An Adaptive and Explainable Monitoring Method for Cascade Metallurgical Processes

Abstract

The present invention relates to the technical field of cascade metallurgical process monitoring, and specifically discloses an adaptive and interpretable monitoring method for cascade metallurgical processes. By constructing a hybrid model including an incremental learning trigger based on the BSOM model and a trained VGATEPi model, and constructing an anomaly-aware localization mechanism for fault localization and explanation, continuous learning is effectively carried out from dynamically changing cascade metallurgical process variables to achieve interpretable monitoring results. The method of the present invention can significantly improve the process monitoring performance, especially in reducing the false alarm rate in process monitoring, and solves the problems of traditional monitoring methods based on static graph models, which are prone to false alarms and missed alarms when monitoring non-stationary data because they do not automatically update according to time-varying characteristics. In addition, the present invention can provide interpretable causal relationships between fault variables, providing important maintenance basis for abnormal processes.

Description

Technical Field

[0001] This application relates to the technical field of cascade metallurgical process monitoring, and specifically discloses an adaptive and interpretable monitoring method for cascade metallurgical processes. Background Art

[0002] With the continuous progress of equipment automation and integration in non-ferrous metallurgy factories, more efficient and reliable online process monitoring methods can reduce maintenance costs and ensure industrial stability. Through online process detection, the occurrence of faults can be monitored in a timely manner and the fault location can be accurately traced, which can ensure the continuous stability of product quality, optimize the production process, and maximize production efficiency and profitability.

[0003] Existing non-ferrous metallurgical process monitoring methods mainly include model-based methods, knowledge-based methods, and data-driven methods. For data-driven cascade metallurgical process monitoring methods, they include multivariate statistical techniques and advanced artificial intelligence methods. Among them, multivariate statistical monitoring methods use the correlation between variables for monitoring.

[0004] In the prior art, there is a process anomaly detection method based on a graph-guided masked autoencoder. This method enhances the accuracy and transparency of anomaly detection by introducing a graph structure to model the topological relationship between variables. By adopting a masking and patching mechanism to reduce data redundancy, the model is forced to extract the most valuable intrinsic information from the data, and at the same time, the KL divergence loss is used to ensure the consistency of the input and output data distributions. This method can capture the time dependence in process variables and the complex relationships between variables, thereby effectively identifying and locating anomalies.

[0005] However, due to the fluctuations in the feed conditions of modern cascade metallurgical processes, they usually exhibit time-varying characteristics. The model built by the graph structure in the above-mentioned graph autoencoder-based process anomaly detection method belongs to a static graph model, which is a time-invariant monitoring method and will not be automatically updated after deployment, resulting in a gradual decline in monitoring performance. During the production process, fluctuations in raw materials and product grade conversion will affect the equipment performance and service life, and the operating conditions may change due to changes in internal mechanisms or external environmental factors, thereby causing the entire system to change over time. This change will bring new monitoring patterns, and the data related to these new patterns is often non-stationary. In this case, the monitoring method based on the static graph model is prone to false alarms and missed alarms.

[0006] The present invention provides an adaptive and interpretable monitoring method for cascade metallurgical processes to solve the above problems. Summary of the Invention

[0007] The object of the present invention is to solve the problem that the traditional monitoring method based on the static graph model is prone to false alarms and missed alarms when monitoring non-stationary data because it does not automatically update according to the time-varying characteristics.

[0008] To achieve the above object, the basic solution of the present invention provides a cascaded metallurgical process adaptive and interpretable monitoring method, including the following steps:

[0009] Step A1: Construct a BSOM model based on the Bayesian self-organizing map network and a VGATEPi model with a variational graph attention autoencoder combined with probability inference;

[0010] Step A2: Collect the historical normal operation data, prior information, and mutual information of the cascaded metallurgical process, and process the prior information and mutual information to obtain graph structure data;

[0011] Step A3: Input the historical normal operation data into the BSOM model for training, use the trained BSOM model as an incremental learning trigger, and input the graph structure data into the VGATEPi model for training to generate an adaptive threshold that conforms to the input data;

[0012] Use the incremental learning trigger and the trained VGATEPi model as a hybrid model, and construct an anomaly-aware localization mechanism for fault localization and interpretation;

[0013] Step A4: Collect the real-time data of the cascaded metallurgical process and input it into the hybrid model. When the input data triggers the incremental learning trigger, calculate and compare the difference with the adaptive threshold by the VGATEPi model updated based on the input data. When it exceeds the adaptive threshold, it is judged as an abnormal state, otherwise it is a normal state;

[0014] When the input data does not trigger the incremental learning trigger, calculate and compare the difference with the adaptive threshold by the non-updated VGATEPi model. When it exceeds the adaptive threshold, it is judged as an abnormal state, otherwise it is a normal state;

[0015] When it is judged as an abnormal state, the anomaly-aware localization mechanism cooperates with the prior information to perform an interpretable analysis of the abnormal state.

[0016] Further, in step A2, the graph structure data takes variables as nodes, including a prior information matrix expressed by an adjacency matrix, an information matrix expressed by an adjacency matrix, and a spatial structure combining the prior information matrix and the information matrix.

[0017] Further, in step A3, when inputting the historical normal operation data into the BSOM model, the steps of training with the historical normal operation data as input samples are as follows:

[0018] Step B1: Randomly initialize the weights of all neurons;

[0019] Step B2: Use historical normal operation data as input data, calculate the best-matching neuron with the input data using Euclidean distance, calculate the similarity between each neuron and the input data through Euclidean distance, and select the best-matching unit among all neurons to determine the winner neuron;

[0020] Step B3: Weight update based on Bayesian regularization: After selecting the best-matching unit among the found neurons, update the weights of all neurons within the neighborhood of the winner neuron, aligning the spatial distribution of the connection weight vector with the probability distribution of the input data;

[0021] Step B4: Generate an incremental trigger strategy using the quantization error as the incremental decision criterion, where the quantization error is a mapping of the data distribution of the input data.

[0022] Furthermore, in Step B4, the expression of the generated incremental trigger strategy is as follows:

[0023] ;

[0024] In the formula, e q is the quantization error of the real-time data, is the average value of the quantization errors of the historical normal operation data;

[0025] In the formula, when , when , ;

[0026] In the formula, when , it indicates triggering incremental learning to promote the update of the BSOM model and the VGATEPi model. When , it indicates that incremental learning is not triggered.

[0027] Furthermore, in Step A3, the training of the VGATEPi model based on graph-structured data includes the following steps:

[0028] Step C1: Receive graph-structured data, encode it based on the graph attention network to obtain graph attention encoding;

[0029] Step C2: Fit the distribution parameters through variational inference and combine the graph attention encoding to construct a latent variable matrix;

[0030] Step C3: Decode to obtain the reconstructed spatial relationship and distribution parameters;

[0031] Step C4: Calculate the reconstruction probability and obtain the parameters of the original variable distribution through the latent variable;

[0032] Step C5: Establish a training loss function;

[0033] Step C6: Generate a decision metric set based on the established training loss function, and estimate the adaptive threshold for cascade reaction process monitoring based on the generated decision metrics.

[0034] Furthermore, in Step C5, the establishment of the training loss function is based on the calculation of the loss of the variational inference process, the loss of spatial structure reconstruction, and the loss of node reconstruction.

[0035] Furthermore, in Step C6, the adaptive threshold f k (T) is adaptively updated based on variational Bayesian inference, and the expression is as follows:

[0036] ;

[0037] ;

[0038] In the formula, is the weighting coefficient, , S k represents the moving average calculated within the sliding window, k = 1, 2, 3, … l, and l is the number of encoder layers of the VGATEPi model;

[0039] In the formula, T k represents the threshold of the input data, represents the mean of the normal distribution , represents the normal distribution variance of, , represents the quantile of the t-distribution, which is used to establish the confidence interval of the threshold.

[0040] Furthermore, in Step A3, the anomaly perception and localization mechanism integrates the attention mechanism and the node reconstruction attribute. The anomaly perception and localization mechanism uses the following formula to aggregate the paths of the edges to calculate the fault score of each variable node:

[0041] ;

[0042] In the formula, n represents the total number of neighbors of node i, represents the attention coefficient between node i and node j, L RECON_N represents the loss of node reconstruction.

[0043] The principle and effect of this solution are as follows:

[0044] Compared with the prior art, the present invention provides an adaptive and interpretable monitoring method for cascade metallurgical processes, aiming to effectively perform continuous learning from dynamically changing cascade metallurgical process variables and achieve interpretable monitoring results. The method of the present invention can significantly improve the process monitoring performance, especially in reducing the false alarm rate in process monitoring. In addition, the present invention can provide interpretable causal relationships between fault variables, solving the problems of false alarms and missed alarms that are prone to occur in the monitoring of non-stationary data by traditional monitoring methods based on static graph models because they do not automatically update according to time-varying characteristics. Brief Description of the Drawings

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1 Shows a flowchart of an adaptive and interpretable monitoring method for cascade metallurgical processes proposed in an embodiment of the present application;

[0047] Figure 2 Shows a schematic diagram of an adaptive and interpretable monitoring method for cascade metallurgical processes proposed in an embodiment of the present application;

[0048] Figure 3 Shows a schematic diagram of the purification process of a smelter proposed in an embodiment of the present application;

[0049] Figure 4 Shows a schematic diagram of the detection results of performance indicators proposed in an embodiment of the present application. Among them, (a) is a schematic diagram of the detection results of the performance indicators of the method PCA_T² when FAR = 51.88% and FDR = 100%, (b) is a schematic diagram of the detection results of the performance indicators of the method PCA_SPE when FAR = 46% and FDR = 100%, (c) is a schematic diagram of the detection results of the performance indicators of the method K-SVD when FAR = 32.77% and FDR = 100%, (d) is a schematic diagram of the detection results of the performance indicators of the method AE when FAR = 38.55% and FDR = 100%, (e) is a schematic diagram of the detection results of the performance indicators of the method VARRP when FAR = 7.66% and FDR = 100%, and (f) is a schematic diagram of the detection results of the performance indicators of the method proposed in the present application when FAR = 0.44% and FDR = 100%;

[0050] Figure 5Shows the explanatory diagram proposed in the embodiments of the present application. Among them, (a1) is the abnormal variable proportion diagram, and (a2) is the schematic diagram of the abnormal variable perception amount;

[0051] Figure 6 Shows the schematic diagram of the variable path propagation relationship for explaining interpretability proposed in the embodiments of the present application;

[0052] Figure 7 Shows the schematic diagram of the variable nodes for explaining interpretability proposed in the embodiments of the present application;

[0053] Figure 8 Shows the schematic diagram of the propagation path of the maximum perceived variable for explaining interpretability proposed in the embodiments of the present application. Detailed implementation manners

[0054] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following combines the drawings and preferred embodiments to detail the specific implementation manners, structures, features and their effects of the present invention as follows.

[0055] A cascaded metallurgical process adaptive and interpretable monitoring method, as shown in the embodiments Figure 1 and Figure 2 shown:

[0056] Includes the following steps:

[0057] Step A1: Construct a BSOM (Bayesian regularized self-organizing map) model based on the Bayesian self-organizing map network and a VGATEPi (Variational graph attention autoencoder with probabilistic inference) model with variational graph attention autoencoder combined with probabilistic inference;

[0058] Step A2: Collect the historical normal operation data, prior information and mutual information of the cascaded metallurgical process, and process the prior information and mutual information to obtain graph structure data;

[0059] Step A3: Input the historical normal operation data into the BSOM model for training, use the trained BSOM model as an incremental learning trigger, and input the graph structure data into the VGATEPi model for training to generate an adaptive threshold that conforms to the input data;

[0060] Use the incremental learning trigger and the trained VGATEPi model as a hybrid model, and construct an abnormal perception and localization mechanism for fault location and explanation;

[0061] Step A4: Collect the real-time data of the cascade metallurgical process and input it into the hybrid model. When the input data triggers the incremental learning trigger, the VGATEPi model updated based on the input data calculates and compares the difference with the adaptive threshold. When the difference exceeds the adaptive threshold, it is judged as an abnormal state; otherwise, it is a normal state.

[0062] When the input data does not trigger the incremental learning trigger, calculate and compare the difference with the adaptive threshold based on the non-updated VGATEPi model. When the difference exceeds the adaptive threshold, it is judged as an abnormal state; otherwise, it is a normal state.

[0063] When it is judged as an abnormal state, the abnormal perception and localization mechanism cooperates with the prior information to perform an interpretable analysis of the abnormal state.

[0064] Specifically, when it is judged as an abnormal state, the abnormal perception and localization mechanism calculates through the graph attention coefficient and the node reconstruction loss, and cooperates with the prior information of the constructed graph structure to perform an interpretable analysis of the abnormal state: the suspicious points of the abnormal occurrence, the propagation relationship, and the propagation path.

[0065] In this embodiment, both the historical normal operation data and the real-time data of the cascade metallurgical process are variables in the cascade metallurgical process.

[0066] In step A2, the graph structure data is constructed based on the prior information and the mutual information, including the information matrix and the prior information matrix expressed by the adjacency matrix, as well as the spatial structure of the combined information matrix and the prior information matrix. In this embodiment, the mutual information is a characterization of the energy-mass flow and information attributes in the cascade metallurgical process. For the collected variables and variable , based on the chain rule of entropy, it is determined that the variable characteristics of variable and variable are:

[0067] ;

[0068] In the formula, and are the entropies of variable and variable respectively, represents the conditional entropy given variable N, represents the conditional entropy given variable M respectively.

[0069] The difference between the entropies of variable and variable is expressed as the mutual information between variable and variable , variable and variable The calculation of the difference between the entropies is as follows:

[0070] ;

[0071] And expand according to the definition of entropy We get:

[0072] ;

[0073] In the formula, and respectively represent the marginal distributions of variables and variable . represents the joint distribution of variables and variable . The relative entropy between the joint distribution and the marginal distributions , , is quantified by the mutual information

[0074] In this embodiment, taking variables as nodes, a threshold ζ is established based on the mean value of mutual information, and an information matrix A expressed by an adjacency matrix is constructed through the threshold ζ ij , and the expression is as follows:

[0075] ;

[0076] Then, a prior information matrix expressed by an adjacency matrix is constructed based on prior information . When there is an association between entities, , otherwise .

[0077] By combining the information matrix A ij and the prior information matrix into a spatial structure, the spatial structure E contains the probability information and prior information of the connections between nodes. The expression of the spatial structure E is as follows:

[0078] ;

[0079] When performing step A3, the historical normal operation data is input into the BSOM model, and the steps of training with the historical normal operation data as input samples are as follows:

[0080] Step B1: Initialization, randomly initialize the weights of all neurons , and set the initial learning rate and the initial neighborhood radius .

[0081] Step B2: Competitive learning. Using the historical normal operation data as the input data, calculate the best-matching neuron with the input data by Euclidean distance, calculate the similarity between each neuron and the input data through Euclidean distance, and select the best-matching unit among all neurons to determine the winning neuron. The expression is:

[0082] ;

[0083] In the formula, is the vector representation of the input data, , N represents the total number of neurons, represents the Euclidean norm.

[0084] Step B3: Weight update based on Bayesian regularization: After selecting the best-matching unit among the neurons, update the weights of all neurons within the neighborhood of the winning neuron, aligning the spatial distribution of the connection weight vector with the probability distribution of the input data.

[0085] In this embodiment, a modified vector objective function is established, and the mean square value of the neuron weight vector is added to the modified vector as a complexity penalty term. Then, Bayesian regularization is used to optimize the parameter adjustment of the modified vector objective function to ensure that the modified vector objective function can achieve the best weight update effect. The update rule of the modified vector objective function is as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, represents the connection weight vector of neuron j at iteration t, j = 1, 2, 3,..., t = 1, 2, 3,..., is the learning rate, which decreases as t increases to ensure network convergence, represents the neighborhood function, using a Gaussian function to ensure that each input data only activates one neuron, and the neighborhood function reaches its maximum value at the best-matching unit selected among the neurons, represents the position of the best-matching unit selected among the neurons, r i is the position of the i-th neuron in the output layer, i = 1, 2, 3,..., is the neighborhood radius, represents the modified vector, The expression of

[0090] ;

[0091] In the formula, Represents the corrected objective function, which combines two regularization parameters and parameter , and determines the distribution of the corrected objective function through parameter and parameter . The expression is:

[0092] ;

[0093] In the formula, represents the correction of the weight of the d-th dimension of neuron i at the iteration number t, and the expression is as follows:

[0094] ;

[0095] represents the mean square value of the weight of neuron i, and the expression is as follows:

[0096] ;

[0097] Then, the corrected vector is used to adjust the neuron connection weights. A smaller value of the corrected objective function indicates a higher similarity between the connection weights and the input data. Bayesian regularization is used to optimize the corrected vector objective function so that can reach its optimal solution.

[0098] In this embodiment, to ensure generality, parameters and parameter are set to follow a Gaussian distribution. According to Bayesian inference, by maximizing the posterior probability, the optimal values of the parameters and parameter of the corrected vector objective function solution can be obtained, and the expressions are as follows:

[0099] ;

[0100] ;

[0101] Step B4: Determine the incremental decision criterion and generate an incremental trigger strategy.

[0102] To evaluate the difference between the real-time data features and the features extracted by the BSOM model, the quantization error e q is used as the incremental decision criterion, and the quantization error is a mapping of the data distribution of the input data.

[0103] Specifically, after the input data is trained by the Bayesian self-organizing mapping network, each input data will be mapped to a neuron unit in the Bayesian self-organizing mapping network, and the Euclidean distance between the input data and its corresponding best matching unit is calculated. For each input data, the quantization error e q-training is the distance between the input data and its best matching unit. During the training phase, the mapping of the data distribution of the input data is used as the quantization error mean , which describes the distance error between the input data set and the corresponding set of best matching units. The mean of the quantization errors of all input data is the quantization error during the training phase.

[0104] Therefore, by comparing the quantization error e q of the subsequent real-time data with the mean of the quantization errors of the historical normal operation data , it can be determined whether the BSOM model understands the measurement data as known data, so as to make an incremental decision, as shown in the following formula:

[0105] ;

[0106] In the formula, when , , when , ;

[0107] In the formula, when , it indicates that incremental learning is triggered, prompting the update of the BSOM model and the VGATEPi model. When , it indicates that incremental learning is not triggered.

[0108] The updated input data consists of real-time data and historical data extracted by dynamic sampling, and the expression is:

[0109] ;

[0110] Among them, represents the concatenation operation, represents the control sampling ratio.

[0111] In step A3, the training of the VGATEPi model based on the graph structure data includes the following steps:

[0112] Step C1: Receive the graph structure data, encode it based on the graph attention network in the VGATEPi model to obtain the graph attention encoding. After the graph data passes through the th layer of encoding, the node i is represented as:

[0113] ;

[0114] Wherein, V i represents the initial input data of node i, and E i represents the set of neighbor nodes of node i, which is obtained through the spatial structure expression, represents the input-to-hidden weight matrix of the encoder hidden layer of the th layer, represents the initial node i before encoding, represents the sigmoid activation function,

[0115] The attention coefficient between node i and node j represents the importance of neighboring nodes to the target node, that is, the attention mechanism. The calculation formula of the attention coefficient between node i and node j is as follows:

[0116] ;

[0117] Wherein, k = 1, 2, 3,... l, represents the attention coefficient between node i and node j in the th layer, and the expression is as follows:

[0118] ;

[0119] Wherein, , , , and , is the trainable parameter of the th layer.

[0120] Step C2: Fit the distribution parameters through variational inference and combine graph attention encoding to construct a latent variable matrix.

[0121] In this embodiment, the fitted distribution parameters are and , and the expressions are as follows:

[0122] ;

[0123] ;

[0124] Wherein, represents the mean vector matrix, represents the variance matrix, and share the training parameter matrix. is obtained through the graph attention encoding of the th layer, is calculated through the spatial structure expression.

[0125] Latent variable vector matrix is generated following a Gaussian distribution, and the expression is as follows:

[0126] ;

[0127] In the formula, is the latent variable vector matrix, and its posterior distribution , is similar to the variational posterior distribution obtained by variational inference. Set as a Gaussian distribution, and its diagonal covariance matrix is shown as follows:

[0128] ;

[0129] ;

[0130] Step C3: Decoding to obtain the reconstructed spatial relationship and distribution parameters, and the expression is as follows:

[0131] ;

[0132] ;

[0133] ;

[0134] In the formula, represents the sigmoid activation function, and share weights. The reconstruction of the above adjacency matrix is generated through the inner product of latent variables and the sigmoid activation function. The decoding process of latent variables is the same as the encoding process to generate the reconstructed parameters and to ensure the consistency of its reconstruction.

[0135] Step C4: Calculate the reconstruction probability to obtain the probability distribution parameters of the original variables through latent variables, so as to evaluate the ability of the model to generate data based on the approximate posterior distribution given the latent variables.

[0136] Specifically, the probability density of the input is defined as:

[0137] ;

[0138] Subsequently, the Monte Carlo method is used for sampling, and the reconstruction probability is obtained as:

[0139] ;

[0140] The reconstruction probability is not only based on the difference between the reconstruction and the original input, but also takes into account the variability of the reconstruction by introducing the variance parameter of the distribution function. This feature enables the model to make selective responses according to the variability of different variables, thus flexibly responding to the different degrees of variability exhibited by different variables.

[0141] Step C5: Establish a training loss function to balance the inference accuracy of the VGATEPi model, effective graph structure reconstruction, and accurate node feature reconstruction through the training loss function.

[0142] The expression of the training loss function is as follows:

[0143] ;

[0144] In the formula, represents the loss of the variational inference process, expressed in the form of a KL divergence term, represents the loss of spatial structure reconstruction, calculated through a binary cross-entropy function, represents the loss of node reconstruction.

[0145] ;

[0146] ;

[0147] ;

[0148] To measure the accuracy of the model in generating node features during the learning process, this loss uses the logarithmic probability density function of the normal distribution to compare the difference between the node features generated by the model and the true features. This integrated loss function enables the entire model to balance various objectives and thus effectively capture the dynamic change characteristics of complex industrial processes during the learning stage.

[0149] Step C6: Calculate the adaptive threshold based on the incremental variational Bayesian inference method.

[0150] Based on the established training loss function, generate a decision index set D = {P1, P2, …, P N}, and estimate the adaptive threshold for cascade reaction process monitoring based on the generated decision indices.

[0151] Based on the incremental variational Bayesian inference method, within the Bayesian framework, predefine the prior distribution p(T), likelihood and prior p(D). By applying Bayes' theorem, the posterior distribution is expressed as follows:

[0152] ;

[0153] In the incremental scenario, the decision metric set D is input sequentially. Therefore, the posterior distribution is specifically expressed as:

[0154] ;

[0155] where D k-1 represents the previously observed metrics, and D k represents the upcoming metrics to be observed.

[0156] In the present invention, the normal distribution is used to approximate the posterior distribution, expressed as the following expression:

[0157] ;

[0158] where and represent the mean and variance of the distribution, respectively.

[0159] By minimizing the and KL divergence, an approximation of can be obtained, and the expression is as follows:

[0160] ;

[0161] where is a constant independent of the parameters. Letting c > 1 and minimizing the KL divergence is transformed into identifying the optimal variational parameters, expressed in the following form:

[0162] ;

[0163] where , representing the log-likelihood of the current data, from the posterior distribution of the previous metric set. In , the posterior of the old set replaces the prior of the new observed metric set. This process reasonably combines the prior and sample data, and can obtain more comprehensive results compared to using only prior information or sample data alone.

[0164] When receiving real-time data, the existing posterior information is used as the prior information for the next calculation, and combined with the newly added sample information to update the posterior information and parameters.

[0165] To adapt to the changes in data distribution and correlation, the present invention adopts a sliding window technique and uses variational Bayesian inference to update the adaptive threshold f k (T). This method dynamically adjusts the adaptive threshold by using the latest real-time data, thereby improving the adaptability, and its expression is as follows:

[0166] ;

[0167] ;

[0168] In the formula, represents the threshold of the input data, represents the quantile of the t-distribution, which is used to establish the confidence interval of the threshold, is the weighting coefficient, , represents the moving average calculated within the sliding window.

[0169] In step A3, the anomaly perception and localization mechanism integrates the attention mechanism and the node reconstruction attribute. The anomaly perception and localization mechanism calculates the failure score of each variable node by aggregating the paths of the edges, so as to determine the faulty variable node. The expression is as follows:

[0170] ;

[0171] In the formula, n represents the total number of neighbors of node i, represents the attention coefficient between node i and node j, represents the loss of node reconstruction.

[0172] In step A4, the real-time data obtained from the real-time measurement of the cascade metallurgical process is input into the incremental learning trigger. When the real-time data meets the trigger condition, the hybrid model is updated, including updating the VGATEPi model and the adaptive threshold;

[0173] When the real-time data does not meet the trigger condition, the hybrid model is not updated, that is, the hybrid model before the current input of real-time data is still used.

[0174] In all cases, the fault detection is based on the current adaptive threshold.

[0175] When an anomaly occurs, the anomaly perception and localization mechanism can be used for causal reasoning of faults and root cause localization, that is, the localization of variable nodes.

[0176] The present invention realizes interpretability and improved feature learning through prior empowerment, adopts a learning paradigm of reconstruction probability, provides an intuitive quantification of "credibility", and performs instance-level self-explanatory causal reasoning based on the inherent transparency of the model. This localization mechanism makes full use of the inherent transparency of the model, does not require an additional surrogate model, reduces the model complexity, and makes it more interpretable and easy to deploy.

[0177] The present invention can adaptively and continuously learn from historical normal operation data, prior information, and mutual information under changing actual conditions and environmental conditions to achieve real-time and accurate monitoring of the system. It can also dynamically adjust the learning strategy according to real-time data to adapt to different operating environments and system state changes, thus providing a promising continuous learning paradigm for real-time adaptive cascade metallurgical process monitoring and overcoming the limitations of traditional methods in dealing with environmental changes.

[0178] Regarding the problem of abnormal identification and root cause analysis in cascade non-ferrous metallurgical processes, the method of the present invention provides a high-level understanding of complex fault situations from the perspective of spatial structure and can identify the root causes of faults. By integrating prior information and probability modeling methods, while ensuring the transparency of the model, it provides accurate and credible instance-level fault attribution explanations, thereby enhancing the interpretability and credibility of the abnormal diagnosis process. The present invention not only improves the accuracy of anomaly detection but also provides effective support for actual anomaly maintenance decisions, helping to reduce the risks of misdiagnosis and missed diagnosis.

[0179] The training process of the present invention only requires normal sample data and can effectively process unlabeled data, thus solving the common problem of unlabeled data in the practical application of cascade non-ferrous metallurgy. In this way, the present invention can still perform effective model training and monitoring in the absence of fault samples, improving the applicability and flexibility of the system in actual operations. This feature enables the framework to achieve fault detection and prediction in scenarios without a large amount of labeled data, providing a solution for the situation of scarce data in cascade non-ferrous metallurgical processes.

[0180] The following is an example of applying an adaptive and interpretable monitoring method for cascade metallurgical processes provided by the present invention to the purification process of a smelter. As Figure 3 shown, the purification process of the smelter is a classic cascade non-ferrous metallurgical process. The main purpose of this process is to remove impurity metal ions in the leaching solution of non-ferrous metal minerals by displacement precipitation to provide high-purity metal electrolyte for subsequent processes. It consists of two major chemical reaction-related processes: the copper removal process and the cobalt removal process. Figure 3 In the figure, B is a silo, R is a reactor, W is a conveyor, H is a heat exchanger, F represents the feed condition variable, TH is a thickener, ST is a storage tank, PF is a pressure filter, V is an operating variable, and Q is an outlet variable. Usually, factors such as fluctuations in the physical and chemical properties of the feed and external disturbances will cause the problem of time-varying dynamic characteristics of the process and will also have associated effects.

[0181] In this example, 16 variables were measured from the factory, and their specific physical meanings are shown in Table 1:

[0182] Table 1

[0183]

[0184] In this embodiment, a total of 600 normal state samples are used for training, and 1200 mixed samples are collected as the test data set according to a ratio of 1:2. Among the mixed samples, the first 900 are normal samples and the last 300 are faulty samples. The normal operation data contains various types of time-varying characteristics, such as feed changes and equipment aging, while the faulty data mainly exhibits a step type.

[0185] The present invention verifies the effectiveness of the method by comparing the detection performance of different methods. The comparison methods include the classical multivariate statistical method PCA, the dictionary learning method K-Singular Value Decomposition (K-SVD), the deep learning method Autoencoder (AE), and the combined optimization model Variational Autoencoder with Reconstruction Probability (VAERP). The false alarm rate (FAR) and the fault detection rate (FDR) are used as the performance indicators to detect the performance of the method. The results are as Figure 4 shown. The present invention integrates adaptive and spatial topology learning, achieving excellent performance. The FAR is 0.44% and the FDR is 100%. Compared with VAERP, the FAR is reduced by 94.26%, showing excellent performance in a continuous non-stationary process. This indicates that the proposed method can continuously learn and adapt to new patterns.

[0186] The present invention also illustrates the interpretability through visual and quantitative evaluations, as Figure 5 、 Figure 6 、 Figure 7 and Figure 8 shown.

[0187] From Figure 5 (a1) and Figure 5 (a2), it can be seen that the total inlet flow rate F11 of the copper removal process has the highest perception value and proportion, followed by the redox potential V11 of the copper removal process in the reactor R11. Next are the total inlet flow rate F21 of the cobalt removal process and the zinc powder addition amount V14 of the copper removal process. This indicates that there are potential abnormal positions in both interconnected processes. In addition, Figure 6 it shows that the transmission relationship related to the total inlet flow rate of the copper removal process is complex and has a wide impact on other variables. Based on the relationship path, the present invention visualizes the first four key variables and their transmission relationships according to the ratio of their LC values, as Figure 8As shown. It can be inferred that the total inlet flow rate F11 of the copper removal process is the main cause of the failure. This transmission relationship is due to the abnormal total flow rate in the copper removal process, resulting in abnormal oxidation-reduction potential in the copper removal process. Additionally, the operation variables are not adjusted in a timely manner, leading to a decline in the overall performance of the copper removal process. Since this is a continuous process, the abnormal flow rate in the copper removal process also causes the abnormal total flow rate in the cobalt removal process, thereby affecting the operating conditions of cobalt removal, such as abnormal oxidation-reduction potential, and ultimately having a negative impact on the cobalt removal performance. These observations emphasize the ability of the present invention to consider the relationships between all variables from a global perspective. These observations emphasize that the method of the present invention can consider the relationships between variables from a global perspective, demonstrating its ability to handle indirect abnormal propagation relationships in complex industrial processes, especially in cases involving common influencing factors and indirect transmission paths.

[0188] To better evaluate the quality of the explanations of the hybrid model, the present invention uses the Fidelity (the larger the better) and Stability metrics (the smaller the better) for quantitative evaluation. For fair comparison, all evaluations are carried out under the condition of a sparsity of 0.75, and the experimental results are shown in Table 2 in the appendix:

[0189] Table 2

[0190]

[0191] The method proposed by the present invention obtains the highest Fidelity score, indicating its ability to accurately identify the key features affecting the system state, thereby enhancing the operator's trust in the monitoring model. In addition, its stability value is 0.0034, indicating that the explanation results have a high degree of consistency. This confirms that the method proposed by the present invention is superior to other methods in terms of both explanation effectiveness and explanation stability.

[0192] The above are only the preferred embodiments of the present invention and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content within the scope of the technical solution of the present invention to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. An adaptive and interpretable monitoring method for cascade metallurgical processes, characterized in that, It includes the following steps: Step A1: Construct a BSOM model based on a Bayesian self-organizing map network and a VGATEPi model with a variational graph attention autoencoder with combined probability inference; Step A2: Collect historical normal operation data, prior information, and mutual information of the cascaded metallurgical process, and process the prior information and mutual information to obtain graph structure data; Step A3: Input the historical normal operation data into the BSOM model for training, use the trained BSOM model as an incremental learning trigger, and input the graph structure data into the VGATEPi model for training to generate an adaptive threshold that conforms to the input data; Use the incremental learning trigger and the trained VGATEPi model as a hybrid model, and construct an anomaly-aware localization mechanism for fault localization and explanation; Step A4: Collect real-time data of the cascaded metallurgical process and input it into the hybrid model. When the input data triggers the incremental learning trigger, calculate and compare the difference from the adaptive threshold by the VGATEPi model updated based on the input data. When it exceeds the adaptive threshold, it is judged as an abnormal state, otherwise it is a normal state; When the input data does not trigger the incremental learning trigger, calculate and compare the difference from the adaptive threshold based on the non-updated VGATEPi model. When it exceeds the adaptive threshold, it is judged as an abnormal state, otherwise it is a normal state; When it is judged as an abnormal state, the anomaly-aware localization mechanism cooperates with the prior information to perform an interpretable analysis of the abnormal state.

2. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 1, wherein In step A2, the graph structure data takes variables as nodes, including a prior information matrix expressed by an adjacency matrix, an information matrix expressed by an adjacency matrix, and a spatial structure combining the prior information matrix and the information matrix.

3. The self-adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 1, characterized in that, In step A3, when inputting the historical normal operation data into the BSOM model, the steps of training with the historical normal operation data as input samples are as follows: Step B1: Randomly initialize the weights of all neurons; Step B2: Use the historical normal operation data as input data, calculate the best-matching neuron with the input data using the Euclidean distance, calculate the similarity between each neuron and the input data through the Euclidean distance, and select the best-matching unit among all neurons to determine the winner neuron; Step B3: Weight update based on Bayesian regularization: After selecting the best-matching unit among the found neurons, update the weights of all neurons in the neighborhood of the winner neuron, and align the spatial distribution of the connection weight vector with the probability distribution of the input data; Step B4: Use the quantization error as an incremental decision criterion to generate an incremental trigger strategy, where the quantization error is a mapping of the data distribution of the input data.

4. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 3, characterized in that In step B4, the expression of the generated incremental trigger strategy is as follows: ； where e q is the quantization error of the real-time data, and is the mean value of the quantization errors of the historical normal operation data; where, when then , when then ; Wherein, when , it indicates that incremental learning is triggered to promote the update of the BSOM model and the VGATEPi model. When , it indicates that incremental learning is not triggered.

5. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 2, characterized in that In step A3, the training of the VGATEPi model based on the graph structure data includes the following steps: Step C1: Receive the graph structure data, encode it based on the graph attention network to obtain graph attention encoding; Step C2: Fit the distribution parameters through variational inference and combine the graph attention encoding to construct a latent variable matrix; Step C3: Decode to obtain the reconstructed spatial relationship and distribution parameters; Step C4: Calculate the rework probability and obtain the parameters of the original variable distribution through latent variables; Step C5: Establish a training loss function; Step C6: Generate a decision metric set based on the established training loss function, and estimate the adaptive threshold for cascade reaction process monitoring based on the generated decision metrics.

6. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 5, characterized in that, In Step C5, the establishment of the training loss function is based on the calculation of the loss of the variational inference process, the loss of spatial structure reconstruction, and the loss of node reconstruction.

7. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 5, wherein In step C6, the adaptive threshold f k (T) is adaptively updated based on variational Bayesian inference, and the expression is as follows: ; ; Wherein, is a weighting coefficient, , S k represents the moving average calculated within the sliding window, k = 1, 2, 3, … l, where l is the number of encoder layers of the VGATEPi model; where, T k represents the threshold of the input data, represents the mean of the normal distribution and represents the variance of the normal distribution and , represents the quantile of the t-distribution and is used to establish the confidence interval of the threshold.

8. The adaptive and interpretable monitoring method for a cascaded metallurgical process according to claim 6, characterized in that, In Step A3, the anomaly perception and localization mechanism integrates the attention mechanism and the node reconstruction attribute. The anomaly perception and localization mechanism uses the following formula to aggregate the paths of edges to calculate the fault scores of each variable node: ； where n represents the total number of neighbors of node i, represents the attention coefficient between node i and node j, and L RECON_N represents the loss of node reconstruction.

Images