A chemical industrial park noise prediction method based on DTW-DCRNN

Abstract

The application discloses a chemical industry park noise prediction method based on DTW-DCRNN, belongs to the field of signal and information processing, and solves the long-term global noise prediction problem of the chemical industry park. The time dynamic regularization theory is introduced into the diffusion convolution recurrent neural network, and a space-time noise prediction network model composed of a time dynamic regularization method, a diffusion convolution recurrent neural network and a Kalman filter is established. The DTW algorithm is improved by using a penalty coefficient, the spatial relationship of the neural network is reconstructed through the DTW model, and the output noise prediction of the neural network is dynamically adjusted by using the Kalman method in combination with the traffic flow characteristics. According to the method, the spatial relationship is reconstructed based on the time sequence similarity first, then the data of all stations are input into the model for prediction and dynamic correction, so that the multi-station noise prediction, long-term noise prediction and global noise level prediction are realized, and the risk of noise disturbance and health damage can be avoided in advance.

Description

Technical Field

[0001] This invention belongs to the field of signal and information processing, specifically relating to a noise prediction method for chemical industrial parks based on DTW-DCRNN. Background Technology

[0002] With the continuous development of industry, environmental noise pollution has become increasingly severe due to the rapid development of urban society and economy, and the main problems in noise pollution prevention and control have become more and more apparent. Meanwhile, with the continuous progress in neural network research, there is an urgent need for a neural network-based method for long-term prediction of environmental noise.

[0003] The rise of the "smart chemical industrial park" concept has laid a solid foundation for the accurate and effective collection of various types of data. The Dongming Engineering Plastics Industrial Park has a relatively complete smart park construction, with comprehensive and abundant monitoring data, providing a good data foundation for conducting research on noise prediction in chemical industrial parks. Using deep learning methods to model, analyze, and predict data, and to mine noise data patterns, can help park managers take preventative measures in advance during periods of high noise levels. These measures include requiring park workers to wear noise-canceling headphones during designated times, setting up noise barriers in advance, or placing excessively noisy equipment away from residential areas. These measures, on the one hand, provide a relatively safe working environment for workers through predictive alerts and early warning analysis, and on the other hand, reduce the impact of noise on the lives of residents near the park. Furthermore, noise levels vary in different locations within the park, making global prediction of environmental noise extremely important for occupational hazard assessment and environmental risk prediction within the park. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a noise prediction method for chemical industrial parks based on DTW-DCRNN. ​​From a spatiotemporal prediction perspective, a spatiotemporal noise prediction network model is constructed, consisting of an improved time dynamic warping algorithm, a diffuse convolutional recurrent neural network, and a Kalman filter, to perform long-term, global noise prediction for chemical industrial parks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A noise prediction method for chemical industrial parks based on DTW-DCRNN is proposed. This method employs an improved time dynamic warping algorithm and reconstructs spatial relationships through the graph structure of monitoring stations to achieve long-term spatiotemporal prediction of noise in chemical industrial parks. The specific steps include the following:

[0007] Step 1: Collect and process monitoring station information and vehicle information data within the chemical industrial park;

[0008] Step 2: Introduce a penalty coefficient to improve the original time dynamic warping algorithm, and introduce it into the diffusion convolutional recurrent neural network. Then, combine it with Kalman filtering to construct a noise prediction network model.

[0009] Step 3: Train the constructed noise prediction network model and output the trained model;

[0010] Step 4: Collect monitoring station information in real time and predict noise in real time based on the trained model.

[0011] Furthermore, in step 1, a gate and several monitoring stations are set up in the chemical industrial park to collect data in real time and record the location information data of each monitoring station simultaneously; the gate and each monitoring station are equipped with sensors; the gate collects information data of vehicles entering and exiting, including license plate number, vehicle entry and exit time, vehicle type and traffic flow characteristics; the monitoring stations collect noise and natural environment data in the chemical industrial park in real time.

[0012] Site location data, noise and natural environment data, and vehicle entry and exit information data are all transmitted to the gateway device, and then exchanged with the database server. During noise prediction, data is extracted in real time and processed using the 3σ criterion.

[0013] Furthermore, the specific process of step 2 is as follows:

[0014] Step 2.1: Introduce a penalty coefficient to improve the original time dynamic warping algorithm and calculate the sequence similarity distance;

[0015] The principle of the improved time dynamic warping algorithm is as follows: The optimal path and the number of common subsequences are obtained to calculate the penalty coefficient, and the similarity distance between stations is calculated based on the penalty coefficient; the specific process is as follows:

[0016] Step 2.1.1: Calculate the distance matrix;

[0017] Suppose we have two noise sequences S and T, as shown in equation (1):

[0018]

[0019] Sequence S has length o, and sequence T has length m, where o, m ∈ Z. + Construct a distance matrix d[s][t] based on the m×o noise sequence, where the i-th sequence point s i (i = 1, 2, ..., o) and the j-th sequence point t j The distance between (j = 1, 2, ..., m) is the matrix (s) i ,t j The value of the element, Euclidean distance: dis(s) i ,t j )=(si -t j ) 2 The matrix distance calculation standard is shown in equation (2):

[0020]

[0021] Step 2.1.2: Find the optimal path in the matrix;

[0022] The optimal path is the path with the smallest accumulated distance that satisfies the lower boundary condition. The lower boundary condition is: starting from the upper right corner of the matrix d[o][m], find the smallest of the three points in the lower left corner as the next node, and continue until the lower left corner of the matrix d[0][0]. In addition, it is necessary to ensure continuity and monotonicity, not to cross or omit any point for matching, and to only align or match with adjacent points. The order of the time series cannot be changed.

[0023] Step 2.1.3: Calculate the number of common subsequences and the length of the optimal path sequence, and then obtain the weights and penalty coefficients;

[0024] The length of the optimal path sequence is the length of the optimal path in the matrix obtained in step 2.1.2.

[0025] The calculation process for the number of common subsequences is as follows: First, construct an empty record list `record`. Iterate through sequences s1 and s2. When a common subsequence is found, use the `record` list to record the length of the common subsequence `ls1` or `ls2` and its position `g`. Extract the common subsequence from the s1 sequence and store it in the `s_sum` array. After the iteration is complete, the length of `s_sum` is the number of common subsequences. When calculating the number of common subsequences, set a threshold of 1, and count subsequences with a length greater than 1 as common subsequences. Finally, count the total number of common subsequences to get the required number of common noise subsequences.

[0026] The formula for calculating the weight w is as shown in equation (3):

[0027]

[0028] Where subseq is the length of the common noise subsequence, and seq is the sequence length of the matrix optimal path;

[0029] The calculation process for setting the penalty coefficient α is shown in equation (4), where x represents the number of common subsequences, and w i This represents the weight of the i-th sequence point.

[0030]

[0031] Step 2.1.4: Represent the optimal path as (r1, r2, ..., r seq ), ri (i = 1, 2, ..., seq) represents the value of the i-th sequence point in the optimal path sequence. Finally, the improved sequence similarity distance dis is obtained by multiplying the penalty coefficient by the sum of the values ​​in the optimal path calculated by the original DTW. DTW ;

[0032]

[0033] Step 2.2: Construct a matrix from the noise sequence similarity distances between each monitoring station and calculate its adjacency matrix to build a graph relationship topology; the specific process is as follows:

[0034] Step 2.2.1: Based on the noise sequence similarity distance information of each monitoring station, calculate the similarity distance matrix M between the stations. Matrix M is as follows:

[0035]

[0036] Where 'a' represents the number of monitoring stations;

[0037] Step 2.2.2: Calculate the standard deviation λ of the similarity distance matrix M. c Let μ be the value of the c-th element in the matrix, N be the number of elements in matrix M, and μ be the value of the c-th element. M The mean of the matrix is ​​calculated as follows:

[0038]

[0039] Step 2.2.3: Analyze the similarity between monitoring stations based on the similarity distance matrix M, and construct the adjacency matrix M using the standard deviation λ of all non-infinite numbers in the similarity distance matrix. d Adjacency matrix M d The elements in The calculation process is as follows:

[0040]

[0041] Where c represents the matrix element index, the larger the similarity distance, the better. The smaller the value, the lower the threshold should be. If the threshold is set to 0.1, then... If the similarity between the two sites is less than the threshold, they are considered to have too low a similarity and no mutual influence relationship. They are regarded as non-adjacent sites, and their weight in the adjacency matrix is ​​zero, so they do not constitute an adjacency relationship.

[0042] Step 2.3: Introduce a Graph Convolutional Neural Network (GCN) to construct an adjacency matrix graph structure. Input this adjacency matrix graph structure into a diffusing convolutional recurrent neural network for noise prediction to obtain preliminary prediction results. The specific process is as follows:

[0043] The propagation mechanism between layers of the GCN convolutional neural network is given by equation (12), where L represents the Lth layer:

[0044]

[0045] Where σ is a nonlinear activation function; W L It is a trainable weight matrix. The degree matrix is ​​constructed as follows: I represents the I-th row of adjacency matrix A, and J represents the J-th column of adjacency matrix A. This indicates that the site's own feature information is retained, where H is the feature extracted by the current layer, and if it is an input layer, then X = H;

[0046] Based on the GCN graph convolutional neural network, a diffuse convolutional recurrent neural network is used to model the spatiotemporal relationship of noisy sequences;

[0047] The stationary distribution of the diffusion process is represented as a weighted combination of infinite random walks on the graph. The diffusion process is expressed by equation (13) and calculated in closed form:

[0048]

[0049] Where W is the node similarity matrix, It is the inverse of the out-degree matrix, β∈[0,1] represents the restart probability, k is the diffusion degree, and ε represents the probability of diffusion from the node. In the DCRNN model, the diffusion process is bidirectional.

[0050] Therefore, in terms of spatial relationships, based on the graph signal feature matrix With filter f θ The bidirectional diffusing convolution operation is defined as Equation (14):

[0051]

[0052] Where θ represents the filter parameters, G represents the graph G, p represents the p-th feature dimension, P represents the total number of feature dimensions, K represents the finite K-step truncation of the diffusion process, ★G represents the diffusion convolution on the graph G, X:, p θ represents the convolution operation of all nodes with respect to the p-th feature. k,1 θ represents the kernel parameter used to calculate the output degree. k,2 This represents the kernel parameters used for in-degree calculation. and Let f(x) represent the transition matrices for diffusion and anti-diffusion, respectively. After defining the convolution operation, a diffusion convolutional layer with a mapping relationship is constructed as shown in equation (15):

[0053]

[0054] Where q represents the q-th feature, Q represents the total number of mapped features, and H:, q This represents the spread convolution operation performed on the q-th feature by all nodes, and the feature matrix... As input, the features extracted by the current layer For output, Let Θ represent the filter, Θ be the parameter tensor, and σ be the activation function;

[0055] In terms of time, a combination of GRU and diffraction is used to construct the diffraction-gated recurrent unit DCGRU, which is expressed as Equation (16):

[0056]

[0057] ★G represents the spread convolution over graph G, r (t) Θ represents the reset gate at time t. r★G The parameter tensor X represents the reset gate. (t) H represents the input at time t. (t-1) This represents the output at time t-1, b r The bias vector representing the reset gate; u (t) Θ represents the update gate at time t. u★G The parameter tensor of the update gate, b u This represents the bias vector of the update gate; C (t) Θ represents the hidden state of the next cell at time t. C★G The parameter tensor representing the hidden state, b C H represents the bias vector of the hidden state. (t) This represents the output at time t;

[0058] Step 2.4: Extract the traffic flow characteristics at the gate corresponding to the current time in real time, and dynamically adjust the DCRNN noise prediction value using the Kalman filtering method;

[0059] Kalman filtering uses the estimate of the previous state to predict the current state; finally, it uses the observation of the current state to correct the predicted value obtained in the prediction stage to obtain a new estimate that is closer to the true value. Specifically, it calculates the Kalman gain of noise and drift by using prior estimation and covariance matrix prediction of normalized traffic flow features and noise sequence, and uses the Kalman gain to correct and update the covariance matrix, thereby correcting the noise value predicted at the current time.

[0060] Furthermore, the specific process of step 3 is as follows:

[0061] Historical data from each monitoring station was extracted to reconstruct the spatiotemporal relationships of the data using the improved DTW algorithm. The noise prediction network model DTW-DCRNN was trained, with 70% of the dataset used as the training set, 10% as the validation set, and the last 20% as the test set. In addition, training results were not recorded in the first 60 epochs, and training results and model parameters were recorded every 10 epochs from 60 to 100 epochs. After 100 epochs, training results and model parameters were recorded every 5 epochs. Finally, training was stopped early based on the measured validation loss to ensure that the model was captured when it was about to overfit.

[0062] The beneficial technical effects of this invention are as follows:

[0063] The original temporal dynamic rule algorithm is improved by introducing a penalty coefficient. The spatial distance relationship is represented by an adjacency matrix constructed using a similarity distance matrix, which effectively compensates for the limitations of the original physical distance method. The prediction results adjusted by the Kalman filter are also closer to the true values ​​and have a compensatory property. The 3σ noise processing method effectively avoids large deviations in the model training prediction results. The spatiotemporal noise prediction network model established in this invention, consisting of a temporal dynamic warping method, a diffuse convolutional recurrent neural network, and a Kalman filter, can achieve multi-site noise prediction, long-term noise prediction, and global noise level prediction, thereby enabling early avoidance of noise pollution and health risks. Attached Figure Description

[0064] Figure 1 This is a flowchart of the noise prediction method for chemical industrial parks based on DTW-DCRNN of the present invention;

[0065] Figure 2 This is a schematic diagram of the data acquisition process of the present invention;

[0066] Figure 3 This is a schematic diagram of the noise prediction deep learning network model of the present invention;

[0067] Figure 4 This is a schematic diagram of the improved DTW algorithm of the present invention;

[0068] Figure 5 This is a flowchart illustrating the calculation of the number of common subsequences in this invention;

[0069] Figure 6 This is a schematic diagram of the structure of the diffusion convolutional recurrent neural network model of the present invention;

[0070] Figure 7 This is a comparison diagram of the changes in the adjacency matrix in Experiment 1 of this invention; where (a) shows the adjacency matrix constructed by the original DTW algorithm, and (b) shows the adjacency matrix constructed by the improved DTW algorithm.

[0071] Figure 8 This is a graph showing the trend of index changes based on time steps in Experiment 1 of this invention.

[0072] Figure 9 This is a curve comparison of the performance metrics MAE and RMSE of the prediction model and the DCRNN model in Experiment 2 of this invention.

[0073] Figure 10 This is a curve comparison of the long-term performance index MAPE of the prediction model of this invention and the DCRNN model in Experiment 2 of this invention. Detailed Implementation

[0074] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0075] This invention addresses the problems in noise prediction in chemical industrial parks from four main aspects: First, the presence of zero values ​​in noise data significantly interferes with prediction accuracy. Therefore, for noise sequence monitoring data processing, the 3σ criterion is used to remove zero values, filtering them and reducing the unbiased ratio standard deviation, thereby lowering prediction error. Second, noise prediction at a single monitoring station is unrelated to other stations, and weights cannot be shared. Therefore, to address the issue of data incompatibility between single-station predictions and the high computational cost of individual station predictions, a graph neural network is used to achieve weight sharing for data information prediction between stations. Third, noise similarity between non-adjacent monitoring stations is not currently included in the calculation. Therefore, to address the problem of calculating station similarity between adjacent noise monitoring stations and non-adjacent stations, the Time Dynamic Warping (DTW) algorithm is used to reconstruct spatial relationships from the perspective of time series similarity. Fourth, given the additive nature of noise, factors with high correlation to noise cannot play a role in spatiotemporal prediction. Since Kalman filtering can adjust and correct prediction results by incorporating other influencing factors, this invention uses the Kalman method to dynamically adjust prediction results based on the characteristics of relevant factors.

[0076] Specifically, this method involves introducing the Time Dynamic Warping (DTW) theory into the Diffused Convolutional Recurrent Neural Network (DCRNN) to establish a spatiotemporal noise prediction network model composed of the DTW method, the DCRNN, and Kalman filtering. A penalty coefficient is used to improve the Time Dynamic Warping (DTW) algorithm, and the spatial relationships of the neural network are reconstructed through the DTW model. The Kalman method is then used to dynamically adjust the output noise prediction of the neural network in conjunction with traffic flow characteristics. This method first reconstructs spatial relationships based on time series similarity, and then feeds data from each station into the model for prediction and dynamic correction, thereby achieving multi-site noise prediction, long-term noise prediction, and global noise level prediction, thus enabling the early avoidance of noise pollution and health risks.

[0077] Based on the characteristics of the DTW algorithm and the DCRNN model, this invention proposes a spatiotemporal noise prediction network model that overcomes the limitations of constructing spatial relationships based on distance. It can effectively utilize the location information of the park and the advantages of the two methods to achieve higher-precision noise prediction.

[0078] like Figure 1 As shown, a noise prediction method for chemical industrial parks based on DTW-DCRNN specifically includes the following steps:

[0079] Step 1: Collect and process monitoring station information and vehicle information data within the chemical industrial park.

[0080] The chemical industrial park is equipped with a gate and several monitoring stations for real-time data collection and synchronous recording of location information data from each monitoring station. Sensors are installed at both the gate and each monitoring station. The gate collects information data on vehicles entering and exiting, including license plate number, vehicle entry and exit time, vehicle type, and other traffic flow characteristics. The monitoring stations collect noise and natural environment data from within the chemical industrial park in real time.

[0081] like Figure 2 As shown, site location data, noise and natural environment data, and vehicle entry and exit information data are all transmitted to the gateway device, and then interacted with the database server for transmission. When performing noise prediction, data is extracted in real time and processed using the 3σ criterion.

[0082] The 3σ criterion first assumes that a set of data contains only random errors, calculates the standard deviation, determines the interval according to probability, and considers any error exceeding this interval as gross error rather than random error. Data containing such errors should be discarded or replaced. The noise value is defined as noise ∈ (u - 3σ). 噪 ,u+3σ 噪 The interval accounts for approximately 99.74%, where u represents the noise mean and σ... 噪 Let u be the noise standard deviation and noise be the noise value. The noise range is defined as 0 ≤ noise < u - 3σ. 噪 (dB) is replaced with the mean.

[0083] The original noise sequence acquired by the sensor contains sparse abrupt zero-values. After processing with the 3σ criterion, the unbiased standard deviation of the samples is reduced, and the number of abrupt changes is decreased, thereby improving the accuracy of neural network noise prediction.

[0084] Step 2: Introduce a penalty coefficient to improve the original time dynamic warping algorithm, and then introduce it into the diffusion convolutional recurrent neural network. Finally, combine it with Kalman filtering to construct a noise prediction network model.

[0085] like Figure 3As shown, the principle of this noise prediction network model is as follows: First, based on the extracted data from each monitoring station, the original time dynamic warping algorithm is used to calculate the optimal path. Then, the original time dynamic warping algorithm is improved by introducing a penalty coefficient. The improved time dynamic warping algorithm is used to calculate the sequence similarity distance, and the monitoring stations are arranged into a matrix to construct an adjacency matrix graph structure. Then, the adjacency matrix is ​​input into a diffusing convolutional recurrent neural network for preliminary noise prediction. Finally, the noise prediction results are dynamically adjusted by combining traffic flow characteristics and Kalman filtering to obtain the final noise prediction result. The specific process is as follows:

[0086] Step 2.1: Introduce a penalty coefficient to improve the original time dynamic warping algorithm and calculate the sequence similarity distance;

[0087] The principle of the improved time dynamic warping algorithm is as follows: obtain the optimal path and the number of common subsequences to calculate the penalty coefficient, and combine the penalty coefficient to calculate the similarity distance between stations.

[0088] Main workflow as follows Figure 4 The specific process is as follows:

[0089] Step 2.1.1: Calculate the distance matrix;

[0090] Suppose we have two noise sequences S and T, as shown in equation (1):

[0091]

[0092] Sequence S has length o, and sequence T has length m, where o, m ∈ Z. + Traditional Euclidean distance can be directly calculated when the overlap of peaks and troughs is high, but it cannot be used to calculate the distance of periodic sequences with different phases. Therefore, this invention constructs a distance matrix d[s][t] based on a noisy sequence m×o, where the i-th sequence point s i (i = 1, 2, ..., o) and the j-th sequence point t j The distance between (j = 1, 2, ..., m) is the matrix (s) i ,t j The value of the element, i.e., the Euclidean distance: dis(s) i ,t j )=(s i -t j ) 2 The matrix distance calculation standard is shown in equation (2):

[0093]

[0094] Step 2.1.2: Find the optimal path in the matrix;

[0095] In the distance matrix calculated in section 2.1.1 above, the rows of the matrix represent the sequence S, and the columns represent the sequence T. The rows and columns correspond one-to-one with the sequence order, but a certain point in sequence S corresponds to multiple points in sequence T. The optimal path is obtained through dynamic regularization. The path selection must satisfy the following boundary conditions: it must start from the upper right corner d[o][m] point, find the smallest of the three lower left points as the next node, and continue until the lower left corner d[0][0] matrix point. In addition, it must ensure continuity and monotonicity, not cross or omit any point for matching, and can only align or match with adjacent points. The order of the time series cannot be changed. The optimal path is the path with the smallest accumulated distance that satisfies the above conditions.

[0096] The DTW algorithm alone is not ideal for calculating distances in unaligned periodic sequences. Therefore, improvements to the DTW algorithm are needed. The improvement involves first using the DTW algorithm to generate a distance matrix and find the optimal path: taking point (i,j) as an example, find the minimum value among the three points to the left and right. If this value is the value of the lower left diagonal, then (i-1,j-1) is the next node; if this value is the value of the left adjacent point, then the next node is (i-1,j); if this value is the value of the lower adjacent point, then (i,j-1) is the next node, and so on until there is no next node. The obtained optimal path provides a basis for verification when calculating the common subsequence of the noisy sequence.

[0097] Step 2.1.3: Calculate the number of common subsequences and the length of the optimal path sequence, and then obtain the weights and penalty coefficients;

[0098] The length of the optimal path in the matrix obtained in step 2.1.2 is the required length of the optimal path sequence.

[0099] The steps for calculating the number of noisy common subsequences are as follows: Figure 5 As shown, first, an empty record list (record) is constructed. Sequences s1 and s2 are iterated through. When a common subsequence is found, its length ls1 or ls2 and its position g are recorded in the record list. This common subsequence is extracted from sequence s1 and stored in the s_sum array. After the iteration is complete, the length of s_sum is the number of common subsequences. When calculating the number of common subsequences, a threshold of 1 is set, meaning that subsequences with a length greater than 1 can be considered common subsequences. Finally, the total number of common subsequences is the required number of common noise subsequences.

[0100] The calculation method is shown in Algorithm 1:

[0101]

[0102] The number of all common noise subsequences is used to calculate the penalty coefficient;

[0103] First, calculate the weight w, as shown in equation (3):

[0104]

[0105] Where subseq is the length of the common noise subsequence, and seq is the sequence length of the matrix optimal path;

[0106] Then, the calculation process for the penalty coefficient α is set as shown in equation (4), where x represents the number of common subsequences, and w i The weight of the i-th sequence point satisfies the criteria that the longer the optimal path sequence is, the more common noise subsequences there are, and the smaller the penalty coefficient is.

[0107]

[0108] Step 2.1.4: Represent the optimal path as (r1, r2, ..., r seq ), r i (i = 1, 2, ..., seq) represents the value of the i-th sequence point in the optimal path sequence. Finally, the improved sequence similarity distance dis is obtained by multiplying the penalty coefficient by the sum of the values ​​in the optimal path calculated by the original DTW. DTW .

[0109]

[0110] Step 2.2: Construct a matrix from the noise sequence similarity distances between each monitoring station and calculate its adjacency matrix to build a graph relationship topology. The specific process is as follows:

[0111] Step 2.2.1: Based on the noise sequence similarity distance information of each monitoring station, calculate the similarity distance matrix M (unit: km) between the stations. The matrix M is as follows:

[0112]

[0113] Where 'a' represents the number of monitoring stations;

[0114] For example, in this embodiment of the invention, a total of 11 monitoring stations are set up in the chemical industrial park, and the matrix formed by the stations is shown below.

[0115] Step 2.2.2: Calculate the standard deviation λ of the similarity distance matrix M. c Let μ be the value of the c-th element in the matrix, N be the number of elements in matrix M, and μ be the value of the c-th element. M The mean of the matrix is ​​calculated as follows:

[0116]

[0117] Step 2.2.3: Analyze the similarity between monitoring stations based on the similarity distance matrix M, and construct the adjacency matrix M using the standard deviation λ of all non-infinite numbers in the similarity distance matrix. d Adjacency matrix M d The elements in The calculation process is as follows:

[0118]

[0119] Where c represents the matrix element index, the larger the similarity distance, the better. The smaller the value, the lower the threshold should be. If the threshold is set to 0.1, then... If the similarity between the two sites is less than the threshold, they are considered to have too low a similarity and no mutual influence relationship. They are regarded as non-adjacent sites, and their weight in the adjacency matrix is ​​zero, so they do not constitute an adjacency relationship.

[0120] The above method and steps meet the similarity calculation requirements for periodic noise time series. Using this method to reconstruct spatial relationships lays the foundation for providing a more complete graph representation of relationships in subsequent graph convolutional neural network work.

[0121] Step 2.3: Introduce a graph convolutional neural network (GCN) to construct an adjacency matrix graph structure. Input the adjacency matrix graph structure into a diffusing convolutional recurrent neural network for noise prediction to obtain preliminary prediction results.

[0122] First, let's understand convolution from the perspective of signal processing. The word "convolve" originally means to flip, but in convolution it is translated as "convolve". "Convolution" itself is a kind of operation, meaning "product". Noise can be classified into discrete noise or continuous noise. Convolution generally refers to continuous signals. Sound monitoring is achieved through sensors. The noise collected by the sensors is a discrete value at 30-second intervals. Therefore, we will only discuss the convolution of discrete noise here.

[0123] Define the pulse signal as: n represents the number of displacements, l represents the discrete time, and the discrete noise x[n] can be represented by equation (9):

[0124]

[0125] Where x[l] represents the discrete noise at time l;

[0126] The discrete noise convolution process can be understood as first flipping the noise sequence, then shifting it, and finally multiplying and summing the results. This means that the system's response to the input discrete noise sequence x[n] is characterized by its unit impulse response h[n], and the convolution of the discrete noise convolution is the sum of the convolutions y[n]. dis [n] is from equation (10).

[0127]

[0128] In this process, the unit impulse response is first inverted to obtain h[-l] during convolution, and then shifted by n to obtain the shifted function h[nl].

[0129] In contrast, the definition of convolution in a Convolutional Neural Network (CNN) lacks the flipping process compared to the convolution principle described above; a direct shift of n yields h[n+l]. The convolution and y in a CNN... cnn [n] is represented by equation (11),

[0130]

[0131] It is not a complete convolution but rather a cross-correlation in signal processing; therefore, the convolution kernel is also a filter. The fundamental purpose of using convolution in convolutional neural networks is weighted summation and feature extraction; flipping is unnecessary.

[0132] Noise sequences from multiple monitoring stations can form a topological graph. This graph is irregular and lacks translation invariance, thus requiring the introduction of a Graph Convolutional Neural Network (GCN) to address graph relationship issues. In this embodiment, eleven monitoring stations are treated as eleven nodes, each with its own characteristics. All nodes form an 11×11 dimensional adjacency matrix A. Assuming the noise sequence length is Y, the node features are 11*Y dimensional feature matrices X, where X and A are the inputs to the graph convolutional network. The propagation mechanism between GCN network layers is given by equation (12), where L represents the Lth network layer.

[0133]

[0134] Where σ is a nonlinear activation function; W L It is a trainable weight matrix. The degree matrix is ​​constructed as follows: I represents the I-th row of adjacency matrix A, and J represents the J-th column of adjacency matrix A. This indicates that the site's own characteristic information is retained. 11 X represents an 11-dimensional identity matrix, λ is the site noise feature weight, λ = 1 indicates that the noise feature of this site is as important as the noise feature of the neighboring sites, H is the feature extracted by the current layer, and if it is the input layer, then X = H.

[0135] CNN models extract very limited features without training, while GCNs, using randomly initialized parameters, extract far superior features. Using GCNs to extract spatial relationships is currently the optimal method in research.

[0136] To formalize the problem of spatiotemporal noise sequence prediction, this invention uses a diffusing convolutional recurrent neural network (DCRNN) to model the spatiotemporal relationships of noise sequences, building upon the graph convolutional neural network (GCN). By associating noise monitoring stations with the diffusion process, the spatial relationships between stations are modeled, thus capturing the dynamic and stochastic nature of noise. The structure of the DCRNN model is as follows: Figure 6 As shown, DCRNN is a recurrent neural network based on diffusion graph convolution. The original input of the model is a spatiotemporal noise data structure that combines the adjacency matrix of the physical distance of the monitoring station distribution with the time series noise data. This invention proposes an improved DTW algorithm to calculate the similarity distance of the sequence and reconstruct the spatiotemporal relationship between stations based on the similarity distance. The advantage of this model is that diffusion convolution can take into account the walk relationship between noises, and an encoder is added to the historical noise sequence. The true value or predicted value of the previous moment is delayed by one time step as the input of the current decoder to predict the noise at the current time. Among them, ReLU is the activation function. If the input is positive, it will output directly; otherwise, it will output zero. The stationary distribution of the diffusion process can be represented as a weighted combination of infinite random walks on the graph. The diffusion process can be represented by Equation (13) and calculated in closed form:

[0137]

[0138] W is the node similarity matrix. It is the inverse of the out-degree matrix, β∈[0,1] represents the restart probability, where k is the diffusion degree and ε represents the probability of diffusion from the node. The diffusion process in the DCRNN model is bidirectional, and the environmental impact of the noise monitoring station on both sides can be considered.

[0139] Therefore, in terms of spatial relationships, based on the graph signal feature matrix With filter f θ (θ is the filter parameter) The bidirectional diffusion convolution operation can be defined as Equation (14):

[0140]

[0141] Where G represents graph G, p represents the p-th feature dimension, P represents the total number of feature dimensions, K represents the finite K-step truncation of the diffusion process, ★G represents the diffusion convolution on graph G, X:, p θ represents the convolution operation of all nodes with respect to the p-th feature. k,1 θ represents the kernel parameter used to calculate the output degree. k,2 This represents the kernel parameters used for in-degree calculation. and Let represent the transition matrices for diffusion and anti-diffusion, respectively. After defining the convolution operation, construct a diffusion convolutional layer (P-dimensional to Q-dimensional) with a mapping relationship as shown in equation (15):

[0142]

[0143] Where q represents the q-th feature, Q represents the total number of mapped features, and H:, q This represents the spread convolution operation performed on the q-th feature by all nodes, and the feature matrix... As input, the features extracted by the current layer For output, Let Θ be the parameter tensor and σ be the activation function. The spatial relationship based on bidirectional diffraction convolution is then constructed.

[0144] In terms of time, a combination of GRU and diffusing convolution is used to construct a diffusing convolution gated recurrent unit (DCGRU). Its essential principle is to replace matrix multiplication in GRU with diffusing convolution. Based on the LSTM principle, combined with the principles of GRU and diffusing convolution, DCGRU can be expressed as Equation (16):

[0145]

[0146] ★G represents the spread convolution of graph G, r (t) Θ represents the reset gate at time t. r★G The parameter tensor X represents the reset gate. (t) H represents the input at time t. (t-1) This represents the output at time t-1, b r The bias vector representing the reset gate; u (t) Θ represents the update gate at time t. u★G The parameter tensor of the update gate, b u This represents the bias vector of the update gate; C (t) Θ represents the hidden state of the next cell at time t. C★G The parameter tensor representing the hidden state, b C H represents the bias vector of the hidden state. (t) This represents the output at time t.

[0147] Spatiotemporal modeling maximizes the capture of spatiotemporal correlations in noise sequences through bidirectional propagation. However, current work relies on distance-based spatial correlations, and spatial features constructed based on physical distance cannot fully represent the correlations between noise sequences at different sites. Therefore, better methods are needed to address this issue. To improve the prediction accuracy of spatial relationships, a deeper understanding of spatial relationships is required to solve the problem.

[0148] Step 2.4: Extract the traffic flow characteristics at the gate corresponding to the current time in real time, and dynamically adjust the DCRNN noise prediction value using the Kalman filtering method.

[0149] Kalman filtering uses the estimate of the previous state to predict the current state. Finally, it uses the observations of the current state to correct the predicted value obtained in the prediction stage, thus obtaining a new estimate that is closer to the true value. Specifically, it calculates the Kalman gain of noise and drift by using prior estimation and covariance matrix prediction of normalized traffic flow features and noise sequence. The Kalman gain is then used to update the covariance matrix, thereby correcting the predicted noise value for the current time step.

[0150] Step 3: Train the constructed noise prediction network model and output the trained model. The specific process is as follows:

[0151] Historical data from each monitoring station was extracted to reconstruct the spatiotemporal relationships of the data using the improved DTW algorithm. The DTW-DCRNN noise prediction network model was trained, using 70% of the dataset as the training set, 10% as the validation set, and the last 20% as the test set. Furthermore, training results were not recorded for the first 60 epochs; from 60-100 epochs, training results were recorded and model parameters were output every 10 epochs; after 100 epochs, training results were recorded and model parameters were output every 5 epochs. Finally, training was stopped early based on the measured validation loss to ensure the model was captured before it overfitted.

[0152] Step 4: Collect monitoring station information in real time and predict noise in real time based on the trained model.

[0153] This invention first establishes a similarity distance matrix based on the improved DTW algorithm according to the location of each station, abandoning the traditional method of constructing a matrix based on physical distance. Instead, the similarity distance matrix is ​​constructed as an adjacency matrix representation. The spatiotemporal dataset based on similarity distance is fed into the DCRNN model to obtain the DTW-DCRNN model based on the spatial relationship of similarity distance. Finally, real-time traffic flow data is combined, and the prediction output of the DCRNN is dynamically updated through a Kalman filter.

[0154] Constructing an adjacency matrix using a similarity distance matrix to characterize spatial distance relationships can effectively compensate for the limitations of the original physical distance method. The prediction results adjusted by the Kalman filter are also closer to the true values ​​and have a compensatory property. The 3σ noise processing method effectively avoids large deviations in the model training prediction results.

[0155] The following experiments were conducted to demonstrate the feasibility and superiority of the present invention.

[0156] Experiment 1: Comparison Experiment of DTW Algorithm Before and After Improvement

[0157] Figure 7(a) shows an example of an adjacency matrix based on the original DTW algorithm, and the improved DTW algorithm as follows: Figure 7 (b) Compared to the original method, it increases the correlation between monitoring station 1 and monitoring station 3.

[0158] This invention uses Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Square Error (RMSE) as evaluation metrics. The DTW algorithm and the improved DTW algorithm were pre-trained with the DCRNN model, and the average prediction results over twelve time steps were calculated. The metrics are shown in Table 1. In the pre-trained model results, MAE noise decreased by 0.06 dB, accuracy improved by 2.8%; MAPE decreased by 0.11%, accuracy improved by 1.9%; and RMSE improved by 0.02 dB, accuracy decreased by 0.5%. The improved MAE and MAPE are both lower than the original method, while the RMSE is slightly higher. This means that the improved method has lower accuracy in predicting outliers and anomalies than the original method, but overall, the prediction accuracy of the new DTW algorithm is higher than that of the original DTW algorithm.

[0159] Table 1 Comparison of DTW algorithm before and after improvement

[0160]

[0161] Plot the indicators based on the time step, as follows: Figure 8 Further observation of the improved DTW algorithm and the original algorithm reveals that the horizontal axis represents the time step, and the vertical axis on the right side of the image represents the MAPE index. The graph shows that from the fourth time step onwards, the improved DWT algorithm exhibits a smaller mean absolute percentage error (MASE), indicating better performance in the MAPE index. The vertical axis on the left side of the image represents the RMSE and MAE indices, measured in decibels (dB). The graph shows that from the sixth time step onwards, the improved DWT algorithm exhibits a smaller root mean square error (RMSE), indicating better performance in the RMSE index. Furthermore, from the fourth time step onwards, the improved DWT algorithm exhibits a smaller mean absolute error (MAE), indicating better performance in the MAE index.

[0162] In summary, the improved DTW algorithm not only increases the overall accuracy of predictions without increasing the computational time cost, but also uncovers deeper sequence correlations. Therefore, it can be innovatively applied to graph convolution work to improve prediction accuracy.

[0163] Experiment 2: Comparison of the prediction model proposed in this invention with other models

[0164] To further verify the predictive performance of the method of this invention, three evaluation metrics—RMSE, MAE, and MAPE—were used to evaluate various methods or models, including the HA historical average prediction model, the VAR vector autoregressive model, the STGCN spatiotemporal convolutional neural network, the DCRNN diffusing convolutional recurrent neural network, DTW-STGCN, and DTW-DCRNN(N). The DTW-DCRNN(N) method was used without Kalman filtering and dynamic adjustment of traffic flow data. All experiments were based on predictions over 12 time steps. The average of the prediction results over these 12 time steps was used as the metric to measure the long-term predictive performance of the model. According to Table 2, the prediction results of this invention are superior to other prediction methods.

[0165] HA: Due to the strong periodicity of noise and the many abrupt change points, the accuracy of the prediction results of this model is low, but the training speed is fast.

[0166] STGCN: Due to the randomness of the noise sequence, the accuracy of noise prediction results is low.

[0167] DCRNN: This model outperforms STGCN. The time-series noisy data is highly time-dependent, which reflects that GRU has a greater advantage than CNN in time series modeling.

[0168] VAR: The prediction results of the vector autoregressive model are better than STGCN because the noise sequence has strong periodicity, indicating that neural networks do not always perform better.

[0169] DTW-STGCN: The DTW algorithm, when applied to the STGCN model, still outperforms the STGCN model in prediction. This indirectly verifies that considering spatial relationships from the perspective of time series similarity is meaningful.

[0170] DTW-DCRNN(N): Compared to DTW-DCRNN(N), the method of this invention adds Kalman filtering, which has a corrective and adjustment effect on the data, and its effect is the best. Compared with the original DCRNN model, the RMSE and MAE accuracy of the method of this invention are improved by 11.1% and 5% respectively, and MAPE is improved by 5.26%.

[0171] Table 2 Baseline Model Prediction Results

[0172]

[0173] In addition, the experiment also found that the method of the present invention performs very well in long-term prediction. The specific data is shown in Table 3. It can be seen that the prediction accuracy of the method of the present invention is lower than that of DCRNN in the first time step, but the performance becomes more stable and the accuracy is higher in the following 11 time steps.

[0174] Table 3 Analysis of Long-Term Forecast Results

[0175]

[0176] To facilitate observation of specific forecast trends, the table above will be visualized as an image. Figure 9 The graph shows the mean absolute error (MAE) and root mean square error (RMSE) curves. The right-hand vertical axis represents the RMSE index in decibels (dB). The graph shows that predictions are more accurate at the 2nd, 3rd, and 4th time steps. The left-hand vertical axis represents the MAE index in decibels (dB). Ignoring the first time step, the errors of both methods increase with the number of time steps, but the DTW-DCRNN method of this invention has smaller errors and higher accuracy. Figure 10 The mean absolute percentage error (MAPE) curve shows a similar trend to the MAE metric, but the DTW-DCRNN model of this invention exhibits a smaller mean absolute percentage error. Starting from the second time step (including the second time step), the method of this invention performs better in all three metrics.

[0177] In summary, the improved method of this invention not only improves the overall accuracy of prediction, but also has advantages over the original method in long-term noisy time series prediction. Furthermore, it can provide a new approach and method for graph neural networks and related research in spatiotemporal prediction.

[0178] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A DTW-DCRNN-based chemical industrial park noise prediction method, characterized in that, The improved time dynamic regularization algorithm is adopted, and the spatial relationship is reconstructed by monitoring the site graph structure, so as to realize the long-term space-time prediction of the noise in the chemical industry park, which includes the following steps: Step 1, collect and process the monitoring site information and vehicle information data in the chemical industry park; Step 2, introduce a penalty coefficient to improve the original time dynamic regularization algorithm, and introduce it into the diffusion convolution recurrent neural network, and then combine Kalman filtering to construct a noise prediction network model; the specific process is as follows: Step 2.1, introduce a penalty coefficient to improve the original time dynamic regularization algorithm, and calculate the sequence similarity distance; The principle of the improved time dynamic regularization algorithm is to calculate the penalty coefficient by obtaining the optimal path and the number of common subsequences, and then calculate the similarity distance between the stations; the specific process is as follows: Step 2.1.1, calculate the distance matrix; Step 2.1.2, find the optimal path of the matrix; Step 2.1.3, calculate the number of common subsequences and the length of the optimal path sequence, and then obtain the weight to obtain the penalty coefficient; Among them, the length of the optimal path sequence is the length of the optimal path matrix finally obtained in step 2.1.2; The process of calculating the number of common subsequences is as follows: first, construct a record empty list record, loop through sequences s1 and s2, and record the common sequence length ls1 or ls2 and the sequence position g when a common sequence appears, and extract the common subsequence from the s1 sequence and place it in the s_sum array for storage; After the traversal is completed, the length of s_sum is the number of common subsequences; When calculating the number of common subsequences, set the threshold to 1, and count the common subsequences with a length greater than 1 as common subsequences, and finally count the total number of common subsequences as the number of common noise subsequences required; The calculation formula of the weight w is as formula (3): （3） Where, subseq is the length of the common noise subsequence, and seq is the length of the optimal path matrix; The calculation process of the penalty coefficient a is shown in equation (4), where x represents the number of common subsequences, w i represents the weight of the i-th sequence point, （4） Step 2.1.4, express the optimal path as (r1, r2, …, r seq ), r i (i = 1, 2, …, seq) represents the value of the i-th sequence point in the optimal path sequence, and finally the improved sequence similarity distance is obtained by multiplying the sum of the values in the original DTW calculated optimal path by the penalty coefficient ; （5） Step 2.2, construct a matrix of the noise sequence similarity distance between each monitoring station and calculate it as an adjacency matrix to construct a graph relationship topology; Step 2.3, introduce a graph convolutional neural network GCN to construct an adjacency matrix graph structure, input the adjacency matrix graph structure into the diffusion convolution recurrent neural network for noise prediction, and obtain the preliminary prediction result; Step 2.4, real-time extract the traffic characteristics of the current time corresponding to the gate, and dynamically adjust the DCRNN noise prediction value using Kalman filtering method; Kalman filtering uses the estimation of the previous state to make a prediction of the current state; Finally, the observation value of the current state is used to correct the prediction value obtained in the prediction stage to obtain a new estimated value closer to the true value; The specific implementation is to calculate the noise and drift Kalman gain by predicting the prior estimation and covariance matrix of the normalized traffic characteristics and noise sequence, and then correct and update the covariance matrix with the Kalman gain to correct and update the noise value at the current time; Step 3, train the constructed noise prediction network model, and output the trained model; Step 4, real-time collect the monitoring site information and perform real-time noise prediction based on the trained model.

2. The DTW-DCRNN-based chemical industrial park noise prediction method according to claim 1, characterized in that, The step 1, the chemical industry park is provided with a barrier gate and a plurality of monitoring sites for real-time data collection, and the position information data of each monitoring site is recorded synchronously; the barrier gate and each monitoring site are provided with sensors; the barrier gate collects information data of vehicles entering and leaving, including license plate number, vehicle entering and leaving time, vehicle flow characteristics of vehicle type; The monitoring site real-time collects noise and natural environment data in the chemical industry park; The site location data, noise and natural environment data, and vehicle information data are transmitted to the gateway device, and then interactively transmitted to the database server. When performing noise prediction, real-time data extraction is performed, and the data is processed using the 3 criteria.

3. The DTW-DCRNN-based chemical industrial park noise prediction method according to claim 1, characterized in that, The step 2.1.1 specifically comprises: Suppose there are two noise sequences S and T, as formula (1): （1） S is of length o, T is of length m, , a distance matrix d[s][t] is constructed based on the noise sequence m x o, wherein the distance between the i-th sequence point s i (i = 1, 2, …, o) and the j-th sequence point t j (j = 1, 2, …, m) is the value of the matrix (s i ,t j ) element, the Euclidean distance: dis(s i ,t j )= (s i -t j ) 2 , the matrix distance calculation standard is shown in equation (2): （2） The step 2.1.2 specifically is: the optimal path is the path with the minimum accumulated distance satisfying the lower boundary condition, and the lower boundary condition is: starting from the right upper corner d[o][m] point of the matrix, finding the minimum point among the left lower three points as the next node until the left lower corner d[0][0] matrix point is reached, in addition to ensuring continuity and monotonicity, it cannot jump over or omit a certain point for matching, and can only align or match with adjacent points, and the order of time sequence cannot be changed; The specific process of the step 2.2 is as follows: Step 2.2.1, according to the noise sequence similarity distance information of each monitoring site, the similarity distance matrix M between the sites is calculated, and the matrix M is as follows: （6） Wherein, a is the number of monitoring sites; Step 2.2.

2. Calculate the standard deviation of the similarity distance matrix M , is the value of the cth element in the matrix, N is the number of elements in the matrix M, is the mean of the matrix, and the calculation process is as follows: （7） Step 2.2.

3. Analyzing the similarity between monitoring sites from the similarity distance matrix M, using the standard deviation of all non-infinite numbers in the similarity distance matrix Constructing the adjacency matrix , the adjacency matrix The calculation process of each element in the adjacency matrix is as follows: （8） Where c represents the matrix element index, the larger the similarity distance, the better. The smaller the value, the lower the threshold should be. If the threshold is set to 0.1, then... If the similarity between the two sites is too low and there is no mutual influence, they are considered non-adjacent sites with a weight of zero in the adjacency matrix and do not constitute an adjacency relationship. The specific process of the step 2.3 is as follows: The propagation mode between the network layers of the graph convolutional neural network GCN is formula (12), and L represents the Lth network layer: （12） wherein, is a non-linear activation function; is a trainable weight matrix, is a degree matrix, which is constructed as: , I denotes the I-th row of the adjacency matrix A, J denotes the J-th column of the adjacency matrix A, denotes the feature information of the reserved station itself, H is the feature extracted by the current layer, and if it is an input layer, X = H; On the basis of the GCN graph convolutional neural network, a diffusion convolution recurrent neural network is used to model the space-time relationship of the noise sequence; The stationary distribution of the diffusion process is represented as a weighted combination of infinite random walks on the graph, and the diffusion process is represented as formula (13) and calculated in a closed form: （13） where W is the node similarity matrix, is the inverse matrix of the out-degree matrix, denotes the restart probability, k is the diffusion degree, denotes the possibility of diffusion from the node, the diffusion process in the DCRNN model is bidirectional; Therefore, in the spatial relationship, based on the graph signal feature matrix and filter The bidirectional diffusion convolution operation is defined as formula (14): （14） where, is the filter parameter, G denotes the graph G, p denotes the p-th feature dimension, P denotes the total number of feature dimensions, K denotes the finite K-step truncation of the diffusion process, ⋆G denotes the diffusion convolution on graph G, denotes the convolution operation on all node pairs for the p-th feature, denotes the convolution kernel parameter for out-degree calculation, denotes the convolution kernel parameter for in-degree calculation, and denote the over-matrix for diffusion and inverse diffusion respectively, after defining the convolution operation, the diffusion convolution layer with mapping relationship is constructed as formula (15): （15） where q denotes the qth feature, Q denotes the total number of mapped features, denotes the qth feature diffusion convolution operation for all nodes, the feature matrix is the input, the feature extracted by the current layer is the output, denotes the filter, is the parameter tensor, is the activation function; In terms of time relationship, a diffusion convolution gated recurrent unit DCGRU is constructed by combining GRU and diffusion convolution, and the DCGRU is represented as formula (16): （16） ⋆G denotes a diffusion convolution over the graph G, denotes the reset gate at time t, denotes the parameter tensor of the reset gate, denotes the input at time t, denotes the output at time t-1, denotes the bias vector of the reset gate; denotes the update gate at time t, denotes the parameter tensor of the update gate, denotes the bias vector of the update gate; denotes the hidden state of the next cell at time t, denotes the parameter tensor of the hidden state, denotes the bias vector of the hidden state; denotes the output at time t.

4. The DTW-DCRNN-based chemical industrial park noise prediction method according to claim 1, characterized in that, The specific process of the step 3 is as follows: Extract the historical data of each monitoring site for the improved DTW algorithm to reconstruct the space-time relationship of the data, train the noise prediction network model DTW-DCRNN, select 70% of the data set as the training set, 10% as the validation set, and the last 20% as the test set; in addition, no training results are recorded in the first 60 Epochs, and the training results are recorded once every 10 Epochs in the 60-100 Epochs, and the model parameters are outputted, and the training results are recorded once every 5 Epochs after 100 Epochs, and the model parameters are outputted; finally, the training is stopped in advance according to the measured validation loss, so as to capture the model when the model is about to be over-fitted.

Images