Multi-modal industrial process monitoring method based on energy model

Abstract

The invention discloses a multi-modal industrial process monitoring method based on an energy model. For multi-modal data collected by an industrial sensor, an energy model is adopted to carry out parametric modeling on probability density distribution of the multi-modal data, and modals are identified. The non-normalized density is estimated by combining an energy function of a full connection layer and residual connection, and the non-normalized density is converted into a probability density function by using a partition function. A model target is composed of a maximum likelihood function of modal classification and a data reconstruction error function, and Langevin dynamics is adopted in the training process. And during monitoring, hidden layer features are extracted to establish secondary statistics, a control limit is set by using kernel density estimation, a sample mode is identified online, whether the statistics exceed the limit or not is judged, and if the statistics exceed the limit, a fault occurs. Meanwhile, a fault isolation method based on an energy model is provided, and a fault variable contribution graph is generated through a difference value between reconstructed data and real data to recognize fault variables. The method is suitable for multi-mode industrial monitoring, guarantees the safety of industrial production, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to a data-driven process monitoring method for process industry, and in particular to a multi-modal industrial process monitoring method based on an energy model. Background Art

[0002] With the development of intelligent manufacturing, the safety of industrial production has become particularly critical. Modern industrial processes are becoming increasingly complex and expanding in scale, showing multi-condition and multi-modal characteristics. In this case, the probability of failures in the production process has increased significantly. If these failures cannot be effectively detected and controlled, it may lead to a decline in product quality, an increase in production costs, and even cause serious problems such as equipment damage.

[0003] Data-driven single-modal process monitoring methods mainly consist of machine learning methods and deep learning methods. Machine learning methods include methods based on principal component analysis (PCA), methods based on linear discriminant analysis (LDA), and methods based on partial least squares (PLS). These methods can monitor process data by calculating the statistical characteristics of process data. Deep learning methods include methods based on convolutional neural networks (CNN), methods based on recurrent neural networks (RNN), and methods based on autoencoders (AE). These deep learning methods perform process monitoring by extracting hidden information from data.

[0004] However, these single-modal-based methods have difficulty processing multimodal process data. When data from different modalities are input into the model, omissions or false alarms will occur. In order to solve this problem, a multimodal process monitoring method is proposed. Most of these multimodal process monitoring methods model the data of each modality separately and combine these models, such as a multimodal principal component analysis process monitoring method and a multimodal autoencoder process monitoring method. They model based on the statistical characteristics or hidden characteristics of multimodal data, ignoring the distribution characteristics of the multimodal data itself, so there may be problems such as missing key data information. Summary of the invention

[0005] In order to solve the technical problems existing in the background technology, the present invention provides a multimodal industrial process monitoring method based on an energy model. The probability distribution of multimodal industrial data is parameterized and modeled by an energy model, and used for online industrial process monitoring. The present invention can classify the mode to which the sample data belongs, and use the constructed quadratic statistics to determine whether the process data is fault data. When a fault is detected, the fault variable can be determined based on the difference between the data reconstructed by the energy model and the real data. This method is applicable to the problems in the field of multimodal industrial process monitoring, and is of great significance in improving the fault detection accuracy of multimodal industrial process monitoring. The present invention provides reliable and effective technical support for industrial process monitoring.

[0006] The technical solution adopted by the present invention is:

[0007] Step 1, collecting industrial process data of different modes under normal working conditions as a training set, and collecting industrial process data under the working conditions to be tested as a test set; in step 1, the specific steps are:

[0008] The normal data collected by the sensor is used as the training set, denoted as X∈R n×d , the data to be tested collected by the sensor is taken as the test set, denoted as Where n is the total number of samples in the training set, m is the total number of samples in the test set, and d is the sample dimension.

[0009] Step 2, input the training set data and the marked mode into the energy model for training; in step 2, the specific steps are:

[0010] Step 2.1: label the modality of the training data collected in step 1, X = {X (1) , X (2) , ...X (k) , ..., X (K)}, Where K represents the total number of K modes in the training data, k represents the index of the mode, and n k Represents the total number of samples of the kth mode.

[0011] Step 2.2: Based on the training data collected in step 1 and the corresponding modal design energy function f obtained in step 2.1, θ (x), f θ (x) is constructed as a fully connected neural network with residual connections, θ is the parameter of the neural network, x∈R d×1 is a sample in the training data X. The energy function f θ (x) is an input data sample, f θ The hidden layer output of (x) is h, which is the modal category of the current sample And the reconstruction sample of the data

[0012] Step 2.3, based on the energy function f constructed in step 2.2 θ (x) Establish energy model p θ (x) = exp(-f θ (x)) / Z, where Z = ∫exp(-f θ (x)) is the partition function, so that the final probability density p θ (x) is between 0 and 1.

[0013] Step 2.4, design the loss function according to the energy model established in step 2.3. The loss function consists of two parts, one is the cross entropy loss function, which minimizes the cross entropy between the mode output by the model and the normal mode, and the other is the mean square error, which minimizes the mean square error between the reconstructed sample output by the model and the real sample.

[0014]

[0015] Where l is the model loss function, CE is the cross entropy loss function, is the mean square error.

[0016] Step 2.5: Continuously train the model based on the training data to minimize the loss function in step 2.4.

[0017] Step 3, calculate the quadratic statistics under each mode in the training data, and use kernel density estimation to calculate the control limits of each quadratic statistic. In step 3, the specific steps are:

[0018] Step 3.1, construct the quadratic statistic D of the latent space, the quadratic statistic of the latent space is f θ The hidden layer output of (x) is the inner product of h, D = h T h.

[0019] Step 3.2, construct the quadratic statistic R of the residual space. The quadratic statistic of the residual space is the inner product of the difference between the reconstructed value output by the energy model and the true value.

[0020] Step 3.3, use kernel density estimation to estimate the distribution of the two quadratic statistics in steps 3.1 and 3.2, and set the 99% quantile as the control limit of the quadratic statistics under each mode, and get and

[0021] Step 4: Perform online process monitoring on the trained energy model. When new data input is collected and input into the energy model, the energy model can output the mode to which the data belongs, the quadratic statistics of the data, and the reconstructed data. If the quadratic statistics of the data exceed the control limit of the statistics, it means that the model has detected a fault. In step 4, the specific steps are:

[0022] Step 4.1: Apply the energy model trained in step 3 to online process detection. After being collected, it is input into the energy model to obtain the mode to which the data belongs. And the reconstructed data Calculate the quadratic statistic for the data and

[0023] Step 4.2: Determine the quadratic statistics of the data and Whether it exceeds the control limit of the corresponding mode, if so, or If yes, it means that the energy model has detected a fault, otherwise it means that no fault has been detected.

[0024] Step 5: When the model detects a fault, the difference between the reconstructed data and the real data sample is calculated to obtain the fault contribution graph, thereby identifying the fault variable.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. Using energy model-based methods to monitor multimodal industrial process data, we have achieved innovation in process monitoring;

[0027] 2. Model the probability distribution of multimodal data through energy models to achieve multimodal modeling of a single model;

[0028] 3. The quadratic statistics constructed through the energy model can accurately identify faults in multimodal industrial process data and ensure the safety of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The modal identification result involved in the present invention;

[0030] Figure 2 The fault detection result involved in the present invention;

[0031] Figure 3 The data reconstruction result involved in the present invention;

[0032] Figure 4 The fault variable contribution diagram results involved in the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0034] This implementation case adopts the embodiment of the present invention and its implementation process is as follows:

[0035] Taking the over-temperature failure of the water-cooled wall pipe of a coal-fired power plant as an example, the multimodal industrial process monitoring method based on the energy model is described in detail.

[0036] The water wall is a key component in the boiler system. It is composed of neatly arranged metal tubes and fits closely inside the furnace. Its main function is to absorb the radiant heat generated by the high-temperature flame or flue gas in the furnace, generate steam or hot water in the tube, thereby reducing the temperature of the furnace wall and protecting the furnace wall. In order to ensure the stable operation of coal-fired power plants, the temperature of the water-cooled wall tubes is usually set within a certain temperature threshold. In this example, the limit temperature is set to 450°C. If the temperature of the water-cooled wall tube exceeds this threshold for a long time, the metal pipe may be burned. Therefore, keeping the pipe temperature below the limit temperature is an important measure to ensure safety. The key is to monitor the variables that affect the temperature of the water-cooled wall tubes. This process will be described in detail below.

[0037] Table 1 Variables and descriptions of boiler water wall overtemperature fault process

[0038]

[0039] In this process, a total of 8 variables were considered and listed in Table 1. Data samples were collected every 15 seconds for each variable.

[0040] Step 1: Collect industrial process data of different modes under normal working conditions as training sets, and collect 3606 samples under normal working conditions in total; collect industrial process data under the working conditions to be tested as test sets, and collect 1003 samples to be tested in total;

[0041] Step 2: Input the training set data and the marked modes into the energy model for training. There are three modes for normal samples, of which 1 to 900 is mode 1, 901-1806 is mode 2, and 1807-3606 is mode 3.

[0042] Step 3, calculate the quadratic statistics under each mode in the training data, and use kernel density estimation to calculate the control limits of each quadratic statistic;

[0043] Step 4: Perform online process monitoring on the trained energy model. When new data input is collected and input into the energy model, the energy model can output the mode to which the data belongs, the quadratic statistics of the data, and the reconstructed data. If the quadratic statistics of the data exceed the control limit of the statistics, it means that the model has detected a fault.

[0044] Figure 1 The modal recognition results of this case are shown. The modalities of all samples are correctly identified, and the modal recognition accuracy is 100%;

[0045] Figure 2 The fault detection results in this case are shown. When the water-cooled wall temperature exceeds the designed control limit, the quadratic statistic instantly exceeds the designed control limit, which means that the method proposed in the present invention can accurately identify faults in multimodal industrial data.

[0046] Figure 3 The data reconstruction results in this case are shown. Under normal data, the reconstructed data value is close to the true value. When a fault occurs, there is a significant difference between the reconstructed data value and the true value.

[0047] Step 5: When the model detects a fault, the difference between the reconstructed data and the real data sample is calculated to obtain the fault contribution graph, thereby identifying the fault variable. Figure 4 The fault variable contribution graph is shown, in which variables 4, 6, 7, and 8 have the largest contribution values, indicating that these four variables are fault variables.

[0048] In summary, the multimodal industrial process detection method based on energy model proposed in the present invention can effectively detect and identify faults in multimodal industrial data and has high application value.

Claims

1. A multimodal industrial process monitoring method based on energy model, characterized in that: The method specifically comprises the following steps: Step 1: Collect industrial process data of different modes under normal working conditions as a training set, and collect industrial process data under the working conditions to be tested as a test set; Step 2, input the training set data and the marked modes into the energy model for training; Step 3, calculate the quadratic statistics under each mode in the training data, and use kernel density estimation to calculate the control limits of each quadratic statistic; Step 4: Use the trained energy model for online process monitoring. When new data input is collected and input into the energy model, the energy model can output the mode to which the data belongs, the quadratic statistics of the data, and the reconstructed data. If the quadratic statistics of the data exceed the control limit of the statistics, it means that the model has detected a fault; Step 5: When the model detects a fault, the difference between the reconstructed data and the real data sample is calculated to obtain the fault contribution graph, thereby identifying the fault variable.

Images