the Sigmoid activation function to generate the output 64-dimensional magnitude-gated vector. Each element's value range is smoothly restricted to a closed interval between 0 and 1, representing the magnitude and intensity ratio of the auxiliary calibration information injection allowed for each hidden neuron dimension.

[0083] Within the polarization gate calculation channel With the independently initialized second linear transformation matrix Perform matrix multiplication mapping, the second linear transformation matrix The shape parameter is 64×129, and the matrix multiplication results are accumulated into a second bias vector with a shape of 64. Unlike amplitude gates, this method uses a hyperbolic tangent activation function for nonlinear projection calculations, outputting a 64-dimensional polarization-gated vector whose element values ​​are constrained to the range of -1 to 1. .when A value greater than 0 in a specific dimension indicates that the current auxiliary information positively enhances the hidden state of smoke emission evolution in that dimension; a value less than 0 indicates that the auxiliary process parameters are inversely opposed to the original smoke emission concentration trend, thus providing bidirectional polarity constraint regulation capability for state updates through reverse control. The output response curves of the amplitude gate and polarization gate are shown below. Figure 2 As shown.

[0084] S3, the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step are projected onto a preset latent semantic space, respectively. The cosine similarity between the two projection vectors is calculated, and the semantic association strength is generated based on the cosine similarity. The semantic association strength is multiplied by the output of the amplitude gate to generate a calibration information flux. The output of the polarization gate is used to determine the enhancement or control effect of the calibration information flux on the hidden state update. Based on the updated hidden state of the backbone prediction network at the final time step, the predicted value of smoke concentration at future time steps is generated by decoding through a fully connected layer.

[0085] Two linear mapping layers with the same output dimension are set up to process the hidden state vector of the backbone prediction network at the previous time step and the auxiliary feature vector output by the auxiliary analysis network at the current time step, respectively. The hidden state vector is input into the first linear mapping layer, and the auxiliary feature vector is input into the second linear mapping layer, resulting in a first projection vector and a second projection vector with the same dimension, so that the backbone temporal information and auxiliary feature information are mapped to the same latent semantic space. The cosine similarity scalar value of the two projection vectors in the feature dimension is calculated, and this cosine similarity scalar value is mapped to the semantic association strength with a value between 0 and 1.

[0086] The semantic association strength is multiplied element-wise with the tensor output by the amplitude gate to generate a calibration information flux. The polarization tensor output by the polarization gate is then multiplied element-wise with the calibration information flux. If the polarization gate output is positive, it enhances the current hidden state update; if it is negative, it controls it. Let the semantic association strength be... The amplitude gate vector is The polarization gate vector is Then the calibration adjustment amount can be expressed as , where ⊙ denotes element-wise multiplication. When it is a scalar, it can be multiplied with the amplitude-gated vector through a broadcast mechanism. Because and All are non-negative quantities, and the direction of increase or control of the calibration adjustment quantity is from... The sign is determined. The calibration adjustment amount is then... The candidate cell state update terms at the current moment are added to the backbone prediction network to obtain the cell state update results after calibration with auxiliary information. Based on the cell state update results, the hidden state at the current moment is generated, and the iteration of the hidden state at the current moment is completed.

[0087] At the last time step of the sliding window, the final hidden state vector after the backbone prediction network completes the gating update is extracted, and this final hidden state vector is input to the output decoding layer. The output decoding layer converts the high-dimensional hidden state vector into a one-dimensional normalized scalar prediction value through linear mapping and normalized activation processing; then, based on the target smoke concentration normalization parameters recorded during the training phase, the one-dimensional normalized scalar prediction value is de-normalized to obtain the smoke concentration prediction value at future time steps.

[0088] In an optional embodiment, the step of projecting the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculating the cosine similarity between the two projection vectors, and generating semantic association strength based on the cosine similarity includes:

[0089] The hidden state vector from the previous time step and the auxiliary feature vector from the current time step are multiplied by their respective weight matrices to obtain a first projection vector and a second projection vector of the same dimension.

[0090] Calculate the L2 norm of the first projection vector and the second projection vector respectively, and use them as their respective vector magnitudes;

[0091] Calculate the dot product of the first projection vector and the second projection vector, divide the dot product result by the product of the magnitudes of the two vectors to obtain a cosine similarity scalar with a value range between -1 and +1, and map the cosine similarity scalar to a range between 0 and 1 to obtain the semantic association strength.

[0092] In the algorithm architecture, a projection dimension of the pre-defined latent semantic space is set. =32. Within each time-series update step t, the hidden state vector of the previous time step carried by the backbone prediction network is extracted. The vector has a dimension of 64, and the auxiliary feature vector extracted at the current time by the auxiliary analysis network. The vector has a dimension of 64. The first learned weight matrix with a size parameter configured to 32×64 is invoked. right Perform matrix multiplication to reduce dimensionality, and calculate the first projection vector with dimension 32. Call the second independent learning weight matrix of the same size, 32×64. right Perform a linear transformation to calculate the second projection vector, which is also 32-dimensional. This maps and aligns two types of heterogeneous information into a higher-order space with the same geometric distribution metric.

[0093] After obtaining the same-dimensional representation, calculate the first projection vector respectively. With the second projection vector The L2 norm in geometric space. Taking the square root of the sum of the squares of all elements of a 32-dimensional vector yields the scalar modulus representing the geometric length of each element. and To reduce numerical instability and division-by-zero errors caused by the magnitude being close to a minimum or even zero in the computational graph of a neural network, a minimum smoothing constant is added to the denominator of the product of the two magnitudes. The preferred setting is 1×10 -8 .

[0094] implement and The operation involves multiplying and summing elements in corresponding dimensions to obtain the inner product dot product. This dot product value is then divided by the product of the magnitudes of the two vectors after adding a smoothing constant. This step outputs a scalar value. ,Will Mapped to semantic association strength without directional polarity Preferred options are preferred. Examples of the distribution of semantic association strength are as follows: Figure 3 As shown.

[0095] In an optional embodiment, the step of generating predicted smoke concentration values ​​for future times through decoding via a fully connected layer based on the updated hidden state of the backbone prediction network at the final time step includes:

[0096] Construct a multilayer perceptron without time loop connections as a decoder;

[0097] The final hidden state vector output by the backbone prediction network after processing all historical time series data at all time steps is input into the decoder;

[0098] By mapping the fully connected weight matrix of the multilayer perceptron and performing nonlinear activation calculation, the high-dimensional hidden state vector is reduced to a one-dimensional scalar, and the one-dimensional scalar is processed by a non-negative activation function to serve as the output predicted value of smoke concentration at future time.

[0099] A multilayer perceptron (MLP) was constructed as the end-point quantity reconstruction decoder. As a neural network structure, the decoder's input is the final hidden state vector of the backbone prediction network, and its output is the predicted smoke concentration at future time steps. To improve the immediacy of the inference process and reduce the computational cost of additional recurrent feedback, the decoder network does not contain any time-dependent feedback loops. The decoder network topology is configured as a cascaded two-layer fully connected feedforward layer: the first hidden layer receives the input from the front end, with a weight matrix of 32×64, and is enhanced with a Mish activation function that provides smooth gradients to improve the nonlinear mapping margin; the second output layer is responsible for dimensionality reduction, with a weight matrix of 1×32.

[0100] At the end of the inference cycle, after the backbone prediction network has sequentially traversed and processed all historical data frame sequences with a window length of W=128, the network will output the final hidden state vector generated by the last time step update. The tensor dimension is 64. This vector now incorporates global temporal evolution memory, semantic gating control, and auxiliary multi-scale polarization calibration information. This vector is truncated and extracted, and used as the sole tensor input to the MLP decoder at the current time step.

[0101] After entering the decoder, the 64-dimensional final hidden state vector is transformed into a 32-dimensional high-dimensional nonlinear representation through matrix multiplication and Mish activation in the hidden layer. Spatial compression is then performed by the fully connected weights of the output layer, resulting in a one-dimensional floating-point scalar. Combined with the aforementioned normalization process where the target smoke concentration has been mapped to the 0-1 interval, this one-dimensional floating-point scalar is mapped to a normalized smoke concentration prediction value within the 0-1 interval using the Sigmoid activation function. Subsequently, inverse normalization is performed based on the minimum and maximum target smoke concentration values ​​recorded during the training phase, yielding a future smoke concentration prediction value with actual physical dimensions. The future time can be set to 5 minutes or 15 minutes after the end of the current window, depending on the prediction task. During the training phase, the actual smoke concentration corresponding to this prediction step size is used as the supervision label.

[0102] In the second embodiment, the present invention also proposes an industrial smoke concentration prediction system based on deep learning, comprising the following modules:

[0103] An extraction module is used to acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of target smoke emission concentration and time-series data of auxiliary process parameters. Based on a historical causal sliding window, a continuous wavelet transform is performed on the time-series data of the auxiliary process parameters to obtain the fluctuation coefficients representing the process at different time scales, and the magnitude of the fluctuation coefficients is extracted as auxiliary features.

[0104] A construction module is used to construct a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the cyclic computation unit of the backbone prediction network, a hyperbolic tangent polarization gate unit is constructed. The gate unit includes an amplitude gate activated by the Sigmoid function and a polarization gate activated by the hyperbolic tangent function.

[0105] The computation module is used to project the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculate the cosine similarity between the two projection vectors, generate semantic association strength based on the cosine similarity, multiply the semantic association strength with the output of the amplitude gate to generate a calibration information flux, use the output of the polarization gate to determine the enhancement or control effect of the calibration information flux on the hidden state update, and generate the predicted value of smoke concentration for future time steps through decoding by a fully connected layer based on the updated hidden state of the backbone prediction network at the final time step.

[0106] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and no limitation is imposed herein.

[0107] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for predicting industrial smoke concentration based on deep learning, characterized in that, Includes the following steps: Acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of target smoke emission concentration and time-series data of auxiliary process parameters. Perform continuous wavelet transform on the time-series data of auxiliary process parameters based on historical causal sliding window to obtain the fluctuation coefficients representing the process at different time scales, and extract the magnitude of the fluctuation coefficients as auxiliary features. A backbone prediction network and an auxiliary analysis network are constructed, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate is constructed, wherein the gate includes an amplitude gate activated by the sigmoid function and a polarization gate activated by the hyperbolic tangent function. The hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step are projected onto a preset latent semantic space. The cosine similarity between the two projection vectors is calculated, and the semantic association strength is generated based on the cosine similarity. The semantic association strength is multiplied by the output of the amplitude gate to generate a calibration information flux. The output of the polarization gate is used to determine the enhancement or control effect of the calibration information flux on the hidden state update. Based on the updated hidden state of the backbone prediction network at the final time step, the predicted value of smoke concentration at future time steps is generated by decoding through a fully connected layer.

2. The method according to claim 1, characterized in that, The step of performing continuous wavelet transform on the time-series data of the auxiliary process parameters based on a historical causal sliding window to obtain fluctuation coefficients representing the process at different time scales, and extracting the magnitude of the fluctuation coefficients as auxiliary features, includes: Choose complex Morlet wavelet as the mother wavelet basis function; A continuously varying set of scaling parameters and a set of translation parameters are defined. The scaling parameter set is used to perform a scaling transformation on the mother wavelet basis function, and the translation parameter set is used to perform a translation transformation to construct a wavelet basis function family. The auxiliary process parameter time series data within the historical causal sliding window are respectively multiplied with each basis function in the wavelet basis function cluster to calculate a complex frequency domain feature matrix containing real and imaginary parts. Calculate the absolute value of the modulus of each element in the complex frequency domain feature matrix to obtain a real-number two-dimensional time-frequency fluctuation coefficient matrix. Stack the two-dimensional time-frequency fluctuation coefficient matrices corresponding to each channel along the channel dimension to form an auxiliary feature tensor. Input the auxiliary feature tensor as the auxiliary feature into the auxiliary analysis network.

3. The method according to claim 1, characterized in that, The process involves constructing a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time-series data, and the auxiliary analysis network is used to process the auxiliary features, including: A recurrent neural network containing multiple layers of long short-term memory neurons is used as the backbone prediction network to extract the temporal dependencies in the historical time series data and propagate the hidden states forward along the time steps. A convolutional neural network containing two-dimensional convolutional layers, batch normalization layers, and max pooling layers along the scale dimension is used as the auxiliary analysis network to extract the local receptive field of the auxiliary feature tensor and reduce its dimensionality along the scale dimension, outputting an auxiliary feature vector sequence that retains the time dimension.

4. The method according to claim 2 or 3, characterized in that, Within the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate is constructed. This gate includes an amplitude gate activated by a sigmoid function and a polarization gate activated by a hyperbolic tangent function. At the current moment, the current input vector of the historical time series data, the hidden state vector of the previous moment, and the output of the auxiliary analysis network at the current moment are concatenated to obtain the joint input vector; The joint input vector is input to the amplitude gate, linearly mapped by the first linear transformation matrix and a first bias term is added. After processing by the Sigmoid function, the output value range is restricted to an amplitude-gated vector between 0 and 1. The joint input vector is input to the polarization gate, linearly mapped by the second linear transformation matrix and a second bias term is added. After processing by the hyperbolic tangent function, the output value range is limited to the polarization gate vector between -1 and +1.

5. The method according to claim 1, characterized in that, The step of projecting the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculating the cosine similarity between the two projection vectors, and generating semantic association strength based on the cosine similarity includes: Multiply the hidden state vector from the previous time step and the auxiliary feature vector from the current time step by their respective weight matrices to obtain a first projection vector and a second projection vector of the same dimension. Calculate the L2 norm of the first projection vector and the second projection vector respectively, and use them as their respective vector magnitudes; Calculate the dot product of the first projection vector and the second projection vector, divide the dot product result by the product of the magnitudes of the two vectors to obtain a cosine similarity scalar with a value range between -1 and +1, and map the cosine similarity scalar to a range between 0 and 1 to obtain the semantic association strength.

6. The method according to claim 1, characterized in that, The step of generating predicted smoke concentration values ​​for future times by decoding through a fully connected layer based on the final hidden state of the updated backbone prediction network includes: Construct a multilayer perceptron without time loop connections as a decoder; The final hidden state vector output by the backbone prediction network after processing all historical time series data at all time steps is input into the decoder; By mapping the fully connected weight matrix of the multilayer perceptron and performing nonlinear activation calculation, the high-dimensional hidden state vector is reduced to a one-dimensional scalar, and the one-dimensional scalar is processed by a non-negative activation function to serve as the predicted value of smoke concentration at future time.

7. A deep learning-based industrial smoke concentration prediction system, characterized in that, Includes the following modules: An extraction module is used to acquire multi-source time-series data of the industrial production process. The multi-source time-series data includes historical time-series data of target smoke emission concentration and time-series data of auxiliary process parameters. Based on a historical causal sliding window, a continuous wavelet transform is performed on the time-series data of the auxiliary process parameters to obtain the fluctuation coefficients representing the process at different time scales, and the magnitude of the fluctuation coefficients is extracted as auxiliary features. A construction module is used to construct a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time series data, and the auxiliary analysis network is used to process the auxiliary features. Inside the cyclic computation unit of the backbone prediction network, a hyperbolic tangent polarization gate unit is constructed. The gate unit includes an amplitude gate activated by the Sigmoid function and a polarization gate activated by the hyperbolic tangent function. The computation module is used to project the hidden state of the backbone prediction network at the previous time step and the output of the auxiliary analysis network at the current time step onto a preset latent semantic space, calculate the cosine similarity between the two projection vectors, generate semantic association strength based on the cosine similarity, multiply the semantic association strength with the output of the amplitude gate to generate a calibration information flux, use the output of the polarization gate to determine the enhancement or control effect of the calibration information flux on the hidden state update, and generate the predicted value of smoke concentration for future time steps through decoding by a fully connected layer based on the updated hidden state of the backbone prediction network at the final time step.

8. The system according to claim 7, characterized in that, The step of performing continuous wavelet transform on the time-series data of the auxiliary process parameters based on a historical causal sliding window to obtain fluctuation coefficients representing the process at different time scales, and extracting the magnitude of the fluctuation coefficients as auxiliary features, includes: Choose complex Morlet wavelet as the mother wavelet basis function; A continuously varying set of scaling parameters and a set of translation parameters are defined. The scaling parameter set is used to perform a scaling transformation on the mother wavelet basis function, and the translation parameter set is used to perform a translation transformation to construct a wavelet basis function family. The auxiliary process parameter time series data within the historical causal sliding window are respectively multiplied with each basis function in the wavelet basis function cluster to calculate a complex frequency domain feature matrix containing real and imaginary parts. Calculate the absolute value of the modulus of each element in the complex frequency domain feature matrix to obtain a real-number two-dimensional time-frequency fluctuation coefficient matrix. Stack the two-dimensional time-frequency fluctuation coefficient matrices corresponding to each channel along the channel dimension to form an auxiliary feature tensor. Input the auxiliary feature tensor as the auxiliary feature into the auxiliary analysis network.

9. The system according to claim 7, characterized in that, The process involves constructing a backbone prediction network and an auxiliary analysis network, wherein the backbone prediction network is used to process the historical time-series data, and the auxiliary analysis network is used to process the auxiliary features, including: A recurrent neural network containing multiple layers of long short-term memory neurons is used as the backbone prediction network to extract the temporal dependencies in the historical time series data and propagate the hidden states forward along the time steps. A convolutional neural network containing two-dimensional convolutional layers, batch normalization layers, and max pooling layers along the scale dimension is used as the auxiliary analysis network to extract the local receptive field of the auxiliary feature tensor and reduce its dimensionality along the scale dimension, outputting an auxiliary feature vector sequence that retains the time dimension.

10. The system according to claim 7, characterized in that, Within the recurrent computation unit of the backbone prediction network, a hyperbolic tangent polarization gate is constructed. This gate includes an amplitude gate activated by a sigmoid function and a polarization gate activated by a hyperbolic tangent function. At the current moment, the current input vector of the historical time series data, the hidden state vector of the previous moment, and the output of the auxiliary analysis network at the current moment are concatenated to obtain the joint input vector; The joint input vector is input to the amplitude gate, linearly mapped by the first linear transformation matrix and a first bias term is added. After processing by the Sigmoid function, the output value range is restricted to an amplitude-gated vector between 0 and 1. The joint input vector is input to the polarization gate, linearly mapped by the second linear transformation matrix and a second bias term is added. After processing by the hyperbolic tangent function, the output value range is limited to the polarization gate vector between -1 and +1.

Images