Safety monitoring method of hazardous chemicals based on big data

Abstract

The present invention relates to the technical field of hazardous chemical safety monitoring methods, and more specifically, to a hazardous chemical safety monitoring method based on big data, comprising step 1. heterogeneous sensor data acquisition and preprocessing; step 2. associated spatiotemporal feature extraction; step 3. dual autoencoder dictionary learning: designing a dual sparse autoencoder network with joint local-global constraints, while minimizing data reconstruction error and dictionary size, and using an alternating iterative proximal gradient descent algorithm to solve the objective function; step 4. dictionary parameter optimization based on quantization coding; step 5. multi-level joint decoding; step 6. abnormality diagnosis based on enhanced reconstruction sequence, achieving multiple synergistic efficiencies in multimodal heterogeneous data compression, spatiotemporal association modeling, dictionary learning and encoding, reconstruction optimization, and abnormality diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of hazardous chemical safety monitoring methods, and more specifically, to a hazardous chemical safety monitoring method based on big data. Background Art

[0002] With the rapid development of modern industry, the production scale of hazardous chemicals continues to expand, the production process is becoming increasingly complex, and the risks and challenges faced by safe production are also increasing. At present, high-risk industries such as petroleum, chemical, and coal generally use automatic control systems such as DCS and SIS, which use sensors to collect key process parameters such as temperature, pressure, flow, and liquid level in real time, and set safety thresholds to achieve abnormal alarms and chain control.

[0003] However, the traditional safety monitoring methods based on rules and thresholds have many limitations: First, large-scale chemical plants usually contain thousands of monitoring points, and the signals of various sensors are complicated. It is difficult to fully consider complex working conditions when setting thresholds manually, and there is a lack of adaptive update mechanism for thresholds, which leads to frequent false alarms and missed alarms. Secondly, there is an intricate coupling relationship between production parameters. The over-limit of a single variable does not necessarily lead to an accident, but the coupling of slight abnormalities of multiple parameters may lead to serious consequences. Traditional methods ignore the dynamic relationship between parameters and lack overall control of production status. Thirdly, process fluctuations and measurement noise are common in actual production, which seriously interfere with the accuracy of abnormal judgment. Traditional methods lack an effective identification mechanism for noise and abnormalities, making it difficult to achieve early warning of faults.

[0004] In recent years, academia and industry have actively explored new methods for intelligent safety monitoring. The rise of machine learning and data mining technology has provided new ideas for deep mining of massive production data. A representative method is multivariate statistical process control (MSPC), which extracts explicit variables of monitoring data through dimensionality reduction methods such as principal component analysis (PCA) and partial least squares (PLS), and constructs statistical indicators such as T2 and SPE to characterize the overall variation of the production process. This type of method takes into account the correlation between variables to a certain extent and can capture small drifts more sensitively. However, the MSPC method generally uses simple linear mapping, which is difficult to characterize complex nonlinear relationships, and relies on the normal distribution assumption of data. It is sensitive to noise and outliers, and its practical application effect is limited.

[0005] Another type of method is knowledge-based fault diagnosis, which extracts the correspondence between fault symptoms and production mechanisms through manual summary or machine learning, builds fault trees, diagnostic rule bases, etc., and realizes fault tracing and safety warning. This type of method integrates engineering experience, and the diagnostic results are explainable, but it is highly dependent on domain knowledge, and the knowledge acquisition and update cycle is long, making it difficult to cope with complex faults and frequent changes in production status.

[0006] It is worth mentioning that in recent years, deep learning methods have made great progress in the field of anomaly detection. In 2016, Malhotra et al. proposed an LSTM-based time series prediction model EncDec-AD, which predicts the output of the next M steps through the observations of the previous N time steps, and uses the reconstruction error as the anomaly score. This method can characterize the long-range dependency of time series, but does not consider the correlation between different physical quantities, and full supervision training requires a large amount of labeled data. In 2018, Kieu et al. proposed a VAE-LSTM hybrid model, which introduced a variational autoencoder to constrain the LSTM hidden state and enhance its generalization, but the encoding process lacks interpretability and is easily disturbed by outliers. In 2019, Zhang et al. proposed a twin network TCN for time series anomaly detection, which characterized the dynamic evolution characteristics of time series by measuring the similarity between different subsequences, but the network structure is single and lacks modeling of global spatiotemporal features.

[0007] In summary, the existing anomaly detection methods have broken through the limitations of traditional threshold monitoring to a certain extent, but the following challenges still exist: 1) It is difficult to effectively integrate multi-source heterogeneous data, and there is a lack of global correlation representation of production status; 2) It lacks robustness to small sample anomalies and is easily disturbed by noise and outliers; 3) The efficiency of data compression and real-time decoding is insufficient, and it is difficult to meet the needs of massive sensor data monitoring; 4) The ability of reconstructing errors to characterize anomalies is limited, and there is a lack of sensitive capture of small offset trends.

[0008] Therefore, there is an urgent need to explore a new paradigm of abnormal diagnosis driven by heterogeneous data, to achieve early intelligent discovery of abnormalities under complex working conditions through deep mining and mechanism fusion of industrial big data, and to improve the inherent safety level of the hazardous chemicals production process. Summary of the invention

[0009] In response to the above-mentioned technical problems, the present invention proposes a hazardous chemical safety monitoring method based on big data, aiming to break through the limitations of traditional anomaly detection methods, and innovatively integrates cutting-edge artificial intelligence technologies such as multi-source heterogeneous data feature extraction, dictionary learning and encoding, variational inference and metric learning, to build an efficient, collaborative, hierarchical and progressive anomaly diagnosis system framework, providing new solutions for ensuring the safety of hazardous chemical production.

[0010] The present invention provides a method for safety monitoring of hazardous chemicals based on big data, comprising the following steps:

[0011] Step 1: Heterogeneous sensor data collection and preprocessing: collect various sensor data in the production process, and perform spatiotemporal alignment, noise removal, missing value interpolation, and normalization preprocessing operations to form a unified spatiotemporal data matrix;

[0012] Step 2. Extraction of correlation spatiotemporal features: A correlation-enhanced recurrent neural network is constructed using a bidirectional gated recursive unit to adaptively model the dynamic correlation features of multi-source time series data, and differential hidden layer states are introduced to characterize short-term multi-scale dynamics;

[0013] Step 3. Dual autoencoder dictionary learning: Design a dual sparse autoencoder network with joint local-global constraints, minimize the data reconstruction error and dictionary size, and use the alternating iterative proximal gradient descent algorithm to solve the objective function;

[0014] Step 4. Dictionary parameter optimization based on quantization coding: perform product quantization on the sparse dictionary, minimize the quantization coding loss by alternately optimizing the codebook and sparse coding, and use the alternating least squares algorithm to solve;

[0015] Step 5. Multi-level joint decoding: At the monitoring end, the low-rank dictionary is restored through the codebook tensor product operation, and then the local reconstruction residual and global prior knowledge are combined to adaptively compensate for the quantization loss to obtain the state estimate;

[0016] Step 6. Anomaly diagnosis based on enhanced reconstruction sequence: For the reconstructed state data, an anomaly diagnosis algorithm combining variational autoencoder and twin network is designed to adaptively determine the degree of abnormality of each time slice through reconstruction error and similarity score.

[0017] Specifically, the correlation spatiotemporal feature extraction step uses a bidirectional gated recurrent unit (BiGRU) to construct a correlation enhanced recurrent neural network (CorrRNN), adaptively modeling the dynamic correlation features of multi-source time series data, and the forward and backward hidden layer states in the correlation enhanced recurrent neural network are and Calculated by:

[0018]

[0019] Among them, X t ∈R D is the multi-sensor input vector at time t, h t ∈R K is the hidden layer state at time t, Represents the vector concatenation operation, ⊙ represents the Hadamard product; the gating function in the gated recursive unit unit adopts the gated linear unit form:

[0020] z t =σ(W z X t +U z h t-1 +b z )

[0021] r t =σ(Wr X t +U r h t-1 +b r )

[0022]

[0023] Among them, z t is the update gate, r t To reset the gate, is the candidate hidden layer state, σ(·) is the sigmoid activation function, and the differential hidden layer state ΔS is introduced t =h t -h t-1 To characterize short-term multi-scale dynamics.

[0024] Specifically, the dual autoencoder dictionary learning step is based on the spatiotemporal feature matrix H∈R extracted by the associated enhanced recurrent neural network N×T×K Based on this, a double sparse autoencoder network (DoubleSAE) with joint local-global constraints is designed to minimize the data reconstruction error and dictionary size. The objective function is as follows:

[0025]

[0026] Among them, ||·||2,1 is the L2,1 mixing norm, D∈R K×M To complete the dictionary, is the sparse coding coefficient, λ1 and λ2 are the sparse regularization parameters of samples and atoms respectively, is the atomic vector of the i-th column of D; the above problem is solved by the alternating iterative proximal gradient descent algorithm (ISTA), and the update formulas of A and D are:

[0027]

[0028]

[0029] Where, f(A)=||H-DA|| 2，1 , g(D)=||H-DA|| 2，1 +λ2||D|| 2，1 , L is the Lipschitz constant of f(A), γ is the update step size of D, j is the number of iterations, SoftThresh α (·) is the soft threshold function, and its component form is: SoftThresh α (x) i =sign(x i )max(|x i |-α, 0).

[0030] Specifically, the dictionary parameter optimization step based on quantization coding is to optimize the sparse dictionary Perform product quantization and express it as the Cartesian product of multiple codebooks:

[0031]

[0032] in, is the lth codebook, represents the Kronecker product, is the codebook dimension, n l is the number of codewords in the lth codebook; under the premise of ensuring the lower bound of the reconstruction error ∈, the codebook is optimized alternately and sparse coding Minimize the quantization coding loss, the optimization goal is:

[0033]

[0034] Alternating least squares (ALS) algorithm is used to solve: Fixed Then A is obtained through LASSO regression:

[0035]

[0036] Fixed A, then each codebook C l Update via low-rank tensor decomposition:

[0037]

[0038] in, The above subproblems can be solved by the high-order power method (HOPM).

[0039] Specifically, the multi-level joint decoding step is at the monitoring end, given the quantization codebook And sparse code A, recover the low-rank dictionary through codebook tensor product operation:

[0040]

[0041] Then, the local reconstruction residual and global prior knowledge are combined to adaptively compensate for the quantization loss and obtain the state estimate:

[0042] H′=D′A+ΞΘ(HD′A)+ΨΦ(D′A)

[0043] Among them, H∈R N×T×K is the state prior mean, Ξ∈R K×K and Ψ∈R K×K are the local and global compensation matrices, respectively,

[0044] Θ(·) and Φ(·) are the corresponding gated activation functions:

[0045] ΞΘ(E)=(σ(W θ E+b θ )⊙E)W ξ

[0046] ΨΦ(Z)=(σ(W φ Z+b φ )⊙Z)W ψ .

[0047] Specifically, the abnormality diagnosis step based on the enhanced reconstruction sequence is for the reconstructed state data H′∈R N×T×K , design an abnormality diagnosis algorithm that combines variational autoencoders and twin networks; first, the state data is probabilistically encoded and generated through a variational autoencoder (VAE):

[0048]

[0049] p θ (H′|z)=Bern(H′;μ θ (z))

[0050] Among them, z is a hidden variable, φ and θ are the parameters of the encoder and decoder respectively, and conditional LSTM is used to implement the mean and variance network Then build a twin network to evaluate the similarity between the reconstructed sequence and the original sequence:

[0051] E i =CNN(BiLSTM(H′ i :i+w))

[0052] L s (E i , E j )=||E i -E j || 2

[0053] Among them, H′ i :i+w is the reconstructed subsequence within the i-th time window, width w, CNN(·) is the convolutional neural network feature extractor, L s (·) is the similarity measure of the twin network; finally, the reconstruction error and similarity score are combined to adaptively determine the abnormality of each time slice:

[0054] s i =||H′ i -H i || 2 +λL s (E i , E j)

[0055]

[0056] Among them, u i ∈[0, 1] is the abnormal probability of the ith time slice, and ProbAE(·) is the autoencoder probability estimation network.

[0057] Specifically, in the heterogeneous sensor data collection and preprocessing step, the heterogeneous sensor data includes temperature, pressure, flow, liquid level, component process parameters, and environmental perception data such as video and image; the collected heterogeneous data are subjected to spatiotemporal alignment, noise removal, missing value interpolation, and normalization preprocessing operations to obtain a unified spatiotemporal data matrix. Where N is the number of batch samples, T is the time step, and D is the sensor dimension.

[0058] Specifically, the measuring range of the pressure transmitter is 0-10MPa, and the accuracy is better than 0.075%; the flow rate is measured by a flowmeter or an electromagnetic flowmeter, and the measuring range of the flowmeter is 0-5m / s, and the accuracy is better than 0.5%; the nominal diameter of the electromagnetic flowmeter is DN80~DN300, and the accuracy is better than 0.3%.

[0059] Specifically, the video capture uses a high-definition network camera with a resolution of no less than 4 million pixels, a frame rate of no less than 30fps, and an encoding format of H.265.

[0060] Specifically, a hazardous chemical safety monitoring device based on big data and an edge computing gateway communicatively connected to the device are used. The CPU main frequency of the edge computing gateway is not less than 1.6 GHz, the memory is not less than 8 GB, the storage capacity is not less than 128 GB, and the communication mode supports 4G LTE or 5G LTE.

[0061] The present invention has the following beneficial effects:

[0062] First, the deep fusion of multi-source heterogeneous sensor data realizes the complementarity and gain of global information. Different physical quantities reflect different aspects of the production status, and single modal data is difficult to fully characterize complex working conditions. The present invention adopts strategies such as spatiotemporal alignment and co-domain projection to organically integrate environmental perception data such as video and images with process parameters such as temperature and pressure, establishes a unified spatiotemporal feature matrix, presents a complete picture of the production status from multiple angles, and enhances the information basis for abnormal diagnosis.

[0063] Secondly, the correlation spatiotemporal feature extraction module adaptively learns multi-scale dynamic correlations through a bidirectional gated recurrent network, revealing the temporal dependence and interactive influence of production parameters. Compared with traditional fixed time-delay analysis, this method can characterize long-term and short-term memory, capture local details and global trends, and enhance the dynamic characterization ability of production status. At the same time, the introduction of differential hidden layer states models the dynamic coupling of different physical quantities, explores potential abnormal driving modes, and improves the ability to capture fault precursors.

[0064] Thirdly, dual autoencoding dictionary learning achieves efficient compression and semantic extraction of massive heterogeneous data. By minimizing the local reconstruction error, a compact overcomplete dictionary is learned to describe the diversity of production status; through the global dictionary sparse constraint, the atom selection is further optimized and the dictionary redundancy is reduced. The synergy of the two improves the data compression rate while ensuring the reconstruction accuracy, enhancing the semantic expression and noise robustness of the dictionary, and laying a solid foundation for subsequent diagnosis.

[0065] At the same time, product quantization coding decomposes the dictionary into tensor product form, which significantly reduces the storage and communication load in the encoding stage. Through codebook learning, dictionary atoms are mapped to low-dimensional subspaces, and then reconstructed through Cartesian product combinations, achieving a high multiple of parameter reduction while ensuring approximation accuracy. This "time-for-space" strategy breaks through the bottleneck of massive data compression and real-time decoding, and improves the system's rapid response capability.

[0066] In addition, multi-level joint decoding corrects the reconstruction bias introduced by quantization error in an adaptive compensation manner. Through the interactive fusion of local and global information, the reconstruction result is smoothed using the prior mean of historical data statistics, which suppresses quantization noise while retaining local details. The design of the gating unit further gives the network data adaptive capabilities, and can autonomously adjust the local and global emphasis according to the dynamic changes of the input distribution, achieving a good balance between reconstruction accuracy and robustness.

[0067] Finally, the anomaly diagnosis model cleverly combines variational inference and metric learning paradigms, and describes the early signs of production state deviation from two complementary perspectives: data reconstruction and time series evolution. The variational autoencoder learns the probability distribution of the state sequence through latent space mapping, and probabilistically describes the reconstruction error, which can quantify the uncertainty of the state. The twin network uses the original data as a reference to evaluate the similarity of the reconstructed sequence and capture the dynamic deviation trend of the state. The synergy and complementarity of the two significantly improves the detection rate of early minor anomalies, reduces the false alarm rate, and makes up for the shortcomings of the traditional diagnostic method based on fixed thresholds.

[0068] In summary, the present invention has achieved multiple synergies, complementary advantages, and bottleneck breakthroughs in multimodal heterogeneous data compression, spatiotemporal correlation modeling, dictionary learning and encoding, reconstruction optimization, and abnormal diagnosis through a series of interlocking and progressive method innovations, forming a complete abnormal diagnosis technology system. This design idea of ​​full-process integration and end-to-end optimization has produced an overall emergence effect of "1+1>2" compared to traditional single module optimization, and has a milestone significance in improving the safety monitoring level of hazardous chemicals production processes. At the same time, the present invention adopts a modular and parameterized system architecture with strong versatility, scalability, and flexibility, opening up broad prospects for industrial big data mining and safe intelligent applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The figure is a flow chart of the method for safety monitoring of hazardous chemicals based on big data of the present invention. DETAILED DESCRIPTION

[0070] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, characteristics and effects of the hazardous chemical safety monitoring method based on big data proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0071] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0072] The specific scheme of the hazardous chemical safety monitoring method based on big data provided by the present invention is described in detail below with reference to the accompanying drawings.

[0073] Example 1

[0074] See also Figure 1 The present invention proposes a method for safety monitoring of hazardous chemicals based on big data. The method integrates multi-source heterogeneous sensor data, constructs spatiotemporal correlation features, and adopts strategies such as dual self-encoding dictionary learning and quantitative coding optimization to achieve global dynamic monitoring and abnormal diagnosis of the production process. Specifically, the method includes the following steps:

[0075] Step 1: Heterogeneous sensor data collection and preprocessing. This step collects and fuses multi-source sensor data from all links of hazardous chemicals production, including process parameters such as temperature, pressure, flow, liquid level, components, and environmental sensor data such as video and images. Among them, the pressure transmitter can choose a model with a measurement range of 0-10MPa and an accuracy better than 0.075%; the flow meter can choose a model with a measurement range of 0-5m / s and an accuracy better than 0.5%; the electromagnetic flow meter can choose a model with a nominal diameter of DN80-DN300 and an accuracy better than 0.3%. For video acquisition, a high-definition network camera with a pixel of not less than 4 million, a frame rate of not less than 30fps, and an encoding format of H.265 can be used.

[0076] The collected heterogeneous data is first aligned in time and space, and data from different sources and with different timestamps are synchronized according to a unified time axis and spatial coordinate system. Then, methods such as wavelet transform and EMD are used to denoise the data and eliminate high-frequency interference and background noise. For data with a large number of missing values, algorithms such as interpolation and matrix completion can be used to repair them to ensure data integrity. Finally, the heterogeneous data is normalized to eliminate the influence of different physical dimensions and uniformly map the data to the [0, 1] interval.

[0077] After the above preprocessing, a unified spatiotemporal data matrix can be obtained Where N is the number of batch samples, T is the time step, and D is the sensor dimension. This matrix provides a standardized input for subsequent feature extraction and anomaly diagnosis.

[0078] Step 2: Extraction of correlation spatiotemporal features. In order to fully explore the intrinsic correlation of multi-source heterogeneous data and capture the dynamic evolution of the production process, this step designs a correlation enhanced recurrent neural network CorrRNN. The network uses the bidirectional gated recurrent unit BiGR as the basic component, introduces the differential hidden layer state ΔS, characterizes short-term multi-scale dynamics, and enhances the modeling ability of complex spatiotemporal correlations.

[0079] Specifically, let the multi-sensor input vector at time t be The hidden state is Hidden state of the forward and backward layers in CorrRNN Calculation and update through gating function:

[0080]

[0081] In the above formula, z t is the update gate, r t To reset the gate, is the candidate hidden layer state, σ(·) is the sigmoid activation function. The bidirectional hidden layer states are fused in parallel to obtain the final hidden layer state at time t:

[0082]

[0083] Finally, by and The spatiotemporal feature matrix It can comprehensively characterize the dynamic correlation pattern of the production process.

[0084] Step 3: Double autoencoder dictionary learning. In order to further extract high-level semantic features and achieve data dimensionality reduction and compression, this step designs a double sparse autoencoder network DoubleSAE with joint local-global constraints based on the correlation feature H. The network learns a dictionary that is both expressive and robust by minimizing both data reconstruction error and dictionary size.

[0085] Specifically, the complete dictionary is D∈R K×M , the sparse coding coefficient is A∈R M×NT The objective function of DoubleSAE is:

[0086]

[0087] Among them, ||·|| 2，1 is the L2,1 mixing norm, λ1 and λ2 are regularization parameters, d i is the atomic vector of the i-th column of dictionary D. The above optimization problem can be solved by the alternating iterative ISTA algorithm, and the update formulas of A and D are:

[0088]

[0089] Where, f(A)=||H-DA|| 2，1 , g(D)=||H-DA|| 2，1 +λ2||D|| 2，1 , L is the Lipschitz constant of f(A), γ is the update step size of D, j is the number of iterations, SoftThresh α (·) is a soft threshold operator, and its component form is:

[0090] SoftThresh α (x) i =sign(x i )max(|x i |-α, 0)

[0091] After J iterations, the optimized dictionary D can be obtained. * and sparse coding A *, achieving efficient feature compression. This step innovatively introduces the global dictionary sparse regularization term, which not only extracts the local optimal feature representation, but also collaboratively learns a simpler global dictionary, further improving the data compression performance.

[0092] Step 4: Dictionary parameter optimization based on quantization coding. In order to further reduce the storage overhead of the dictionary and improve data compression efficiency, this step optimizes the sparse dictionary. Perform product quantization and express it as the Cartesian product of multiple codebooks:

[0093]

[0094] in, is the lth codebook, represents the Kronecker product, is the codebook dimension, n l is the number of codewords in the lth codebook. Under the premise of ensuring the lower bound of the reconstruction error ∈, the codebook is optimized alternately. and sparse coding Minimize the quantization coding loss, the optimization goal is:

[0095]

[0096] This problem can be solved by alternating least squares (ALS) method. l}, then the sparse code A can be obtained through LASSO regression:

[0097]

[0098] Fixed A, then each codebook C l It can be updated by low-rank tensor decomposition:

[0099]

[0100] Among them, H (k) is the residual of the kth iteration:

[0101]

[0102] The above subproblems can be solved efficiently by the high-order power method (HOPM). Usually L = 3 to 4, n l =2 8 to 2 12 , while ensuring the reconstruction accuracy, the dictionary parameters can be compressed to 1 / 100 to 1 / 10000 of the original.

[0103] Step 5: Multi-level joint decoding. In this step, the monitoring terminal receives the quantization codebook Decode the sparse code A and restore the production state. First, the low-rank dictionary is losslessly reconstructed through the tensor product operation of the codebook:

[0104]

[0105] Then, the local reconstruction residual and global prior knowledge are combined to adaptively compensate for the quantization loss and obtain the state estimate:

[0106] H′=D′A+ΞΘ(HD′A)+ΨΦ(D′A)

[0107] in, is the state mean of historical data statistics, and are the local and global compensation matrices, Θ(·) and Φ(·) are the corresponding gated activation functions:

[0108] ΞΘ(E)=(σ(W θ E+b θ )⊙E)W ξ

[0109] ΨΦ(Z)=(σ(W φ Z+b φ )⊙Z)W ψ

[0110] The above method combines local reconstruction residuals and global prior knowledge, significantly improving the accuracy and robustness of reconstruction.

[0111] Step 6: Abnormal diagnosis based on enhanced reconstruction sequence. This step is to reconstruct the state data Perform anomaly detection and diagnosis. First, a variational autoencoder (VAE) is used to perform probability modeling on the data:

[0112]

[0113] p θ (H′|z)=Bern(H′;μ θ (z))

[0114] Among them, z is a hidden variable, q φ (z|H′) is the posterior distribution, p θ (H′|z) is the likelihood distribution. The parameters φ and θ of the encoder and decoder are implemented through the conditional LSTM network, which can characterize the temporal dependency of the state sequence.

[0115] Then, the twin network is introduced to measure the similarity between the reconstructed data and the original data. The twin network consists of two sub-networks with shared parameters, which respectively map the input data to the feature space and evaluate the similarity through the distance metric:

[0116] E i =CNN(BiLSTM(H′ i :i+w))

[0117] L s (E i , E j )=||E i -E j || 2

[0118] Among them, E i , E j is the feature vector of the i-th, j-th reconstruction window, CNN is a convolutional neural network, BiLSTM is a bidirectional long short-term memory network, H′ i :i+w represents the i-th window, the width is the reconstructed subsequence within w, L s (·) is the Euclidean distance.

[0119] Finally, the reconstruction error and twin similarity are combined to adaptively determine the degree of abnormality of each time slice:

[0120] s i =||H′ i -H i || 2 +λL s (E i , E j )

[0121]

[0122] Among them, u i ∈[0, 1] is the anomaly probability of the i-th time slice, ProbAE(·) is the autoencoder probability estimation network, which can learn the prior distribution of anomaly scores. i >τ (such as τ = 0.95), it is judged as abnormal and an alarm is triggered. Through online segment-by-segment sliding diagnosis, abnormal deviations in production status can be discovered in a timely manner, providing a decision-making basis for accident prevention. This step innovatively combines variational inference and metric learning paradigms, and through the joint evaluation of reconstruction error and twin similarity, characterizes abnormal features from the two perspectives of data reconstruction and time series evolution, overcoming the limitations of traditional threshold-based rules and significantly improving the timeliness and accuracy of abnormal diagnosis.

[0123] The method for safety monitoring of hazardous chemicals based on big data provided by the present invention has the following advantages:

[0124] (1) Through the deep fusion of cross-modal heterogeneous data, global spatiotemporal correlation features are constructed, providing a new perspective for the overall characterization of production status;

[0125] (2) Dual autoencoding dictionary learning is used to extract implicit multi-scale patterns from massive heterogeneous data, achieving efficient data compression and semantic representation;

[0126] (3) The introduction of product quantization coding significantly reduces dictionary storage and transmission overhead while ensuring reconstruction accuracy, thereby improving the real-time performance of the system;

[0127] (4) Fusion of variational inference and metric learning to characterize anomalies from the perspectives of data reconstruction and time series evolution, thus improving the accuracy of fault diagnosis;

[0128] (5) A hierarchical and progressive process design is adopted to construct an integrated abnormal diagnosis paradigm of "perception-understanding-judgment", which can significantly improve the inherent safety level of the hazardous chemicals production process.

[0129] The method of the present invention can be widely used in production safety monitoring in the petroleum, chemical, coal and other industries, providing important technical support for the prevention and emergency response of major accidents such as leakage of hazardous chemicals, fire and explosion, and generating significant social and economic benefits.

[0130] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for safety monitoring of hazardous chemicals based on big data, characterized in that: The following steps are involved: Step 1: Heterogeneous sensor data collection and preprocessing: collect various sensor data in the production process, and perform spatiotemporal alignment, noise removal, missing value interpolation, and normalization preprocessing operations to form a unified spatiotemporal data matrix; Step 2. Extraction of correlation spatiotemporal features: A correlation-enhanced recurrent neural network is constructed using a bidirectional gated recursive unit to adaptively model the dynamic correlation features of multi-source time series data, and differential hidden layer states are introduced to characterize short-term multi-scale dynamics; Step 3. Dual autoencoder dictionary learning: Design a dual sparse autoencoder network with joint local-global constraints, minimize the data reconstruction error and dictionary size, and use the alternating iterative proximal gradient descent algorithm to solve the objective function; Step 4. Dictionary parameter optimization based on quantization coding: perform product quantization on the sparse dictionary, minimize the quantization coding loss by alternately optimizing the codebook and sparse coding, and use the alternating least squares algorithm to solve; Step 5. Multi-level joint decoding: At the monitoring end, the low-rank dictionary is restored through the codebook tensor product operation, and then the local reconstruction residual and global prior knowledge are combined to adaptively compensate for the quantization loss to obtain the state estimate; Step 6. Anomaly diagnosis based on enhanced reconstruction sequence: For the reconstructed state data, an anomaly diagnosis algorithm combining variational autoencoder and twin network is designed to adaptively determine the degree of anomaly of each time slice through reconstruction error and similarity score; The dual autoencoder dictionary learning step is based on the spatiotemporal feature matrix H∈R extracted by the associated enhanced recurrent neural network N ×T×K Based on this, a dual sparse autoencoder network with joint local-global constraints is designed to minimize data reconstruction error and dictionary size at the same time. The objective function is as follows: Among them, ||·|| 2,1 is the L2,1 mixing norm, D∈R K×M For an overcomplete dictionary, A∈R M×NT is the sparse coding coefficient, λ1 and λ2 are the sparse regularization parameters of samples and atoms respectively, d i ∈R K is the atomic vector of the i-th column of D; the objective function is solved by the alternating iterative proximal gradient descent algorithm, and the update formulas of A and D are: Where, f(A)=||H-DA|| 2,1 ,g(D)=||H-DA|| 2,1 +λ2||D|| 2,1 , L is the Lipschitz constant of f(A), γ is the update step size of D, j is the number of iterations, SoftThresh α (·) is the soft threshold function, and its component form is: SoftThresh α (x) i =sign(x i )max(|x i |-α,0).

2. The method according to claim 1, characterized in that The step of extracting the associated spatiotemporal features uses a bidirectional gated recursive unit to construct an associated enhanced recurrent neural network, adaptively modeling the dynamic associated features of multi-source time series data, and the forward and backward hidden layer states in the associated enhanced recurrent neural network and Calculated by: Among them, X t ∈R D is the multi-sensor input vector at time t, h t ∈R K is the hidden layer state at time t, Represents the vector concatenation operation, ⊙ represents the Hadamard product; the gating function in the gated recursive unit unit adopts the gated linear unit form: z t =σ(W z X t +U z h t-1 +b z ) r t =σ(W r X t +U r h t-1 +b r ) Among them, z t is the update gate, r t To reset the gate, is the candidate hidden layer state, σ(·) is the sigmoid activation function, and the differential hidden layer state ΔS is introduced t =h t -h t-1 To characterize short-term multi-scale dynamics.

3. The method according to claim 1, characterized in that The dictionary parameter optimization step based on quantization coding is performed on the sparse dictionary D∈R K×M Perform product quantization and express it as the Cartesian product of multiple codebooks: in, is the lth codebook, represents the Kronecker product, is the codebook dimension, n l is the number of codewords in the lth codebook; under the premise of ensuring the lower bound of the reconstruction error ∈, the codebook is optimized alternately and sparse coding Minimize the quantization coding loss, the optimization goal is: Alternating least squares (ALS) algorithm is used to solve: Fixed Then A is obtained through LASSO regression: Fixed A, then each codebook C l Update via low-rank tensor decomposition: in, The optimization objective can be solved by a high-order power method.

4. The method according to claim 1, characterized in that The multi-level joint decoding step is at the monitoring end, given a quantization codebook And sparse code A, restore the low-rank dictionary through the codebook tensor product operation: Then, the local reconstruction residual and global prior knowledge are combined to adaptively compensate for the quantization loss and obtain the state estimate: H′=D′A+ΞΘ(HD′A)+ΨΦ(D′A) Among them, H∈R N×T×K is the state prior mean, Ξ∈R K×K and Ψ∈R K×K are the local and global compensation matrices, respectively, Θ(·) and Φ(·) are the corresponding gated activation functions: ΞΘ(E)=(σ(W θ E+b θ )⊙E)W ξ ΨΦ(Z)=(σ(W φ Z+b φ )⊙Z)W ψ 。 5. The method according to claim 1, characterized in that The abnormality diagnosis step based on the enhanced reconstruction sequence is aimed at the reconstructed state data H′∈R N×T×K , design an abnormality diagnosis algorithm combining variational autoencoder and twin network; firstly, the state data is probabilistically encoded and generated through variational autoencoder: p θ (H′|z)=Bern(H′;μ θ (With)) Among them, z is a hidden variable, φ and θ are the parameters of the encoder and decoder respectively, and conditional LSTM is used to implement the mean and variance network μ φ , μ θ ; Then build a twin network to evaluate the similarity between the reconstructed sequence and the original sequence: E i =CNN(BiLSTM(H′ i :i+w)) THE s (AND i ,AND j )=||And i -AND j || 2 Among them, H′ i :i+w is the reconstructed subsequence within the i-th time window and width w, CNN(·) is the convolutional neural network feature extractor, L s (·) is the similarity measure of the twin network; finally, the reconstruction error and similarity score are combined to adaptively determine the abnormality of each time slice: s i =||H′ i -H i || 2 +λL s (AND i ,AND j ) Among them, u i ∈[0,1] is the abnormal probability of the ith time slice, and ProbAE(·) is the autoencoder probability estimation network.

6. The method according to claim 1, characterized in that In the heterogeneous sensor data collection and preprocessing step, the heterogeneous sensor data includes temperature, pressure, flow, liquid level, component process parameters, and video and image environmental perception data; Perform spatiotemporal alignment, noise removal, missing value interpolation, and normalization preprocessing operations on the collected heterogeneous data to obtain a unified spatiotemporal data matrix X∈R N×T×D , where N is the number of batch samples, T is the time step, and D is the sensor dimension.

7. The method according to claim 6, characterized in that The pressure is measured by a pressure transmitter, the measuring range of which is 0-10MPa, and the accuracy is better than 0.075%; the flow rate is measured by a flow meter or an electromagnetic flow meter, the measuring range of which is 0-5m / s, and the accuracy is better than 0.5%; the nominal diameter of the electromagnetic flow meter is DN80-DN300, and the accuracy is better than 0.3%.

8. The method according to claim 7, characterized in that The video is captured using a high-definition network camera with a pixel count of no less than 4 million, a frame rate of no less than 30fps, and an encoding format of H.

265.

9. The method according to any one of claims 1 to 8, characterized in that A hazardous chemical safety monitoring device based on big data and an edge computing gateway that is communicatively connected to the device are used. The CPU main frequency of the edge computing gateway is not less than 1.6 GHz, the memory is not less than 8 GB, the storage capacity is not less than 128 GB, and the communication method supports 4G LTE or 5G LTE.

Images