A Method and System for Monitoring the Preparation Process of Ternary Cathode Materials Based on RVAE

Abstract

This invention discloses a monitoring method and system for the preparation process of ternary cathode materials based on RVAE. It constructs a nonlinear dynamic system model of the sintering process based on a variational autoencoder; assigns different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, derives the loss function of the nonlinear dynamic system model of the sintering process, and trains the model parameters through backpropagation; defines the statistics of the nonlinear dynamic system model of the sintering process based on the cyclic variational autoencoder, and obtains the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; collects online data as a test set for the nonlinear dynamic system model, calculates the monitoring statistics online and compares them with the control limits to determine whether a fault has occurred. This invention can significantly improve the fault detection rate and false alarm rate, providing a strong guarantee for the stable operation of the sintering process.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology in the preparation process of ternary cathode materials, and in particular discloses a monitoring method and system for the preparation process of ternary cathode materials based on RVAE (Recurrent neural network-Variational auto-encoder). Background Technology

[0002] The preparation of ternary cathode materials is a complex process involving the simultaneous coupling of multiple forms of mass transfer and energy exchange, as well as heat convection and diffusion, within a roller kiln. To simplify the system, the sintering process is divided into three temperature zones based on temperature trends: a heating zone, a isothermal zone, and a cooling zone. Each zone is adjusted by the addition of oxygen and temperature according to the chemical reactions occurring. Under specific temperature and oxygen conditions, the materials undergo predetermined reactions within their respective zones. However, deviations from the set sintering regime will lead to substandard product quality or energy waste. Therefore, real-time monitoring of the sintering process is crucial for adjusting the sintering regime, providing timely guidance to operators to adjust operating parameters.

[0003] However, due to the temperature convection and coupling of reactions between the front and rear temperature zones in the roller kiln, the sample at the current sampling moment is affected by the state at the previous moment. The collected process temperature data breaks the traditional static assumption and exhibits strong dynamic characteristics, that is, the data shows time-delay dynamic characteristics. At the same time, due to factors such as raw material composition fluctuations, catalyst deactivation, and silicon carbide rod aging, the relationship between the sintering regime and product quality is complex and exhibits strong nonlinear characteristics.

[0004] Therefore, the time-delay dynamic and nonlinear characteristics exhibited in the preparation process of ternary cathode materials lead to low fault detection rate and high false alarm rate of traditional monitoring methods, which is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides a monitoring method and system for the preparation process of ternary cathode materials based on RVAE, aiming to solve the technical problems of low fault detection rate and high false alarm rate of traditional monitoring methods caused by the time-delay dynamic characteristics and nonlinear characteristics exhibited in the preparation process of ternary cathode materials.

[0006] One aspect of the present invention relates to a method for monitoring the preparation process of ternary cathode materials based on RVAE, comprising the following steps:

[0007] A nonlinear dynamic system model of the sintering process based on a variational autoencoder is constructed.

[0008] An autoregressive equation between latent variables is established through a recurrent neural network. A weighted moving average method is introduced to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process. The loss function of the nonlinear dynamic system model of the sintering process is derived, and the model parameters are trained through backpropagation.

[0009] Based on the nonlinear dynamic system model based on cyclic variational autoencoder, the statistics of the nonlinear dynamic system model of sintering process are defined, and the control threshold of the nonlinear dynamic system model of sintering process is obtained by kernel density estimation.

[0010] Collect online data as a test set for the nonlinear dynamic system model and perform standardization processing;

[0011] The test set is tested based on the constructed nonlinear dynamic system model of the sintering process. The statistics of the test samples are calculated, and the calculated statistics are compared with the preset monitoring and control limits to determine whether a fault has occurred, and the detection results are output.

[0012] Furthermore, the steps for constructing a nonlinear dynamic system model of the sintering process based on a variational autoencoder include:

[0013] The collected historical data is used as the model training set, and each sample in the model training set is standardized.

[0014] Based on the nonlinear framework of variational autoencoders, a recurrent neural network is used to construct the autoregressive equations of latent variables, resulting in a nonlinear dynamic system model of the sintering process based on recurrent variational autoencoders.

[0015] Furthermore, based on the nonlinear framework of the variational autoencoder, and in conjunction with a recurrent neural network to construct the autoregressive equations of the latent variables, the mathematical structure of the obtained nonlinear dynamic system model of the sintering process based on the recurrent variational autoencoder is as follows:

[0016] z(1)=μ(1)+σ(1)×ε,ε~N(0,I)

[0017] z(t)=q(h(t-1))+B(t)μ(t)+σ(t)×ε

[0018] h(t-1)=[z(t-1) T ,z(t-2) T …,z(tT) T ]

[0019]

[0020]

[0021] Where z(1) is the latent variable at time 1, μ(1) and σ(1) are the mean and variance calculated by the inference network at time 1; z(t) is the latent variable at time t; and , respectively, are the process variable and quality variable reconstructed at time t; μ(t) and σ(t) are the mean and variance calculated by the inference network at time t; ε is the reparameter factor, which follows a standard Gaussian distribution; B(t) is the transition matrix at time t; T represents the time delay coefficient of the process; h(t-1) is the state increment containing latent variables of the past T time lengths; q(*), f(*) and g(*) are nonlinear functions approximated by the neural network.

[0022] Furthermore, an autoregressive equation between latent variables is established using a recurrent neural network, and a weighted moving average method is introduced to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process. The loss function of the final nonlinear dynamic system model is derived, and the model parameters are trained through backpropagation.

[0023] According to the forward propagation algorithm, the output of the hidden layer of the recurrent neural network at different time steps is:

[0024] h(t) = f(Uz(t) + Wh(t-1) + b)

[0025] Where U and W are weight matrices, z(t) are the latent variables calculated by the inference network at different times, h(t) are hidden layer units, h(t-1) are the hidden layer outputs at the previous time step, b is the bias term, and f(*) is the activation function.

[0026] The loss function derived for the nonlinear dynamic system model of the sintering process is:

[0027]

[0028] Where J(Φ,Θ) is the loss function, weight(t) is the weight of the sample at time t, x(t) is the process variable at time t, and μ x(t) Let y(t) be the expected mean of the process variable, and μ be the mass variable at time t. y(t) Let D be the expected mean of the quality variable, weight(0) be the initial sample weights, and D be the mean of the quality variable. KL Let z(1) be the KL divergence, z(1) be the initial hidden variable, φ(1) be the initial neural network parameters, N(0,I) be the standard Gaussian distribution, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, φ(t) be the neural network parameters at time t, I be the identity matrix, and q(*) and p(*) be the nonlinear functions approximated by the neural network.

[0029] Furthermore, in the step of detecting the test set based on the constructed nonlinear dynamic system model of the sintering process, calculating the statistics of the test samples, comparing the calculated statistics with the preset monitoring and control limits to determine whether a fault has occurred, and outputting the detection results, the calculated statistics of the test samples are as follows:

[0030]

[0031] Among them, T 2 (t) is the statistic; H(t) = [x(t), z(t)] is the input variable; The singular values ​​are obtained by performing singular value decomposition on H(t); n z The number of its principal components; for The principal space spanned by the corresponding eigenvectors; The transpose of the primary space, H T For the transpose of the input variable, It is the inverse of the inverse diagonal matrix composed of singular values;

[0032] The determination of whether a fault has occurred is based on comparing the statistical values ​​of the online process data with the control limits. The logic for this determination is as follows:

[0033]

[0034] Among them, T 2 (t) is a statistic. For control limits.

[0035] Another aspect of the present invention relates to a monitoring system for the preparation process of ternary cathode materials based on RVAE, comprising:

[0036] A building block is used to construct a nonlinear dynamic system model of the sintering process based on a variational autoencoder;

[0037] The derivation module is used to establish the autoregressive equation between latent variables through a recurrent neural network, and to introduce the weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, derive the loss function of the nonlinear dynamic system model of the sintering process, and train the model parameters through backpropagation.

[0038] The estimation module is used to define the statistics of the nonlinear dynamic system model of the sintering process based on the nonlinear dynamic system model of the cyclic variational autoencoder, and obtain the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation.

[0039] The data collection module is used to collect online data as a test set for nonlinear dynamic system models and to perform standardization processing.

[0040] The comparison module is used to detect the test set based on the constructed nonlinear dynamic system model of the sintering process, calculate the statistics of the test samples, compare the calculated statistics with the preset monitoring and control limits to determine whether a fault has occurred, and output the detection results.

[0041] Furthermore, the building blocks include:

[0042] The processing unit is used to use the collected historical data as a model training set and to standardize each sample in the model training set.

[0043] The acquisition unit is used to construct the autoregressive equation of the latent variables in conjunction with the nonlinear framework based on the variational autoencoder, thereby obtaining the nonlinear dynamic system model of the sintering process based on the cyclic variational autoencoder.

[0044] Furthermore, the mathematical structure of the nonlinear dynamic system model of the sintering process obtained in the acquisition unit is as follows:

[0045] z(1)=μ(1)+σ(1)×ε,ε~N(0,I)

[0046] z(t)=q(h(t-1))+B(t)μ(t)+σ(t)×ε

[0047] h(t-1)=[z(t-1) T ,z(t-2) T …,z(tT) T ]

[0048]

[0049]

[0050] Where z(1) is the latent variable at time 1, μ(1) and σ(1) are the mean and variance calculated by the inference network at time 1; z(t) is the latent variable at time t; and , respectively, are the process variable and quality variable reconstructed at time t; μ(t) and σ(t) are the mean and variance calculated by the inference network at time t; ε is the reparameter factor, which follows a standard Gaussian distribution; B(t) is the transition matrix at time t; T represents the time delay coefficient of the process; h(t-1) is the state increment containing latent variables of the past T time lengths; q(*), f(*) and g(*) are nonlinear functions approximated by the neural network.

[0051] Furthermore, in the derivation module, based on the forward propagation algorithm, the hidden layer outputs of the recurrent neural network at different time points are:

[0052] h(t) = f(Uz(t) + Wh(t-1) + b)

[0053] Where U and W are weight matrices, z(t) are the latent variables calculated by the inference network at different times, h(t) are hidden layer units, h(t-1) are the hidden layer outputs at the previous time step, b is the bias term, and f(*) is the activation function.

[0054] The loss function derived for the nonlinear dynamic system model of the sintering process is:

[0055]

[0056] Where J(Φ,Θ) is the loss function, weight(t) is the weight of the sample at time t, x(t) is the process variable at time t, and μ x(t) Let y(t) be the expected mean of the process variable, and μ be the mass variable at time t. y(t) Let D be the expected mean of the quality variable, weight(0) be the initial sample weights, and D be the mean of the quality variable. KL Let z(1) be the KL divergence, z(1) be the initial hidden variable, φ(1) be the initial neural network parameters, N(0,I) be the standard Gaussian distribution, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, φ(t) be the neural network parameters at time t, I be the identity matrix, and q(*) and p(*) be the nonlinear functions approximated by the neural network.

[0057] Furthermore, in the comparison module (50),

[0058] The calculated statistic for the test sample is:

[0059]

[0060] Among them, T 2 (t) is the statistic; H(t) = [x(t), z(t)] is the input variable; The singular values ​​are obtained by performing singular value decomposition on H(t); n z The number of its principal components; for The principal space spanned by the corresponding eigenvectors; The transpose of the primary space, H T For the transpose of the input variable, It is the inverse of the inverse diagonal matrix composed of singular values;

[0061] The determination of whether a fault has occurred is based on comparing the statistical values ​​of the online process data with the control limits. The logic for this determination is as follows:

[0062]

[0063] Among them, T 2 (t) is a statistic. For control limits.

[0064] The beneficial effects achieved by this invention are as follows:

[0065] This invention provides a monitoring method and system for the preparation process of ternary cathode materials based on RVAE. It constructs a nonlinear dynamic system model of the sintering process based on a variational autoencoder; establishes an autoregressive equation between latent variables using a recurrent neural network; and introduces a weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, deriving the loss function of the nonlinear dynamic system model of the sintering process, and training the model parameters through backpropagation. Based on the nonlinear dynamic system model based on the cyclic variational autoencoder, it defines the statistics of the nonlinear dynamic system model of the sintering process, and obtains the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; collects online data as a test set for the nonlinear dynamic system model, and performs standardization processing; detects the test set based on the constructed nonlinear dynamic system model of the sintering process, calculates the statistics of the test samples, compares the calculated statistics with the preset monitoring control limits to determine whether a fault has occurred, and outputs the detection results. The present invention provides a monitoring method and system for the preparation process of ternary cathode materials based on RVAE, which takes into account both the dynamic characteristics of process time delay and nonlinear features to ensure the stable operation of the ternary cathode material preparation system; it can significantly improve the fault detection rate and false alarm rate, and provide strong support for the stable operation of the sintering process. Attached Figure Description

[0066] Figure 1 A schematic flowchart of an embodiment of the RVAE-based ternary cathode material preparation process monitoring method provided by the present invention;

[0067] Figure 2 for Figure 1 The diagram shows a detailed flowchart of an embodiment of the steps for constructing a nonlinear dynamic system model of the sintering process based on a variational autoencoder.

[0068] Figure 3 The method for monitoring the preparation process of ternary cathode materials based on RVAE provided by this invention includes a supervised NDS model structure.

[0069] Figure 4 The nonlinear dynamic model structure of RVAE in the monitoring method for the preparation process of ternary cathode materials based on RVAE provided by the present invention;

[0070] Figure 5 The RNN model structure in the RVAE-based ternary cathode material preparation process monitoring method provided by this invention;

[0071] Figure 6 The flowchart of the process monitoring based on the RVAE nonlinear dynamic model in the RVAE-based ternary cathode material preparation process monitoring method provided by the present invention;

[0072] Figure 7 The following figures illustrate the monitoring results of different methods for abnormal temperature rise in roller kilns in the RVAE-based ternary cathode material preparation process monitoring method provided by this invention: Figure (a) shows the monitoring results of DPLVM for abnormal temperature rise in roller kilns, Figure (b) shows the monitoring results of DALM for abnormal temperature rise in roller kilns, Figure (c) shows the monitoring results of SVAE for abnormal temperature rise in roller kilns, and Figure (d) shows the monitoring results of RVAE for abnormal temperature rise in roller kilns.

[0073] Figure 8 The following figures illustrate the monitoring results of different methods for abnormal temperature drop in roller kilns in the RVAE-based ternary cathode material preparation process monitoring method provided by this invention: Figure (a) shows the monitoring results of DPLVM for abnormal temperature drop in roller kilns, Figure (b) shows the monitoring results of DALM for abnormal temperature drop in roller kilns, Figure (c) shows the monitoring results of SVAE for abnormal temperature drop in roller kilns, and Figure (d) shows the monitoring results of RVAE for abnormal temperature drop in roller kilns.

[0074] Figure 9 The following are monitoring results of different methods for roller kiln shutdown failure in the RVAE-based ternary cathode material preparation process monitoring method provided by the present invention: Figure (a) shows the monitoring results of DPLVM for roller kiln shutdown failure, Figure (b) shows the monitoring results of DALM for roller kiln shutdown failure, Figure (c) shows the monitoring results of SVAE for roller kiln shutdown failure, and Figure (d) shows the monitoring results of roller kiln shutdown failure.

[0075] Figure 10 A functional block diagram of an embodiment of the RVAE-based ternary cathode material preparation process monitoring system provided by the present invention;

[0076] Figure 11 for Figure 10 The diagram shows a functional module schematic of one embodiment of the building module.

[0077] Explanation of icon numbers:

[0078] 10. Construction Module; 20. Derivation Module; 30. Estimation Module; 40. Collection Module; 50. Comparison Module; 11. Processing Unit; 12. Acquisition Unit. Detailed Implementation

[0079] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0080] like Figures 1 to 9 As shown, the first embodiment of the present invention proposes a method for monitoring the preparation process of ternary cathode materials based on RVAE, including the following steps:

[0081] Step S100: Construct a nonlinear dynamic system model of the sintering process based on a variational autoencoder.

[0082] Based on the nonlinear framework of variational autoencoder, an autoregressive equation for latent variables is constructed by combining a recurrent neural network, resulting in a nonlinear dynamic model of the sintering process based on the recurrent variational autoencoder.

[0083] Step S200: Establish an autoregressive equation between latent variables through a recurrent neural network, and introduce a weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, derive the loss function of the nonlinear dynamic system model of the sintering process, and train the model parameters through backpropagation.

[0084] An autoregressive equation between latent variables is established using a recurrent neural network, and a weighted moving average method is introduced to assign different weights to samples at different times. The loss function of the final nonlinear dynamic model is derived, and the model parameters are trained through backpropagation.

[0085] Step S300: Based on the nonlinear dynamic system model of the sintering process using a cyclic variational autoencoder, define the statistics of the nonlinear dynamic system model of the sintering process, and obtain the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation.

[0086] Based on the nonlinear dynamic system model based on a cyclic variational autoencoder, the T of this model is defined. 2 The t-statistic is used to obtain the control threshold of the model through kernel density estimation.

[0087] Step S400: Collect online data as a test set for the nonlinear dynamic system model and perform standardization processing.

[0088] Collect online data x(k), k = 1, 2, ..., N as the test set for the model, and perform standardization.

[0089] Step S500: Based on the constructed nonlinear dynamic system model of the sintering process, the test set is detected, the statistics of the test samples are calculated, the calculated statistics are compared with the preset monitoring and control limits to determine whether a fault has occurred, and the detection results are output.

[0090] The test set is tested based on the obtained nonlinear dynamic system model, and the T of the test samples is calculated. 2 (t), T 2 (t) and monitoring control limits Compare and output the detection results.

[0091] Furthermore, such as Figure 2 As shown, Figure 2 for Figure 1 The detailed flowchart of step S100 shown in the figure is illustrated in this embodiment. In this embodiment, step S100 includes:

[0092] Step S110: Use the collected historical data as the model training set, and standardize each sample in the model training set.

[0093] Data preparation and preprocessing: Collect a historical data set x h (k), k=1,2,..,T is used as the training set for the model, where x h (k)∈T m The data consists of m-dimensional process observations, where k is the time label and T is the number of samples; each sample is standardized.

[0094] Preprocess or standardize the dataset. The standardization operation is as follows: for each element in the sample set, first subtract the sample mean of the variable to which it belongs, and then divide by the standard deviation of the sample, so that the mean of the data corresponding to each process variable and key quality variable is 0 and the variance is 1.

[0095] Step S120: Based on the nonlinear framework of variational autoencoder, and combined with recurrent neural network, construct the autoregressive equation of latent variables to obtain the nonlinear dynamic system model of sintering process based on recurrent variational autoencoder.

[0096] First, a supervised nonlinear dynamic system model is established within the framework of VAE (Variational Autoencoder). VAE is an unsupervised generative model, consisting of an inference network (encoder) and a generator network (decoder). The inference network maps the input x to the feature space, obtaining the mean and variance of the latent variable z. The generator network randomly generates the input x from the latent variable z. Unlike traditional methods, VAE does not require calculating the joint distribution of the inputs; instead, it fits the target through a neural network. The log-likelihood function of the marginal probabilities of VAE can be expressed as:

[0097]

[0098] In formula (1), q(z|φ) is a variational approximation of the computationally difficult-to-calculate posterior distribution p(z|θ), φ and θ are the parameters to be estimated, ELBO is the variational lower bound of the log-likelihood function, and D... KL (q(z|φ)‖p(z|θ)) represents the KL divergence between q(z|φ) and p(z|θ), which measures the difference in probability distributions between the two.

[0099] Supervised NDS (Nonlinear Dynamic System) model structure as follows Figure 3 As shown, supervised NDS models can effectively extract the nonlinear features of process variables for latent variable regression through inference networks, but the state transitions between latent variables are still achieved through first-order Markov chains. Therefore, NDS models can only extract low-order dynamic characteristics of process variables.

[0100] However, due to factors such as the multi-temperature coupling and long sintering cycle in the ternary cathode material sintering process, the variables exhibit time-delay dynamic characteristics. To fully explore the dynamic time-series characteristics of the variables in the sintering process, an RNN (Recurrent Neural Network) is introduced on top of the NDS to establish an autoregressive equation between the latent variables, constructing a nonlinear dynamic model based on RVAE. The model structure is as follows: Figure 4 As shown, the mathematical structure is as follows:

[0101]

[0102] In formula (2), z(1) is the latent variable at time 1, μ(1) and σ(1) are the mean and variance calculated by the inference network at time 1; z(t) is the latent variable at time t. and Let μ(t) and σ(t) be the process variables and quality variables reconstructed at time t, respectively; μ(t) and σ(t) are the mean and variance calculated by the inference network at time t; ε is the reparameter factor, following a standard Gaussian distribution; B(t) is the transition matrix at time t; and T represents the time delay coefficient of the process. Figure 2 The dynamic model length is represented in the equation; h(t-1) is the state increment containing latent variables over the past T time periods, which can be estimated by an RNN; q(*), f(*), and g(*) are nonlinear functions approximated by the neural network.

[0103] Specifically, the RVAE nonlinear dynamic model is constructed by introducing an RNN into the recognition network, generation network, and latent variable connection part of the supervised NDS model. The latent variables extracted by the sub-model inference network at different time points are used as inputs to the RNN. The RNN performs deep feature extraction on the temporal dynamic features of the latent variables and obtains the estimated value of the latent variable state transition at time T. This completes the construction of the latent variable autoregressive equation based on the supervised NDS model.

[0104] RNN structure as follows Figure 5 As shown, assuming the input unit of the RNN is the latent variable z(t) calculated by the inference network at different times, and the hidden layer unit, also called the memory unit, is represented as h(t), the dynamic temporal relationship of the latent variable z(t) can be extracted through the memory unit. According to the forward propagation algorithm, the output of the hidden layer at different times is:

[0105] h(t)=f(Uz(t)+Wh(t-1)+b)(3)

[0106] In formula (3), U and W are weight matrices, h(t-1) is the hidden layer output at the previous time step, b is the bias term, and f(*) is the activation function, which is often a non-linear activation function. The network parameters can be learned through backpropagation of the gradients unfolded over time.

[0107] In step S200, based on formula (2) and the structure of the RVAE nonlinear dynamic model, the detailed derivation of its joint probability log-likelihood function is shown in formula (4):

[0108]

[0109] In formula (4), T is the length of the dynamic model, θ=[θ(1),……,θ(T)] and φ=[φ(1),……,φ(T)] are the parameters to be estimated, p(x(1),y(1),…,x(t),y(t)|Θ) is the joint probability distribution; q(z(1)|φ(1)) is the initial posterior distribution; p(z(1)|θ(1)) is the initial prior distribution; q(z(t)|h(t-1),φ(t)) is the dynamic posterior distribution; p(z(t)|h(t-1),θ(t)) is the dynamic prior distribution; x(t) is the process variable at time t, y(t) is the quality variable at time t, x(1) is the process variable at time 1, y(1) is the quality variable at time 1, D KLLet z(t) be the KL divergence, z(1) be the initial hidden variable, φ(1) and θ(1) be the initial neural network parameters, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, and φ(t) and θ(t) be the neural network parameters at time t. The log-likelihood function of the nonlinear dynamic system based on RVAE consists of two parts: the variational lower bound ELBO(1) and KL (Kullback-Leibler divergence) divergence of the initial RVAE, and the variational lower bound ELBO(t) and KL divergence of the subsequent dynamic RVAE.

[0110] Assuming x(t) and y(t) are conditionally independent, the variational lower bound of the joint probability log-likelihood function can be further derived as follows:

[0111]

[0112] In formula (5), ELBO(1) is the variational lower bound of the initial RVAE, ELBO(t) is the variational lower bound of the subsequent dynamic RVAE, p(x(1),y(1),…,x(t),y(t)|Θ) is the joint probability distribution; q(z(1)|φ(1)) is the initial posterior distribution; p(z(1)|θ(1)) is the initial prior distribution; q(z(t)|h(t-1),φ(t)) is the dynamic posterior distribution; p(z(t)|h(t-1),θ(t)) is the dynamic prior distribution; x(t) is the process variable at time t, y(t) is the mass variable at time t, x(1) is the process variable at time 1, y(1) is the mass variable at time 1, D KL Let z(t) be the KL divergence, z(1) be the initial hidden variable, φ(1) and θ(1) be the initial neural network parameters, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, and φ(t) and θ(t) be the neural network parameters at time t.

[0113] Similar to VAEs, maximizing the joint probability can be transformed into maximizing the variational lower bound:

[0114]

[0115] In formula (6), ELBO(1) is the variational lower bound of the initial RVAE, and ELBO(t) is the variational lower bound of the subsequent dynamic RVAE. z(t)～q(z(t)|h(t-1),φ(t)) [lnp(x(t),y(t)|z(t),θ(t))] represents the log-expectation of the probability distribution p(x(t),y(t)|z(t),θ(t)) under the latent variable z(t)~q(z(t)|h(t-1),φ(t)).

[0116] Assume the prior distribution at initial t=1 is p(z(1)|θ(1))=N(0,I), and the dynamic prior distribution at t>1 is p(z(t)|h(t-1),θ(t))=N(q(h(t-1)),I); the posterior estimate at initial t=1 is q(z(1)|φ(1))=N(μ(1),σ 2 (1)I), the dynamic posterior estimate when t>1 is q(z(t)|h(t-1)),φ(t)=N(q(h(t-1))+B(t)μ(t),σ 2 (t)I); and p(x(t)|z(t))=N(μ x(t) ,I) and p(y(t)|z(t))=N(μ y(t) If I), then the loss function of the RVAE nonlinear dynamic model can be simplified to:

[0117]

[0118] In formula (7), J(Φ,Θ) is the loss function of the RVAE nonlinear dynamic model, p(z(1)|θ(1)) is the prior distribution when t=1, p(z(t)|h(t-1),θ(t)) is the dynamic prior distribution when t>1, q(z(1)|φ(1)) is the posterior estimate when t=1, and q(z(t)|h(t-1)),φ(t) is the dynamic posterior estimate when t>1.

[0119] Considering the non-uniform dynamics of process variables, a weighted moving average method is introduced in formula (7) to assign different weights to the sub-models at different times, thereby dynamically measuring the importance of samples at different times and fully exploring the dynamic characteristics between latent variables. The weight formula is shown in formula (8), where ξ represents the weight decay factor. The larger ξ is, the faster the weight decays, and the more obvious the distinction between the importance of sub-models at different times. In this paper, ξ = 0.1 is selected.

[0120] weight(t) = 1 - (Tt) × ξ (8)

[0121] In formula (8), weight(t) is the weight of the sample at time t, and ξ represents the weight decay factor.

[0122] Therefore, the loss function can be modified as follows:

[0123]

[0124] In formula (9), J(Φ,Θ) is the loss function, weight(t) is the weight of the sample at time t, x(t) is the process variable at time t, and μ x(t) Let y(t) be the expected mean of the process variable, and μ be the mass variable at time t. y(t)Let D be the expected mean of the quality variable, weight(0) be the initial sample weights, and D be the mean of the quality variable. KL Let z(1) be the KL divergence, z(1) be the initial hidden variable, φ(1) be the initial neural network parameters, N(0,I) be the standard Gaussian distribution, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, φ(t) be the neural network parameters at time t, I be the identity matrix, and q(*) and p(*) be the nonlinear functions approximated by the neural network.

[0125] Once the loss function is determined, the parameters of the RVAE nonlinear dynamic model can be trained by backpropagation of the training samples.

[0126] In step S300, it is necessary to define T for this model. 2 The (t) statistic is used to obtain the fault control limit threshold through kernel density estimation.

[0127] As can be seen from formula (2), the process variable x and the learned f(z(t)) are linearly related. Therefore, for [x,z]∈R N×(v+k) PCA is performed, where the latent variable z is estimated by the mean μ(t) learned by the inference network, N represents the sample size, and v and k are the process variable dimension and latent variable dimension, respectively. The variable space is then decomposed into mutually orthogonal principal component subspaces and residual subspaces, and a T is constructed in the principal component subspace. 2 The statistics are as follows:

[0128]

[0129] In formula (10), H(t) = [x(t), z(t)] are the input variables. Let n be a diagonal matrix of singular values ​​obtained by singular value decomposition of H(t), arranged in ascending order; z The number of its principal components can be determined using the cumulative variance contribution method. for The principal space spanned by the corresponding eigenvectors.

[0130] For the collected normal historical process data, T at each time point is calculated according to formula (10). 2 (t), and obtain T using kernel density estimation. 2 The probability density distribution of the statistic, taking the T value corresponding to a confidence level α = 0.99. 2 The value of the statistic is used as the monitoring control limit. This means that the prior probability of a false alarm is 0.01.

[0131] T is calculated online using process data. 2 (t) statistic and control limits The comparison is used to determine whether a fault has occurred, and the judgment logic is shown in formula (11):

[0132]

[0133] In formula (11), T 2 (t) is a statistic. For control limits.

[0134] The RVAE-based monitoring method for the preparation process of ternary cathode materials provided in this embodiment, compared with the prior art, constructs a nonlinear dynamic system model of the sintering process based on a variational autoencoder; establishes an autoregressive equation between latent variables through a recurrent neural network, and introduces a weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, derives the loss function of the nonlinear dynamic system model of the sintering process, and trains the model parameters through backpropagation; defines the statistics of the nonlinear dynamic system model of the sintering process based on the cyclic variational autoencoder, and obtains the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; collects online data as a test set for the nonlinear dynamic system model of the nonlinear dynamic system and performs standardization processing; detects the test set based on the constructed nonlinear dynamic system model of the sintering process, calculates the statistics of the test samples, compares the calculated statistics with the preset monitoring control limits to determine whether a fault has occurred, and outputs the detection results. The RVAE-based monitoring method for the preparation process of ternary cathode materials provided in this embodiment considers both the dynamic characteristics of the process time delay and the nonlinear characteristics to ensure the stable operation of the ternary cathode material preparation system; it can significantly improve the fault detection rate and false alarm rate, and provide strong support for the stable operation of the sintering process.

[0135] like Figure 10 As shown, Figure 10This is a functional block diagram of an embodiment of the RVAE-based ternary cathode material preparation process monitoring system provided by the present invention. In this embodiment, the RVAE-based ternary cathode material preparation process monitoring system includes a construction module 10, a derivation module 20, an estimation module 30, a collection module 40, and a comparison module 50. The construction module 10 is used to construct a nonlinear dynamic system model of the sintering process based on a variational autoencoder. The derivation module 20 is used to establish an autoregressive equation between latent variables through a recurrent neural network, and introduce a weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, thereby deriving the nonlinear dynamic system model of the sintering process. The loss function is used to train the model parameters through backpropagation; the estimation module 30 is used to define the statistics of the nonlinear dynamic system model of the sintering process based on the nonlinear dynamic system model of the cyclic variational autoencoder, and obtain the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; the collection module 40 is used to collect online data as the test set of the nonlinear dynamic system model and perform standardization processing; the comparison module 50 is used to detect the test set based on the constructed nonlinear dynamic system model of the sintering process, calculate the statistics of the test samples, compare the calculated statistics with the preset monitoring control limits to determine whether a fault has occurred, and output the detection results.

[0136] Module 10 constructs a nonlinear framework based on variational autoencoders, and uses a recurrent neural network to construct autoregressive equations for latent variables, thereby obtaining a nonlinear dynamic model of the sintering process based on a recurrent variational autoencoder.

[0137] The derivation module 20 establishes an autoregressive equation between latent variables through a recurrent neural network, and introduces a weighted moving average method to assign different weights to samples at different times, deriving the loss function of the final nonlinear dynamic model, and training the model parameters through backpropagation.

[0138] Estimation module 30 defines T based on the nonlinear dynamic system model based on a cyclic variational autoencoder. 2 The t-statistic is used to obtain the control threshold of the model through kernel density estimation.

[0139] The collection module 40 collects online data x(k), k = 1, 2, ..., N as the test set for the model and performs standardization processing.

[0140] Comparison module 50 performs testing on the test set based on the obtained nonlinear dynamic system model and calculates the T of the test samples. 2 (t), T 2 (t) and monitoring control limits Compare and output the detection results.

[0141] Preferably, please see Figure 11 , Figure 11 for Figure 10 The diagram shows a functional module schematic of one embodiment of the construction module. In this embodiment, the construction module 10 includes a processing unit 11 and an acquisition unit 12. The processing unit 11 is used to use the collected historical data as a model training set and to standardize each sample in the model training set. The acquisition unit 12 is used to construct an autoregressive equation for the latent variables based on the nonlinear framework of the variational autoencoder and in conjunction with a recurrent neural network, so as to obtain a nonlinear dynamic system model of the sintering process based on the recurrent variational autoencoder.

[0142] Processing Unit 11 Data Preparation and Preprocessing: Collecting a segment of historical data x h (k), k=1,2,..,T is used as the training set for the model, where x h (k)∈T m The data consists of m-dimensional process observations, where k is the time label and T is the number of samples; each sample is standardized.

[0143] Preprocess or standardize the dataset. The standardization operation is as follows: for each element in the sample set, first subtract the sample mean of the variable to which it belongs, and then divide by the standard deviation of the sample, so that the mean of the data corresponding to each process variable and key quality variable is 0 and the variance is 1.

[0144] First, a supervised nonlinear dynamic system model is established within the framework of VAE (Variational Autoencoder). VAE is an unsupervised generative model, consisting of an inference network (encoder) and a generator network (decoder). The inference network maps the input x to the feature space, obtaining the mean and variance of the latent variable z. The generator network randomly generates the input x from the latent variable z. Unlike traditional methods, VAE does not require calculating the joint distribution of the inputs; instead, it fits the target through a neural network. The log-likelihood function of the marginal probabilities of VAE can be expressed as:

[0145]

[0146] In formula (12), q(z|φ) is a variational approximation of the computationally difficult-to-calculate posterior distribution p(z|θ), φ and θ are the parameters to be estimated, ELBO is the variational lower bound of the log-likelihood function, and D... KL (q(z|φ)‖p(z|θ)) represents the KL divergence between q(z|φ) and p(z|θ), which measures the difference in probability distributions between the two.

[0147] Supervised NDS (Nonlinear Dynamic System) model structure as follows Figure 3As shown, supervised NDS models can effectively extract the nonlinear features of process variables for latent variable regression through inference networks, but the state transitions between latent variables are still achieved through first-order Markov chains. Therefore, NDS models can only extract low-order dynamic characteristics of process variables.

[0148] However, due to factors such as the multi-temperature coupling and long sintering cycle in the ternary cathode material sintering process, the variables exhibit time-delay dynamic characteristics. To fully explore the dynamic time-series characteristics of the variables in the sintering process, an RNN (Recurrent Neural Network) is introduced on top of the NDS to establish an autoregressive equation between the latent variables, constructing a nonlinear dynamic model based on RVAE. The model structure is as follows: Figure 4 As shown, the mathematical structure is as follows:

[0149]

[0150] In formula (13), z(1) is the latent variable at time 1, μ(1) and σ(1) are the mean and variance calculated by the inference network at time 1; z(t) is the latent variable at time t. and Let μ(t) and σ(t) be the process variables and quality variables reconstructed at time t, respectively; μ(t) and σ(t) are the mean and variance calculated by the inference network at time t; ε is the reparameter factor, following a standard Gaussian distribution; B(t) is the transition matrix at time t; and T represents the time delay coefficient of the process. Figure 2 The dynamic model length is represented in the equation; h(t-1) is the state increment containing latent variables over the past T time periods, which can be estimated by an RNN; q(*), f(*), and g(*) are nonlinear functions approximated by the neural network.

[0151] Specifically, the RVAE nonlinear dynamic model is constructed by introducing an RNN into the recognition network, generation network, and latent variable connection part of the supervised NDS model. The latent variables extracted by the sub-model inference network at different time points are used as inputs to the RNN. The RNN performs deep feature extraction on the temporal dynamic features of the latent variables and obtains the estimated value of the latent variable state transition at time T. This completes the construction of the latent variable autoregressive equation based on the supervised NDS model.

[0152] RNN structure as follows Figure 5 As shown, assuming the input unit of the RNN is the latent variable z(t) calculated by the inference network at different times, and the hidden layer unit, also called the memory unit, is represented as h(t), the dynamic temporal relationship of the latent variable z(t) can be extracted through the memory unit. According to the forward propagation algorithm, the output of the hidden layer at different times is:

[0153] h(t)=f(Uz(t)+Wh(t-1)+b)(14)

[0154] In formula (14), U and W are weight matrices, h(t-1) is the hidden layer output at the previous time step, b is the bias term, and f(*) is the activation function, which is often a non-linear activation function. The network parameters can be learned through backpropagation of gradients unfolded over time.

[0155] In derivation module 20, based on formula (13) and the structure of the RVAE nonlinear dynamic model, the detailed derivation of its joint probability log-likelihood function is shown in formula (15):

[0156]

[0157] In formula (15), T is the length of the dynamic model, θ = [θ(1), ..., θ(T)] and φ = [φ(1), ..., φ(T)] are the parameters to be estimated, p(x(1), y(1), ..., x(t), y(t)|Θ) is the joint probability distribution; q(z(1)|φ(1)) is the initial posterior distribution; p(z(1)|θ(1)) is the initial prior distribution; q(z(t)|h(t-1), φ(t)) is the dynamic posterior distribution; p(z(t)|h(t-1), θ(t)) is the dynamic prior distribution; x(t) is the process variable at time t, y(t) is the quality variable at time t, x(1) is the process variable at time 1, y(1) is the quality variable at time 1, and D KL Let z(t) be the KL divergence, z(1) be the initial hidden variable, φ(1) and θ(1) be the initial neural network parameters, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, and φ(t) and θ(t) be the neural network parameters at time t. The log-likelihood function of the nonlinear dynamic system based on RVAE consists of two parts: the variational lower bound ELBO(1) and KL (Kullback-Leibler divergence) divergence of the initial RVAE, and the variational lower bound ELBO(t) and KL divergence of the subsequent dynamic RVAE.

[0158] Assuming x(t) and y(t) are conditionally independent, the variational lower bound of the joint probability log-likelihood function can be further derived as follows:

[0159]

[0160] In formula (16), ELBO(1) is the variational lower bound of the initial RVAE, ELBO(t) is the variational lower bound of the subsequent dynamic RVAE, p(x(1),y(1),…,x(t),y(t)|Θ) is the joint probability distribution; q(z(1)|φ(1)) is the initial posterior distribution; p(z(1)|θ(1)) is the initial prior distribution; q(z(t)|h(t-1),φ(t)) is the dynamic posterior distribution; p(z(t)|h(t-1),θ(t)) is the dynamic prior distribution; x(t) is the process variable at time t, y(t) is the mass variable at time t, x(1) is the process variable at time 1, y(1) is the mass variable at time 1, D KL Let z(t) be the KL divergence, z(1) be the initial hidden variable, φ(1) and θ(1) be the initial neural network parameters, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, and φ(t) and θ(t) be the neural network parameters at time t.

[0161] Similar to VAEs, maximizing the joint probability can be transformed into maximizing the variational lower bound:

[0162]

[0163] In formula (17), ELBO(1) is the variational lower bound of the initial RVAE, and ELBO(t) is the variational lower bound of the subsequent dynamic RVAE. z(t)～q(z(t)|h(t -1), φ(t)) [lnp(x(t),y(t)|z(t),θ(t))] represents the log-expectation of the probability distribution p(x(t),y(t)|z(t),θ(t)) under the latent variable z(t)~q(z(t)|h(t-1),φ(t)).

[0164] Assume the prior distribution at initial t=1 is p(z(1)|θ(1))=N(0,I), and the dynamic prior distribution at t>1 is p(z(t)|h(t-1),θ(t))=N(q(h(t-1)),I); the posterior estimate at initial t=1 is q(z(1)|φ(1))=N(μ(1),σ 2 (1)I), the dynamic posterior estimate when t>1 is q(z(t)|h(t-1)),φ(t)=N(q(h(t-1))+B(t)μ(t),σ 2 (t)I); and p(x(t)|z(t))=N(μ x(t) ,I) and p(y(t)|z(t))=N(μ y(t) If I), then the loss function of the RVAE nonlinear dynamic model can be simplified to:

[0165]

[0166] In formula (18), J(Φ,Θ) is the loss function of the RVAE nonlinear dynamic model, p(z(1)|θ(1)) is the prior distribution when t=1, p(z(t)|h(t-1),θ(t)) is the dynamic prior distribution when t>1, q(z(1)|φ(1)) is the posterior estimate when t=1, and q(z(t)|h(t-1)),φ(t) is the dynamic posterior estimate when t>1.

[0167] Considering the non-uniform dynamics of process variables, a weighted moving average method is introduced in formula (18) to assign different weights to the sub-models at different times, thereby dynamically measuring the importance of samples at different times and fully exploring the dynamic characteristics between latent variables. The weight formula is shown in formula (8), where ξ represents the weight decay factor. The larger ξ is, the faster the weight decays, and the more obvious the distinction between the importance of sub-models at different times. In this paper, ξ = 0.1 is selected.

[0168] weight(t) = 1 - (Tt) × ξ (19)

[0169] In formula (19), weight(t) is the weight of the sample at time t, and ξ represents the weight decay factor.

[0170] Therefore, the loss function can be modified as follows:

[0171]

[0172] In formula (20), J(Φ,Θ) is the loss function, weight(t) is the weight of the sample at time t, x(t) is the process variable at time t, and μ x(t) Let y(t) be the expected mean of the process variable, and μ be the mass variable at time t. y(t) Let D be the expected mean of the quality variable, weight(0) be the initial sample weights, and D be the mean of the quality variable. KL Let z(1) be the KL divergence, z(1) be the initial hidden variable, φ(1) be the initial neural network parameters, N(0,I) be the standard Gaussian distribution, z(t) be the hidden variable at time t, h(t-1) be the state increment of the hidden variable at time t-1, φ(t) be the neural network parameters at time t, I be the identity matrix, and q(*) and p(*) be the nonlinear functions approximated by the neural network.

[0173] Once the loss function is determined, the parameters of the RVAE nonlinear dynamic model can be trained by backpropagation of the training samples.

[0174] In step S300, it is necessary to define T for this model. 2 The (t) statistic is used to obtain the fault control limit threshold through kernel density estimation.

[0175] As can be seen from formula (13), the process variable x and the learned f(z(t)) are linearly related. Therefore, for [x,z]∈R N×(v+k) PCA is performed, where the latent variable z is estimated by the mean μ(t) learned by the inference network, N represents the sample size, and v and k are the process variable dimension and latent variable dimension, respectively. The variable space is then decomposed into mutually orthogonal principal component subspaces and residual subspaces, and a T is constructed in the principal component subspace. 2 The statistics are as follows:

[0176]

[0177] In formula (21), H(t) = [x(t), z(t)] are the input variables. Let n be a diagonal matrix of singular values ​​obtained by singular value decomposition of H(t), arranged in ascending order; z The number of its principal components can be determined using the cumulative variance contribution method. for The principal space spanned by the corresponding eigenvectors.

[0178] For the collected normal historical process data, T at each time point is calculated according to formula (21). 2 (t), and obtain T using kernel density estimation. 2 The probability density distribution of the statistic, taking the T value corresponding to a confidence level α = 0.99. 2 The value of the statistic is used as the monitoring control limit. This means that the prior probability of a false alarm is 0.01.

[0179] T is calculated online using process data. 2 (t) statistic and control limits The comparison is used to determine whether a fault has occurred, and the judgment logic is shown in formula (22):

[0180]

[0181] In formula (22), T 2 (t) is a statistic. For control limits.

[0182] The RVAE-based monitoring system for the preparation process of ternary cathode materials provided in this embodiment, compared with existing technologies, constructs a nonlinear dynamic system model of the sintering process based on a variational autoencoder; establishes an autoregressive equation between latent variables through a recurrent neural network, and introduces a weighted moving average method to assign different weights to samples at different times in the constructed nonlinear dynamic system model of the sintering process, deriving the loss function of the nonlinear dynamic system model of the sintering process, and training the model parameters through backpropagation; defines the statistics of the nonlinear dynamic system model of the sintering process based on the cyclic variational autoencoder, and obtains the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; collects online data as a test set for the nonlinear dynamic system model of the nonlinear dynamic system, and performs standardization processing; detects the test set based on the constructed nonlinear dynamic system model of the sintering process, calculates the statistics of the test samples, compares the calculated statistics with the preset monitoring control limits to determine whether a fault has occurred, and outputs the detection results. The RVAE-based monitoring system for the preparation of ternary cathode materials provided in this embodiment considers both the dynamic characteristics of the process time delay and the nonlinear features to ensure the stable operation of the ternary cathode material preparation system; it can significantly improve the fault detection rate and false alarm rate, and provide strong support for the stable operation of the sintering process.

[0183] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for monitoring the preparation process of ternary cathode material based on RVAE, characterized in that, The method comprises the following steps: constructing a sintering process nonlinear dynamic system model based on a variational autoencoder; establishing an autoregressive equation between hidden variables through a recurrent neural network, introducing a weighted moving average method to give different weights to samples at different times in the constructed sintering process nonlinear dynamic system model, deriving a loss function of the sintering process nonlinear dynamic system model, and training model parameters through backpropagation error propagation; according to the nonlinear dynamic system model based on the recurrent variational autoencoder, defining a statistical quantity of the sintering process nonlinear dynamic system model, and obtaining a control threshold of the sintering process nonlinear dynamic system model through kernel density estimation; collecting online data as a test set of the nonlinear dynamic system model and performing standardization processing; detecting the test set based on the constructed sintering process nonlinear dynamic system model, calculating a statistical quantity of a test sample, comparing the calculated statistical quantity with a preset monitoring control limit to determine whether a fault occurs, and outputting a detection result; the step of constructing a sintering process nonlinear dynamic system model based on a variational autoencoder comprises: collecting historical data as a model training set, and performing standardization processing on each specimen in the model training set; based on a nonlinear framework of a variational autoencoder, a recurrent neural network is used to construct an autoregressive equation of hidden variables to obtain a sintering process nonlinear dynamic system model based on a recurrent variational autoencoder; in the step of constructing an autoregressive equation of hidden variables based on a nonlinear framework of a variational autoencoder and a recurrent neural network to obtain a sintering process nonlinear dynamic system model based on a recurrent variational autoencoder, the mathematical structure of the sintering process nonlinear dynamic system model is: in, For the latent variable at time 1, and These are the mean and variance calculated by the inference network at time step 1; For the first Latent variables at each moment; and The first Process variables and quality variables reconstructed at each moment; and For the first The mean and variance calculated by the inference network at each time step; It is a multiple parameter factor that follows a standard Gaussian distribution; For the first Transition matrix at each time step; The time delay coefficient represents the process; For including the past The state increment of a time-length hidden variable; , and It is a nonlinear function approximated by a neural network.

2. The method of claim 1, wherein the process of preparing a ternary cathode material based on RVAE is characterized by, in the step of establishing an autoregressive equation between hidden variables through a recurrent neural network and introducing a weighted moving average method to give different weights to samples at different times in the constructed sintering process nonlinear dynamic system model, deriving a loss function of the final nonlinear dynamic system model, and training model parameters through backpropagation error propagation, according to the forward propagation algorithm, the hidden layer output of the recurrent neural network at different times is: wherein, and is a weight matrix, is a hidden variable computed by the inference network at different time, is a hidden layer unit, is a hidden layer output at the previous time, is a bias term, is an activation function; the loss function of the sintering process nonlinear dynamic system model derived is: where, is the loss function, is the weight sample weight at time t, is the process variable at time t, is the expected mean of the process variable, is the quality variable at time t, is the expected mean of the quality variable, is the initial sample weight, is the KL divergence, is the initial latent variable at time 0, is the initial neural network parameter, is the standard Gaussian distribution, is the latent variable at time t, is the latent variable state increment at time t - 1, is the neural network parameter at time t, is the identity matrix, and is a non-linear function approximated by a neural network.

3. The method of claim 1, wherein the process of preparing a ternary cathode material based on RVAE is characterized by, in the step of detecting the test set based on the constructed sintering process nonlinear dynamic system model, calculating a statistical quantity of a test sample, comparing the calculated statistical quantity with a preset monitoring control limit to determine whether a fault occurs, and outputting a detection result, the calculated statistical quantity of the test sample is: wherein is a statistical quantity; is an input variable; is a singular value obtained by singular value decomposition on is a number of its pivots; is is a pivot space spanned by the corresponding eigenvectors; is a transpose of the pivot space, is a transpose of the input variable, is an inverse matrix of an inverse diagonal matrix composed of the singular values;​ whether a fault occurs is determined by comparing the statistical quantity of the process data calculated online with the control limit, and the determination logic is: wherein is a statistical quantity, is a control limit.

4. A RVAE-based ternary cathode material preparation process monitoring system, characterized in that, comprises: a construction module (10) for constructing a sintering process nonlinear dynamic system model based on a variational autoencoder; the step of constructing a sintering process nonlinear dynamic system model based on a variational autoencoder comprises: collecting historical data as a model training set, and performing standardization processing on each specimen in the model training set; Based on the nonlinear framework of the variational autoencoder, the recurrent neural network constructs the autoregressive equation of the hidden variable, and obtains the nonlinear dynamic system model of the sintering process based on the recurrent variational autoencoder; In the step of obtaining the nonlinear dynamic system model of the sintering process based on the nonlinear framework of the variational autoencoder and the recurrent neural network constructing the autoregressive equation of the hidden variable, the mathematical structure of the nonlinear dynamic system model of the sintering process obtained is: in, For the latent variable at time 1, and These are the mean and variance calculated by the inference network at time step 1; For the first Latent variables at each moment; and The first Process variables and quality variables reconstructed at each moment; and For the first The mean and variance calculated by the inference network at each time step; It is a multiple parameter factor that follows a standard Gaussian distribution; For the first Transition matrix at each time step; The time delay coefficient represents the process; For including the past The state increment of a time-length hidden variable; , and For neural networks, it is a nonlinear function approximated. The derivation module (20) is used to establish the autoregressive equation between the hidden variables through the recurrent neural network, introduce the weighted moving average method to give different weights to the samples at different times in the constructed nonlinear dynamic system model of the sintering process, derive the loss function of the nonlinear dynamic system model of the sintering process, and train the model parameters through backpropagation error propagation; The estimation module (30) is used to define the statistics of the nonlinear dynamic system model of the sintering process according to the nonlinear dynamic system model based on the recurrent variational autoencoder, and obtain the control threshold of the nonlinear dynamic system model of the sintering process through kernel density estimation; The collection module (40) is used to collect online data as a test set of the nonlinear dynamic system model and perform standardization processing; The comparison module (50) is used to detect the test set based on the constructed nonlinear dynamic system model of the sintering process, calculate the statistics of the test sample, compare the calculated statistics with the preset monitoring control limit to determine whether a fault occurs, and output the detection result.

5. The RVAE-based ternary cathode material preparation process monitoring system of claim 4, wherein, The construction module (10) comprises: The processing unit (11) is used to standardize each specimen in the model training set by taking the collected historical data as the model training set; The acquisition unit (12) is used to construct the autoregressive equation of the hidden variable based on the nonlinear framework of the variational autoencoder and the recurrent neural network, and obtain the nonlinear dynamic system model of the sintering process based on the recurrent variational autoencoder.

6. The RVAE-based ternary cathode material preparation process monitoring system of claim 4, wherein, In the derivation module (20), according to the forward propagation algorithm, the hidden layer output of the recurrent neural network at different times is: wherein, and is a weight matrix, is a hidden variable computed by the inference network at different time, is a hidden layer unit, is a hidden layer output at the previous time, is a bias term, is an activation function; The loss function of the derived nonlinear dynamic system model of the sintering process is: where, is the loss function, is the weight sample weight at time t, is the process variable at time t, is the expected mean of the process variable, is the quality variable at time t, is the expected mean of the quality variable, is the initial sample weight, is the KL divergence, is the initial latent variable at time 0, is the initial neural network parameter, is the standard Gaussian distribution, is the latent variable at time t, is the latent variable state increment at time t - 1, is the neural network parameter at time t, is the identity matrix, and is a non-linear function approximated by a neural network.

7. The RVAE-based ternary cathode material preparation process monitoring system of claim 4, wherein, In the comparison module (50), The calculated statistics of the test sample are: wherein is a statistical quantity; is an input variable; is a number of principal components of singular values obtained by singular value decomposition; is a number of principal components thereof; is a principal component space spanned by the corresponding eigenvectors; is a transpose of the principal component space, is a transpose of the input variable, is an inverse matrix of an inverse diagonal matrix composed of the singular values; By comparing the statistics of the process data with the control limit in the online calculation process to determine whether a fault occurs, the judgment logic is: wherein, is the statistical amount, is the control limit.

Images