Process monitoring method for abnormal operation state in safe water treatment process

Abstract

The invention discloses a process monitoring method for an abnormal operation state in a safe water treatment process. The method comprises the following steps: firstly, segmenting original data from the dimension of time by adopting a sequential division method, and standardizing the interiors of blocks; then, a multi-scale feature extraction network is adopted to extract restrictive multi-scale spatio-temporal features from spatial dimensions by adding network function independence and information diversity constraints; and a deep stationary subspace network is established, and decoupling of stationary and non-stationary features is realized by applying difference constraints to network output features. And finally, constructing a process state change statistic by using the non-stationary features, performing data reconstruction by using the stationary features reflecting the process equilibrium state, constructing a statistic of a residual space, and completing process monitoring of the non-stationary process. According to the method, non-stationary faults influencing process safety and stationary faults caused by process mechanism abnormity can be effectively identified, and the method has high practical application value and effectiveness.

Description

Technical Field

[0001] The present invention belongs to the cross application of environmental engineering and process control, and relates to a process monitoring method for abnormal operation status in a safe water treatment process (hereinafter referred to as SWaT, i.e. Secure Water Treatment). Background Art

[0002] SWaT is a laboratory-grade safe water treatment test platform, such as Figure 1 As shown in the figure, the treatment process of the platform is based on ultrafiltration filter (UF) and reverse osmosis (RO) technology, including ultraviolet dechlorination device, ultrafiltration device, chemical dosing station, reverse osmosis device, programmable logic controller (PLC) control cabinet and water tank and other equipment. The sensor components include liquid level indicator and transmitter (LIT), pump (P), attribute indicator and transmitter (AIT), and differential pressure indicator and transmitter (DPIT), which are exactly the same as the real factory water treatment process.

[0003] like Figure 2 As shown, SWaT contains six stages of a real industrial water treatment process, labeled P1 to P6. Each stage is controlled by a set of dual PLCs, one as the master and the other as a backup in case the master PLC fails. Overall, the testbed uses a distributed control approach in normal operation, where each process stage is controlled individually by a local PLC. In the entire production process of SWaT, the first stage (P1) controls the way water flows in by opening or closing the valve connected to the water inlet pipe. Water from the city water pipe is pumped into the raw water tank after chemical dosing, i.e. chlorination treatment, in the second stage (P2). In the third stage (P3), the UF feed pump pumps water from the raw water tank into the UF feed tank. In the fourth stage (P4), the RO feed pump sends water to the UV dechlorination unit through a unit controlled by the PLC. The purpose of this step is to remove any free chlorine before the water passes through the reverse osmosis unit. In the fifth stage (P5), the dechlorinated water passes through a three-stage RO filtration unit. The filtered water from the RO unit is stored in the permeate tank and the wastewater is stored in the UF backwash tank. The sixth stage (P6) is responsible for the membrane cleaning of the UF unit. The backwash cycle is automatically initiated every 30 minutes and takes less than one minute to complete. A differential pressure sensor is installed in the third stage (P3) to measure the pressure drop in the UF unit. If the pressure drop exceeds 0.4 Pa, indicating that the membrane needs to be cleaned immediately, the backwash cycle will also be initiated.

[0004] The entire SWaT test platform completely simulates the real industrial water treatment process. In the entire production process, due to environmental disturbances or abnormalities of the equipment itself, some devices or final products will produce abnormalities. With the increasing complexity of industrial processes, it is extremely important to use process monitoring methods to ensure process safety and production efficiency. In actual industrial processes, in addition to linearity and nonlinearity, stability and non-stationarity are also important characteristics of industrial processes. Among them, stability represents the equilibrium state of the process, represents the time-invariant characteristics of process data, and is reflected in the statistical characteristics that the first-order moment and the second-order moment remain unchanged, while non-stationarity represents the changing state of the process, and its first-order moment and second-order moment change over time. Non-stationarity is very common in industrial processes, and there are many reasons for non-stationarity, such as changes in operating conditions, aging of equipment, and the influence of unknown disturbances. In actual industrial processes, stability and non-stationarity are usually mixed in a complex and unknown form, so how to conduct effective process monitoring in this case is a work worth studying.

[0005] So far, process monitoring methods have been widely studied. Traditional process monitoring methods based on multivariate statistics are a classic method and are well used for process monitoring. Principal Component Analysis (PCA), Independent Component Analysis (ICA), Canonical Correlation Analysis (CCA) and their variants have proven their effectiveness in a variety of complex tasks and form the basis of traditional process monitoring technology. With the development of deep learning technology, network structures such as AutoEncoder (AE), Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are constantly being innovated. Thanks to their powerful feature extraction and nonlinear fitting capabilities, deep learning technology has been widely used in the field of process monitoring. However, the premise of these methods is that the process is under stable conditions and cannot effectively process non-stationary information. However, in actual industrial processes, non-stationarity is a typical state, so these methods may have model mismatch problems, making the monitoring results unreliable.

[0006] Based on the above limitations on assumptions when monitoring processes, stationary subspace analysis (SSA) has begun to attract the attention of some researchers and has gradually been used sporadically in fault monitoring of industrial processes. SSA does not require strict assumptions like previous algorithms. It is based on weak stationarity without time structure and decomposes the original data into a linear combination of stationary signals and non-stationary signals in the form of subspace decomposition. Although SSA has achieved good results in the monitoring of the above industrial processes, SSA and its improved algorithms still have high computational complexity and unstable results. Moreover, with the development of deep learning, SSA, as an offline feature extraction method, cannot be integrated into neural networks to achieve end-to-end online model training. Summary of the invention

[0007] In order to solve the problems existing in the above methods, the present invention discloses a deep stationary subspace network algorithm based on multi-scale spatiotemporal feature enhancement to monitor the non-stationary faults that affect process safety and the stable faults of abnormal process mechanisms in the secure water treatment process (SWaT) of the non-stationary process. This method can not only fully mine the spatiotemporal feature information in the data, but also decouple the stable features and non-stationary features in the process data, making the monitoring results more accurate and ensuring the safe operation of the platform and the stability of product quality.

[0008] The technical solution of the present invention is specifically described as follows.

[0009] The present invention provides a process monitoring method for abnormal operation status in a safe water treatment process, comprising the following steps:

[0010] 1) Given the original data X of a non-stationary process, divide the data into N consecutive time segments X i , and then perform internal standardization within the block to retain the internal time dependency of the data while eliminating the impact of the dimension;

[0011] 2) Use a multi-scale feature extraction network to extract spatiotemporal features from the original data, add network function independence and information diversity constraints to the extracted spatiotemporal features, remove the influence of similarity in the input data space on the functional similarity within the network, and improve the diversity of multi-scale spatiotemporal features;

[0012] 3) Establish a deep stationary subspace network and decouple stationary and non-stationary features by using first-order and second-order moments to impose differential constraints on the output features of the feature extraction network;

[0013] 4) Use non-stationary features to construct process state change statistics, use stationary features that reflect the process equilibrium state to reconstruct data, construct statistics in the residual space, and complete process monitoring of non-stationary processes.

[0014] In the present invention, in step 2), a convolutional neural network with different convolution kernel sizes in the convolution layer is used in combination with a self-attention mechanism and a fully connected network to respectively form three feature extraction networks of different scales, and feature extraction is performed from the spatial dimension; specifically as follows:

[0015] The convolutional neural network extracts data from the same time segment into feature representations of different scales; before the final linear layer mapping, a layer of self-attention mechanism is added to adjust the degree of attention of the feature representation to the two dimensions of time and space; then it is input into a fully connected network to complete the construction of the entire feature extraction network.

[0016] In the present invention, in step 2), dCKA is used to constrain the similarity of network functions of feature extraction networks of three different scales, and the variance of dCKA of the constrained network is minimized to ensure the independence of network functions; combined with the mutual information method, the information difference between features is maximized by constraining the mean and variance of mutual information between each pair to minimize, thereby achieving feature diversity.

[0017] In the present invention, in step 3), the deep stationary subspace network is divided into three groups of network subunits connected in parallel, each group of network subunits includes a non-stationary decoupling subnetwork and a stationary decoupling subnetwork, each non-stationary decoupling subnetwork includes an autoencoder designed using the residual idea, the autoencoder is divided into an encoder and a decoder, and the stationary decoupling subnetwork is a fully connected network that acts as an encoder; each group of network subunits is a series structure composed of an autoencoder and a fully connected network, respectively corresponding to the output of the feature extraction network;

[0018] The multi-scale spatiotemporal features are first passed through an autoencoder in parallel.

[0019]

[0020] Where b represents the sequence number of the feature extraction network, b = 1, 2 or 3, are the nonlinear activation functions of the encoder and decoder of the non-stationary decoupled subnetwork, are the weight matrices of the encoder and decoder of the non-stationary decoupled sub-network, They are the biases of the encoder and decoder of the non-stationary decoupled subnetwork respectively;

[0021] The feature representation obtained by outputting the feature of the b-th feature extraction network at the output of the autoencoder In the process, the maximum stationarity is used as a constraint, so that the network can extract the non-stationary components in the data as much as possible; at this time, the data reconstructed by the decoder through the hidden layer features that maximize the stationarity is contains non-stationary information, borrowing the design idea of ​​skip connection of residual network, and using the bth feature to extract the output feature of the network With the decoder output Subtract, which means that the non-stationary components contained in the data are removed from the data, and the feature representation output by the autoencoder is the non-stationary feature representing the state of process change;

[0022] Furthermore, the multi-scale spatiotemporal features with non-stationary components removed are connected to a fully connected network respectively.

[0023]

[0024] in, is the nonlinear activation function of the smoothly decoupled subnetwork, and are the weight matrix and bias of the stable decoupled subnetwork, Indicates that the skip connection of the residual network is used to filter non-stationary information; in order to ensure that the output fully represents the stationary components in the multi-scale spatiotemporal features, the stationarity of the output of the fully connected network is minimized at this time, so that the feature representation based on the output features of the b-th feature extraction network A stationary characteristic that represents the equilibrium state of a process.

[0025] In the present invention, for non-stationary features, the way to maximize stationarity is to ensure the maximization of KL divergence, which is based on the minimization of negative exponents. The loss function is as follows:

[0026]

[0027] For stationary features, the stationarity of the output of the fully connected network is minimized by minimizing the KL divergence, based on the distribution assumption of minimum prior information. The loss function is as follows:

[0028]

[0029] in: and The time segments X i The mean and covariance matrix of is a multivariate normal distribution with mean 0 and variance the identity matrix.

[0030] In the present invention, in step 4), the non-stationary characteristics are used to construct process state change statistics:

[0031]

[0032] Among them, F N Indicates the non-stationary characteristics of the test data, mean(F TN ) represents the mean of the non-stationary features of the training data, and T represents the transpose.

[0033] In the present invention, in step 4), the step of constructing the statistics of the residual space includes:

[0034] ① Concatenate the stable features that reflect the equilibrium state of the process and then connect them to a fully connected network to reconstruct the input data:

[0035] Z=W T F S +B(10)

[0036] Among them, Z represents the data obtained by decoding and reconstruction, F S represents the stable features obtained by decoupling and splicing the output features of all different scale feature extraction networks, W T and B represent the weight matrix and bias of the fully connected network;

[0037] ②Construct statistics in residual space

[0038] SPE=[ZX] T [ZX](12)

[0039] T stands for transpose.

[0040] In the present invention, in step ①, the loss function of the residual space between the constrained reconstructed data and the training data is expressed as follows:

[0041]

[0042] Compared with the prior art, the present invention has the following technical effects:

[0043] (1) Modeling is done entirely in a data-driven way, without the need to establish an accurate mathematical model, and is highly adaptable and operable;

[0044] (2) For non-stationary processes, data division and feature extraction are completed from the two dimensions of time and space, ensuring the temporal dependency and spatial feature information between data;

[0045] (3) The dCKA technology is used to effectively weaken the impact of input data on the inter-layer similarity constraints during network training, ensuring the functional independence of each layer of the network. Combined with the mutual information method, the mean and variance of the network output features are constrained to maximize the information difference between features, thereby achieving a high diversity of features.

[0046] (4) A deep stationary subspace network is designed, which uses the first-order and second-order moments to differentially constrain the network output features, thereby achieving accurate decoupling of the stationarity representing the process equilibrium state and the non-stationarity representing the process change state. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a real picture of the SWaT test platform.

[0048] Figure 2 This is the six-stage production flow chart of the SWaT platform.

[0049] Figure 3 It is the model process monitoring flow chart.

[0050] Figure 4 are the training and testing error plots.

[0051] Figure 5 is a graph of the learning rate curve.

[0052] Figure 6 This is the monitoring result diagram of fault 1.

[0053] Figure 7 This is the monitoring result diagram of fault 2.

[0054] Figure 8 This is the monitoring result diagram of fault 3. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0056] The present invention provides a fault detection method for non-stationary processes based on deep learning, which is mainly divided into four parts:

[0057] The first part preprocesses the raw data from a time perspective

[0058] By sequentially partitioning and standardizing the blocks, the time-dependent properties of the data are preserved.

[0059] The second part extracts features from the scale of spatiotemporal fusion

[0060] The multi-scale feature extraction network is used to extract the spatiotemporal features in the original data, and the network function independence and information diversity constraints are added to the extracted spatiotemporal features to remove the influence of the similarity in the input data space on the functional similarity within the network, thereby improving the diversity of the multi-scale spatiotemporal features.

[0061] The third part decouples features from the perspective of stability and non-stationarity

[0062] In order to decouple the stationary and non-stationary information in multi-scale spatiotemporal features, the SSA idea is used to establish a deep stationary subspace network, and the difference constraint is used to achieve the decoupling of stationary and non-stationary features.

[0063] The fourth part completes the process monitoring of the SWaT platform through unified model training

[0064] By adding a global decoder and a global loss function, all parts are unified into one network, thus achieving effective monitoring of non-stationary processes.

[0065] Specific embodiments of the present invention are as follows Figure 3 As shown, the details are as follows.

[0066] (1) Time-dependent division and standardization of raw data

[0067] First, we consider processing from the time dimension when preprocessing data. Because we know from the characteristics of non-stationary processes that the time series characteristics in the observed data may contain very important state information for non-stationary process monitoring, we use sequential partitioning instead of random partitioning when preprocessing data. That is, given the original data X of the non-stationary process, we divide the data into N consecutive time segments X. i , in order to preserve the temporal characteristics of the data, and also meet the subsequent deep stationary subspace network for stationary and non-stationary decoupling of data partitioning requirements. Then, the data standardization work is carried out in each time segment obtained by the division.

[0068] (2) Spatiotemporal feature extraction and constraints

[0069] Then, in order to ensure the diversity of features and to represent the data from another dimension, we used a convolutional neural network combined with a self-attention mechanism and a fully connected network to form a feature extraction network to extract features from the spatial dimension. Specifically, we used convolution kernels of three different sizes: 1×1, 1×3, and 1×5 to extract data from the same time segment into three different scales of feature representation.

[0070]

[0071] in, represents the intermediate sub-output of the b-th feature extraction network, represents the convolution operation of the lth sub-feature extraction unit, σ represents the Leaky ReLU activation function. In order to prevent overfitting, a dropout layer is added after each convolution layer, and the ratio is set to 0.5.

[0072] Before the final linear layer mapping, a self-attention mechanism is added to adjust the degree of attention of feature representation to the two dimensions of time and space. Will be respectively Multiply them and we get Q b , K b , V b Three matrices; then they are input into a fully connected network to complete the construction of the entire feature extraction network.

[0073]

[0074] in Represents the final output features of each feature extraction network, and Represents the weights and biases of the fully connected network. d K Denotes the matrix K b feature dimension.

[0075] In order to eliminate the influence of data structure similarity in the same time segment on the final feature representation, dCKA is used to constrain the network function similarity of the three feature extraction networks.

[0076]

[0077] Among them, HSIC stands for Hilbert-Schmidt Independence Criterion (HSIC), Represents the feature representation of network functional similarity after removing the spatial similarity of the input data. Represents confounding factors.

[0078] Furthermore, in order to ensure the diversity of information in the feature representations learned by the three feature extraction networks, mutual information constraints are further imposed on the feature representations of the three networks, thereby ensuring that the final features contain sufficient information.

[0079]

[0080] The final feature representation not only contains the temporal and spatial information in the original data, but also ensures the functionality and diversity of the feature representation.

[0081] (3) Decoupling of spatiotemporal features

[0082] In order to realize process monitoring of non-stationary industrial processes, it is necessary to effectively separate the stationary information and non-stationary information in the non-stationary process. Therefore, a deep stationary subspace network is specially designed, in which the core is a series structure consisting of 3 groups of parallel autoencoders and a fully connected network.

[0083] The deep stable sub-network consists of three groups of network sub-units, each of which corresponds to the output of the previous feature extraction network. Each group includes a non-stationary decoupling sub-network and a stable decoupling sub-network. Each non-stationary decoupling sub-network contains an autoencoder designed with the residual idea (divided into encoder and decoder). The stable decoupling sub-network is a fully connected network that acts as an encoder.

[0084] The deep stationary subspace network adopts the design concept of SSA. The objective function of the algorithm is to minimize the distance between the first two moments of each epoch and their respective means. Here, this distance is measured using the KL divergence between Gaussian distributions. Because according to the maximum entropy principle, this is the assumption with the least restriction on the distribution. By standardizing the data, the mean and covariance of the data are made to be zero mean and unit covariance. In this way, by constraining the scale of the data, the number of parameters in the optimization problem can be reduced to avoid falling into the local minimum problem.

[0085]

[0086] in and is the time segment X i The mean and covariance matrix of is a multivariate normal distribution with mean 0 and covariance as the unit matrix, and I represents the unit matrix;

[0087] The obtained three sets of multi-scale spatiotemporal features with functional and diversity constraints will contain mixed stationary and non-stationary components. Therefore, the three sets of multi-scale spatiotemporal features will first be passed through an autoencoder in parallel.

[0088]

[0089] in, and are the nonlinear activation functions of the encoder and decoder of the non-stationary decoupled subnetwork, respectively. and as well as and are the weight matrix and the bias respectively. The feature representation obtained at the output of the encoding layer of the autoencoder In this case, we will use the method of maximizing stationarity to constrain, so that the network can extract the non-stationary components in the data as much as possible. At this time, the hidden layer (i.e., encoder) features that maximize stationarity continue to be reconstructed by the decoder. In the data, as much non-stationary information as possible will be included. Here, we borrow the design idea of ​​the residual network and subtract the original input data from the output of the decoder. This means that the non-stationary components contained in the data are removed from the data, and the feature representation of the autoencoder output is the non-stationary feature representing the state of process change.

[0090] Furthermore, the three sets of multi-scale spatiotemporal features with non-stationary components removed are then connected to a fully connected network (which can also be considered as the encoder of the autoencoder).

[0091]

[0092] in, is the nonlinear activation function of the smoothly decoupled subnetwork, and are the weight matrix and bias respectively, In order to ensure that the output fully represents the stationary components in the multi-scale spatiotemporal features, the stationarity of the output of the fully connected network is minimized, so that the feature representation It is a stable feature that fully represents the equilibrium state of the process. At this time, the three sets of stable features will be concat spliced. It is then connected to a fully connected network (which can be considered as the decoder of the overall network for global data reconstruction) to reconstruct the input data.

[0093] Z=W T F S +B(10)

[0094] Among them, Z represents the final network reconstruction output, F S represents the stable features obtained by concatenating three feature extraction networks, W T and B represent the weight matrix and bias of the fully connected network.

[0095] (4) Non-stationary process monitoring

[0096] Through the multi-scale spatiotemporal feature-enhanced deep stationary subspace network, data partitioning, feature extraction and decoupling are completed. Then, the non-stationary features are used to construct the process state change statistics.

[0097]

[0098] in, represents the non-stationary characteristics of the test data, F TN It represents the non-stationary characteristics of the training data. In order to prevent the amplitude of abnormal data from affecting the calculation of statistics, the test data is uniformly subtracted from the mean of the non-stationary characteristics of the training data.

[0099] The test data and training data are sampled from the original data in a certain ratio. The training data is divided into a training set and a validation set in a ratio of 8:2. Due to the uncertainty of the occurrence of different faults, the sampling numbers set for the test data and training data are different for different faults.

[0100] The stable characteristics of the equilibrium state of the reaction process are used for data reconstruction, and the statistics of the residual space are constructed to complete the process monitoring of the non-stationary process.

[0101] SPE=[ZX] T [ZX](12)

[0102] Where Z represents the data obtained by decoding and reconstruction, and X is the training data. The control limits of both statistics are constructed using kernel density estimation (KDE), where the confidence level is set to 97.5%.

[0103] The following are specific examples.

[0104] Example 1

[0105] SWaT is a laboratory-level test platform that truly simulates the industrial water treatment process. The data used in this paper comes from a batch of improved data collected by the platform in July 2019. The total amount of data is more than 15,000, including 78 variables measured by sensors at different stages, including 3 hours of normal data and multiple fault simulation data within 1 hour. Among them, a total of 5,000 normal data were extracted. After filtering out non-numeric and constant variables, there are still 26 variables. One variable is set by using the process variable synthesis simulation process mechanism, a total of 27 variables, which are used for model training.

[0106] Table 1 Alternative variables for training

[0107] Serial number Variable Name Serial number Variable Name Serial number Variable Name 1 LIT 101 10 FIT 301 19 FIT 501 2 AIT 201 11 LIT 301 20 FIT 502 3 AIT 202 12 AIT 402 21 FIT 503 4 AIT 203 13 FIT 401 22 FIT 504 5 FIT 201 14 LIT 401 23 PIT 501 6 AIT 301 15 AIT 501 24 PIT 502 7 AIT 302 16 AIT 502 25 PIT 503 8 AIT 303 17 AIT 503 26 FIT 601 9 DPIT 301 18 AIT 504 27 Synthetic variables

[0108] For the first stable fault, this fault indicates that a slight abnormality has occurred in the operating mechanism of the production process. In order to simulate this fault, a mixed variable was constructed using the two variables FIT502 and FIT504, with a total of 1000 data, in which a slight change in the process mechanism was introduced at point 501. The second and third faults used the faults set during the SWaT operation process, which are defined here as non-stable faults. The second fault was caused by the abnormal adjustment of the valve FIT401 from 0.8 to 0.5, which in turn caused an abnormality in the ultraviolet dechlorination device. There were 315 data in total, and the fault was introduced at point 211. The third fault was caused by the abnormal flow measured by LIT301. The flow value changed from 835 to 1024, causing T301 to underflow. There were 522 data in total, and the fault was introduced at point 217.

[0109] (1) Data Division

[0110] Given the original data X of a non-stationary process, divide the data into N consecutive time segments Where l is the window size, m is the dimension of the variable, and N is the number of time slices. Then in each time slice X i Internally, normalize the data

[0111]

[0112] Where j = 1, 2, ..., m, x j is the time segment X i The jth sampled variable, m j is the mean of the variable, s j is the corresponding variance, is the j-th variable after standardization.

[0113] (2) Feature extraction

[0114] After data segmentation, the time segments will be sent to the multi-scale feature extraction network based on CNN to realize spatiotemporal feature extraction at different spatial scales. In one training, the input data segments X corresponding to these three networks are i To ensure that the spatiotemporal features learned by the three networks are real spatiotemporal features and are not affected by the same input time segments, dCKA is added to each network.

[0115]

[0116] in, is the output feature of the b-th feature extraction network, X i is the input time segment. In order to ensure the consistency of the three networks’ ability to learn spatial features, the variance of the dCKA of the three networks is constrained to be minimized:

[0117]

[0118] Ultimately, the structural similarity in the input data is ignored, and only the constraints on the network structural similarity are retained. By removing the structural similarity of the input data from the learned spatiotemporal features, only the network functional similarity learned by different feature extraction networks is retained, thereby achieving the learning of three sets of multi-scale features to obtain the true spatiotemporal features.

[0119] The above dCKA can only ensure that the learned spatiotemporal features remove the influence of the structural similarity of the input data. In order to ensure the diversity of features learned between different networks, more information can be learned from the input data to maximize the information content of the hidden layer features. For the output features of the three feature extraction networks, calculate the mutual information between them.

[0120]

[0121] Where b, c = 1, 2, 3 and b ≠ c are three different feature extraction networks. By constraining the mean and variance of the mutual information between the two to be minimized,

[0122]

[0123] Ensure that different feature extraction networks can learn the diversity of features at different scales to the greatest extent, and maximize the amount of extracted spatiotemporal feature information.

[0124] (3) Feature Decoupling

[0125] The multi-scale spatiotemporal features constrained by network similarity and information diversity will be further fed into the deep stationary subspace network for the decoupling of stationary and non-stationary features. In order to specifically measure the stationarity and non-stationarity of the features, the first and second order moments are used to introduce constraints on stationarity and non-stationarity. In order to limit the time segment X of the finite sample i The estimation error caused by using the empirical mean μ i and the empirical covariance ∑ i To describe the distribution of different time slices. Since it is impossible to estimate the mean and covariance of a single time point, the original time series X is divided into N time slices X1,...,X N If the mean μ1,...,μ of each time segment is N and the covariance matrix ∑1,...,∑ N The same, that is

[0126]

[0127] For all time segments 1≤i,j≤N, then the time series is said to be stationary. To simplify the calculation, the amount of calculation can be reduced by assuming that the mean and covariance matrix of each time segment are equal to the mean of all time segments.

[0128]

[0129] Where 1≤i≤N, are the mean and covariance of all time segments, respectively.

[0130] Since our goal is to minimize the mean and covariance between time periods, we use KL divergence to measure this distance, which is based on the maximum entropy distribution, which is a distribution assumption that requires minimal prior information.

[0131]

[0132] Here we do not need to assume that the data distribution is Gaussian, but only assume that the non-stationarity can be reflected by the first two moments. As for the non-stationarity, we can use the above concepts to maximize the KL divergence, which is achieved here by minimizing the negative exponent.

[0133]

[0134] (4) Global Reconstruction

[0135] The above decoupled stationary feature F S and non-stationary characteristics F N , the non-stationary features will be directly used to construct non-stationary statistics to monitor the operating status of the process, and the stationary features will be spliced ​​and then connected to a fully connected network to decode the features and reconstruct the data. Therefore, it is necessary to use

[0136]

[0137] To constrain the residual space between the reconstructed data and the training data.

[0138] Finally, the total loss function needs to integrate the above sub-loss functions to achieve unified training. The loss changes of the training set and the validation set during the training process are as follows: Figure 4 As shown in the figure, it can be seen that the loss reaches a basically stable state at around 100 epchs, indicating that the above network settings can achieve network convergence quickly. Figure 5 As shown in the figure, the network can basically complete network training and parameter tuning when the learning rate is adjusted about 5-6 times.

[0139] L total =min{αL dCKA +βL MI +γ(L SF +L NF )+L Recon} (twenty three)

[0140] Among them, α, β, and γ are the hyperparameters of the balanced loss function.

[0141] (5) Process monitoring

[0142] After the above preprocessing, the actual test data is input into the trained model for real-time non-stationary process monitoring. The monitoring results of the model for the test data are shown in Table 2. The monitoring diagrams for the three faults are shown in Figure 6-Figure 8 It can be seen from the monitoring results that the method proposed by the present invention can fully realize the effective monitoring of different types of faults in the safe water treatment process.

[0143] Table 2 Summary of monitoring results

[0144]

[0145] As mentioned above, the present invention proposes a non-stationary process monitoring method for deep stationary subspace differential decoupling with restrictive constraints on multi-scale spatiotemporal features. First, the original data is segmented from the time dimension using a sequential division method, and the internal block is standardized to retain the internal time dependency of the data while eliminating the influence of the dimension; then a multi-scale feature extraction network is used from the spatial dimension, and network functional independence and information diversity constraints are added to extract restrictive multi-scale spatiotemporal features. Then a deep stationary subspace network is established, and the decoupling of stationary and non-stationary features is achieved by imposing differential constraints on the network output features. Finally, when the real-time process data is fed into the trained model, this method can effectively identify non-stationary faults that affect process safety and stationary faults caused by abnormal process mechanisms, and has high practical application value and effectiveness.

Claims

1. A process monitoring method for abnormal operation status in a safe water treatment process, characterized in that: The following steps are involved: 1) Given the original data X of a non-stationary process, divide the data into N consecutive time segments X i , and then perform internal standardization within the block to retain the internal time dependency of the data while eliminating the impact of the dimension; 2) Use a multi-scale feature extraction network to extract spatiotemporal features from the original data, add network function independence and information diversity constraints to the extracted spatiotemporal features, remove the influence of similarity in the input data space on the functional similarity within the network, and improve the diversity of multi-scale spatiotemporal features; 3) Establish a deep stationary subspace network and decouple stationary and non-stationary features by using first-order and second-order moments to impose differential constraints on the output features of the feature extraction network; 4) Use non-stationary features to construct process state change statistics, use stationary features that reflect the process equilibrium state to reconstruct data, construct statistics in the residual space, and complete process monitoring of non-stationary processes.

2. The process monitoring method according to claim 1, characterized in that: In step 2), a convolutional neural network with different convolution kernel sizes in the convolution layer is used in combination with a self-attention mechanism and a fully connected network to form three feature extraction networks of different scales, and feature extraction is performed from the spatial dimension; specifically, as follows: The convolutional neural network extracts data from the same time segment into feature representations of different scales; before the final linear layer mapping, a layer of self-attention mechanism is added to adjust the degree of attention of the feature representation to the two dimensions of time and space; then it is input into a fully connected network to complete the construction of the entire feature extraction network.

3. The process monitoring method according to claim 1, characterized in that: In step 2), dCKA is used to constrain the similarity of network functions of feature extraction networks of three different scales, and the variance of dCKA of the constrained network is minimized to ensure the independence of network functions; combined with the mutual information method, the information difference between features is maximized by minimizing the mean and variance of the mutual information between each pair, thereby achieving feature diversity.

4. The process monitoring method according to claim 1, characterized in that: In step 3), the deep stationary subspace network is divided into three groups of network subunits connected in parallel, each group of network subunits includes a non-stationary decoupling subnetwork and a stationary decoupling subnetwork, each non-stationary decoupling subnetwork contains an autoencoder designed with a residual idea, the autoencoder is divided into an encoder and a decoder, and the stationary decoupling subnetwork is a fully connected network that acts as an encoder; each group of network subunits is a series structure composed of an autoencoder and a fully connected network, corresponding to the output of the feature extraction network respectively; The multi-scale spatiotemporal features are first passed through an autoencoder in parallel. Where b represents the sequence number of the feature extraction network, b = 1, 2 or 3, are the nonlinear activation functions of the encoder and decoder of the non-stationary decoupled subnetwork, are the weight matrices of the encoder and decoder of the non-stationary decoupled sub-network, They are the biases of the encoder and decoder of the non-stationary decoupled subnetwork respectively; The feature representation obtained by outputting the feature of the b-th feature extraction network at the output of the autoencoder In the process, the maximum stationarity is used as a constraint, so that the network can extract the non-stationary components in the data as much as possible; at this time, the data reconstructed by the decoder through the hidden layer features that maximize the stationarity is contains non-stationary information, borrowing the design idea of ​​skip connection of residual network, and using the bth feature to extract the output feature of the network With the decoder output Subtract, which means that the non-stationary components contained in the data are removed from the data, and the feature representation output by the autoencoder is the non-stationary feature representing the state of process change; Furthermore, the multi-scale spatiotemporal features with non-stationary components removed are connected to a fully connected network respectively. in, is the nonlinear activation function of the smoothly decoupled subnetwork, and are the weight matrix and bias of the stable decoupled subnetwork, Indicates that the skip connection of the residual network is used to filter non-stationary information; in order to ensure that the output fully represents the stationary components in the multi-scale spatiotemporal features, the stationarity of the output of the fully connected network is minimized at this time, so that the feature representation based on the output features of the b-th feature extraction network A stationary characteristic that represents the equilibrium state of a process.

5. The process monitoring method according to claim 4, characterized in that: For non-stationary features, the way to maximize stationarity is to ensure the maximization of KL divergence, which is based on the minimization of negative exponents. The loss function is as follows: For stationary features, the stationarity of the output of the fully connected network is minimized by minimizing the KL divergence, based on the distribution assumption of minimum prior information. The loss function is as follows: in: and The time segments X i The mean and covariance matrix of is a multivariate normal distribution with mean 0 and variance the identity matrix.

6. The process monitoring method according to claim 1, characterized in that: In step 4), the non-stationary characteristics are used to construct process state change statistics: Among them, F N Indicates the non-stationary characteristics of the test data, mean(F TN ) represents the mean of the non-stationary features of the training data, and T represents the transpose.

7. The process monitoring method according to claim 1, characterized in that: In step 4), the step of constructing the statistics of the residual space includes: ① Concatenate the stable features that reflect the equilibrium state of the process and then connect them to a fully connected network to reconstruct the input data: Z=W T F S +B(10) Among them, Z represents the data obtained by decoding and reconstruction, F S represents the stable features obtained by decoupling and splicing the output features of all different scale feature extraction networks, W T and B represent the weight matrix and bias of the fully connected network; ②Construct statistics in residual space SPE=[Z-X] T [Z-X](12) T stands for transpose.

8. The process monitoring method according to claim 7, characterized in that: In step ①, the loss function of the residual space between the constrained reconstructed data and the training data is expressed as follows:

Images