Self-adaptive and interpretable monitoring method for cascade metallurgy process

Abstract

The invention relates to the technical field of cascade metallurgical process monitoring, and particularly discloses a cascade metallurgical process adaptive and interpretable monitoring method, which comprises the following steps: constructing a hybrid model comprising an incremental learning trigger based on a BSOM model and a trained VGATEPi model, and constructing an anomaly sensing and positioning mechanism for fault positioning and interpretation; according to the method, continuous learning is effectively carried out from dynamically changing cascade metallurgy process variables, interpretable monitoring results are achieved, the process monitoring performance can be remarkably improved, and particularly, in the aspect of reducing the false alarm rate in process monitoring, the problem that a traditional monitoring method based on a static graph model is poor in monitoring effect on non-stable changing data is solved. And the problems of false alarm and missing alarm are easy to occur because the time-varying characteristics cannot be automatically updated at any time. In addition, an interpretable causal relationship between fault variables can be provided, and an important maintenance basis is provided for an abnormal process.

Description

Technical Field

[0001] The present 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 advancement of automation and integration of non-ferrous metallurgical plant equipment, more efficient and reliable online process monitoring methods can reduce maintenance costs and ensure industrial stability. Online process detection can timely monitor the occurrence of faults and accurately track the fault location to ensure continuous and stable product quality, optimize production processes, and maximize production efficiency and profitability.

[0003] Existing nonferrous metallurgical process monitoring methods mainly include model-based methods, knowledge-based methods and data-driven methods. Data-driven cascade metallurgical process monitoring methods 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 graph-guided masked autoencoders. This method introduces a graph structure to model the topological relationship between variables, thereby enhancing the accuracy and transparency of anomaly detection. By adopting mask and patch mechanisms to reduce data redundancy, the model is forced to extract the most valuable intrinsic information from the data, while ensuring the consistency of input and output data distribution through KL divergence loss. This method can capture the time dependency in process variables and the complex relationship between variables, thereby effectively identifying and locating anomalies.

[0005] However, due to the fluctuations in feeding conditions in modern cascade metallurgical processes, they usually exhibit time-varying characteristics. The model built through the graph structure in the above-mentioned process anomaly detection method based on graph autoencoders belongs to a static graph model, which is a time-invariant monitoring method. It will not be automatically updated after deployment, resulting in a gradual decline in monitoring performance. During the production process, raw material fluctuations and product grade conversions will affect equipment performance and service life, and operating conditions may change due to changes in internal mechanisms or external environmental factors, causing the entire system to change over time. This change will bring about new monitoring modes, and the data associated with these new modes are often non-stationary. In this case, the monitoring method based on the static graph model is prone to false positives and false negatives.

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

[0007] The purpose of the present invention is to solve the problem that the traditional monitoring method based on static graph model is prone to false alarm and missed alarm when monitoring non-stationary changing data because the characteristics cannot be automatically updated at any time.

[0008] In order to achieve the above object, the basic scheme of the present invention provides an adaptive and interpretable monitoring method for a cascade metallurgical process, comprising 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 combined with probabilistic reasoning; Step A2: Collect historical normal operation data, prior information and mutual information of the cascade 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; The incremental learning trigger and the trained VGATEPi model are used as a hybrid model, and an abnormality perception and location mechanism for fault location and explanation is constructed; Step A4: Collect 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 adaptive threshold is exceeded, it is judged as an abnormal state, otherwise it is a normal state. When the input data does not trigger the incremental learning trigger, the difference with the adaptive threshold is calculated and compared based on the non-updated VGATEPi model. If it exceeds the adaptive threshold, it is considered to be an abnormal state, otherwise it is a normal state; When an abnormal state is determined, the abnormal perception and positioning mechanism combines prior information to perform an explainable analysis of the abnormal state.

[0009] Further, in step A2, the graph structure data uses 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.

[0010] Further, in 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: Step B1: Randomly initialize the weights of all neurons; Step B2: Use historical normal operation data as input data, use Euclidean distance to calculate the best matching neuron with the input data, calculate the similarity between each neuron and the input data through Euclidean distance, select the best matching unit among all neurons, and thus determine the winner neuron; Step B3: Weight update based on Bayesian regularization: After selecting the best matching unit among the neurons found, the weights of all neurons in the neighborhood of the winner neuron are updated to align the spatial distribution of the connection weight vector with the probability distribution of the input data; Step B4: Generate an incremental trigger strategy using the quantization error as an incremental decision criterion, wherein the quantization error is a mapping of the data distribution of the input data.

[0011] Further, in step B4, the expression of the generated incremental triggering strategy is as follows: ; In the formula, e q is the quantization error of real-time data, is the mean value of the quantization error of historical normal operation data; In the formula, when hour, ,when hour, ; In the formula, when When , it indicates that incremental learning is triggered, prompting the update of BSOM model and VGATEPi model. , it indicates that incremental learning is not triggered.

[0012] Further, in step A3, the training of the VGATEPi model based on the graph structure data includes the following steps: Step C1: receiving graph structure data, encoding based on the graph attention network, and obtaining graph attention encoding; Step C2: Fit the distribution parameters through variational inference combined with graph attention encoding to construct the latent variable matrix; Step C3: decoding to obtain the reconstructed spatial relationship and distribution parameters; Step C4: Calculate the probability of re-modification and obtain the parameters of the original variable distribution through the latent variables; Step C5: Establish training loss function; Step C6: Generate a decision indicator set based on the established training loss function, and estimate an adaptive threshold for monitoring the cascade reaction process based on the generated decision indicators.

[0013] Further, 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 the spatial structure reconstruction and the loss of the node reconstruction.

[0014] Further, in step C6, the adaptive threshold f k (T) Adaptive update based on variational Bayesian inference, the expression is as follows: ; ; In the formula, is the weighting coefficient, , S k represents the moving mean calculated in the sliding window, k=1,2,3,…l, l is the number of encoder layers of the VGATEPi model; Where, T k represents the threshold of the input data, Represents a normal distribution The mean of Represents a normal distribution The variance of , Represents the quantile of the t-distribution and is used to build confidence intervals for thresholds.

[0015] Further, in step A3, the anomaly-aware localization mechanism integrates the attention mechanism and the node reconstruction attribute, and the anomaly-aware localization mechanism aggregates the edge paths to calculate the fault score of each variable node using the following formula: ; 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.

[0016] The principle and effect of this scheme are: Compared with the prior art, the present invention provides an adaptive and interpretable monitoring method for cascade metallurgical processes, which aims to effectively carry out continuous learning from dynamically changing cascade metallurgical process variables to achieve interpretable monitoring results. The method of the present invention can significantly improve process monitoring performance, especially in reducing the false alarm rate in process monitoring. In addition, the present invention can provide an interpretable causal relationship between fault variables, which solves the problem that traditional monitoring methods based on static graphical models are prone to false alarms and missed alarms when monitoring non-stationarily changing data because the characteristics will not be automatically updated with time-varying characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A flow chart of an adaptive and interpretable monitoring method for a cascade metallurgical process proposed in an embodiment of the present application is shown; Figure 2 A schematic diagram of an adaptive and interpretable monitoring method for a cascade metallurgical process proposed in an embodiment of the present application is shown; Figure 3 A schematic diagram of a smelter purification process proposed in an embodiment of the present application is shown; Figure 4 Schematic diagrams of the performance index detection results proposed in the embodiments of the present application are shown, wherein (a) is a schematic diagram of the performance index detection results of the method PCA_T² when FAR=51.88% and FDR=100%, (b) is a schematic diagram of the performance index detection results of the method PCA_SPE when FAR=46% and FDR=100%, (c) is a schematic diagram of the performance index detection results of the method K-SVD when FAR=32.77% and FDR=100%, (d) is a schematic diagram of the performance index detection results of the method AE when FAR=38.55% and FDR=100%, (e) is a schematic diagram of the performance index detection results of the method VARRP when FAR=7.66% and FDR=100%, and (f) is a schematic diagram of the performance index detection results of the method proposed in the present application when FAR=0.44% and FDR=100%; Figure 5 The interpretability explanation diagram proposed in the embodiment of the present application is shown, wherein (a1) is a diagram of the abnormal variable proportion, and (a2) is a schematic diagram of the abnormal variable perception amount; Figure 6 A schematic diagram of a variable path propagation relationship for explaining interpretability proposed in an embodiment of the present application is shown; Figure 7 A schematic diagram of a variable node for explaining interpretability proposed in an embodiment of the present application is shown; Figure 8 A schematic diagram of a propagation path of a maximum perceptual variable for explaining interpretability proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0019] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0020] An adaptive and interpretable monitoring method for cascade metallurgical processes, implementing e.g. Figure 1 and Figure 2 As shown: The steps include: Step A1: Build a BSOM (Bayesian regularized self-organizing map) model based on a Bayesian self-organizing map network and a VGATEPi (Variational graph attention autoencoder with probabilistic inference) model with a variational graph attention autoencoder combined with probabilistic inference; Step A2: Collect historical normal operation data, prior information and mutual information of the cascade 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; The incremental learning trigger and the trained VGATEPi model are used as a hybrid model, and an abnormality perception and location mechanism for fault location and explanation is constructed; Step A4: Collect 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 adaptive threshold is exceeded, it is judged as an abnormal state, otherwise it is a normal state. When the input data does not trigger the incremental learning trigger, the difference with the adaptive threshold is calculated and compared based on the non-updated VGATEPi model. If it exceeds the adaptive threshold, it is considered to be an abnormal state, otherwise it is a normal state; When an abnormal state is determined, the abnormal perception and positioning mechanism combines prior information to perform an explainable analysis of the abnormal state.

[0021] Specifically, when an abnormal state is judged, the abnormal perception and positioning mechanism calculates the graph attention coefficient and the node reconstruction loss, and cooperates with the prior information of the constructed graph structure to perform an explainable analysis of the abnormal state: the suspicious points where the abnormality occurs, the propagation relationship and the propagation path.

[0022] In this embodiment, the historical normal operation data and real-time data collection of the cascade metallurgical process are variables in the cascade metallurgical process.

[0023] In step A2, graph structure data is constructed based on prior information and mutual information, including information matrix and prior information matrix expressed by adjacency matrix, and spatial structure of combined information matrix and prior information matrix. In this embodiment, mutual information is a characterization of energy-mass flow and information attributes in cascade metallurgical process. and variables , based on the entropy chain rule, we can determine the variables and variables The variable characteristics and are: ; In the formula, and The variables are and variables The entropy of represents the conditional entropy when a variable N is given, They represent the conditional entropy when the variable M is given.

[0024] By variable and variables The difference between the entropy of and variables Mutual information ,variable and variables The difference between the entropies is calculated as follows: ; And according to the definition of entropy We get: ; In the formula, and Respectively represent variables and variables The marginal distribution of Representation variables and variables The joint distribution of Quantified joint distribution With marginal distribution , edge distribution The relative entropy between .

[0025] In this embodiment, the variables are taken as nodes, a threshold ζ is established based on the mutual information mean, and the information matrix A expressed by the adjacency matrix is ​​constructed by the threshold ζ. ij , the expression is as follows: ; Based on the prior information, a prior information matrix expressed by the adjacency matrix is ​​constructed. , when there is an association between entities, ,otherwise .

[0026] By combining the information matrix A ij and the prior information matrix It becomes a spatial structure. The spatial structure E contains the probability information and prior information of the connection between nodes. The expression of the spatial structure E is as follows: ; When performing step A3, the historical normal operation data is input into the BSOM model. The steps of training using the historical normal operation data as input samples are as follows: Step B1: Initialization, randomly initialize the weights of all neurons , and set the initial learning rate and the initial neighborhood radius .

[0027] Step B2: Competitive learning, using historical normal operation data as input data, using Euclidean distance to calculate the best matching neuron with the input data, calculating the similarity between each neuron and the input data through Euclidean distance, selecting the best matching unit among all neurons, and thus determining the winner neuron. The expression is: ; In the formula, is the vector representation of the input data, , N represents the total number of neurons, represents the Euclidean norm.

[0028] Step B3: Weight update based on Bayesian regularization: After selecting the best matching unit among the neurons found, the weights of all neurons in the neighborhood of the winner neuron are updated to align the spatial distribution of the connection weight vector with the probability distribution of the input data.

[0029] In this embodiment, a correction vector objective function is established, and the mean square value of the neuron weight vector is added to the correction vector as a complexity penalty term. Then, the parameter adjustment of the correction vector objective function is optimized by Bayesian regularization to ensure that the correction vector objective function can achieve the best weight update effect. The update rule of the correction vector objective function is as follows: ; ; ; In the formula, represents the connection weight vector of neuron j at iteration number t, j=1,2,3,…, t=1,2,3,…, is the learning rate, which decreases as t increases to ensure the network converges, Represents the neighborhood function, using the 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 in the neuron. represents the position of the best matching unit selected in the neuron, r i is the position of the i-th neuron in the output layer, i=1,2,3,…, is the neighborhood radius, represents the correction vector, The expression is as follows: ; In the formula, express The modified objective function combines two regularization parameters and parameters , through the parameters and parameters Determine the distribution of the modified objective function, the expression is: ; In the formula, It represents the modification of the d-th dimension weight by neuron i at iteration number t, and the expression is as follows: ; Represents the mean square value of the weight of neuron i, and the expression is as follows: ; Then, using the correction vector To adjust the neuron connection weights, a smaller value of the modified objective function indicates a higher similarity between the connection weights and the input data. Bayesian regularization is used to optimize the modified vector objective function so that can reach its optimal solution.

[0030] In this embodiment, in order to ensure generality, the parameters are set and parameters Follows Gaussian distribution. According to Bayesian reasoning, by maximizing the posterior probability, the parameters of the corrected vector objective function solution can be obtained and parameters The optimal value of is expressed as follows: ; ; Step B4: Determine the incremental decision criteria and generate the incremental trigger strategy.

[0031] In order to evaluate the difference between the real-time data features and the features extracted by the BSOM model, the quantization error e q As an incremental decision criterion, the quantization error is a mapping of the data distribution of the input data.

[0032] Specifically, after the input data is trained by the Bayesian self-organizing map network, each input data will be mapped to a neuron unit in the Bayesian self-organizing map network, and the Euclidean distance between the input data and its corresponding best matching unit will be calculated. For each input data, the quantization error e q-training It 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 mean quantization error. , which describes the distance error between the input data set and the corresponding best matching unit set, the mean of the quantization errors of all input data It is the quantization error in the training phase.

[0033] Therefore, by comparing the quantization error e of subsequent real-time data q The mean value of the quantitative error compared with the historical normal operation data , it can be determined whether the BSOM model understands the measured data as known data and thus makes incremental decisions, as shown in the following formula: ; In the formula, when hour, ,when hour, ; In the formula, when When , it indicates that incremental learning is triggered, prompting the update of BSOM model and VGATEPi model. , it indicates that incremental learning is not triggered.

[0034] The updated input data consists of real-time data and historical data extracted through dynamic sampling, and the expression is: ; in, Represents a splicing operation, Indicates the control sampling ratio.

[0035] In step A3, the training of the VGATEPi model based on the graph structure data includes the following steps: Step C1: Receive graph structure data, encode it based on the graph attention network in the VGATEPi model, and obtain the graph attention encoding. After layer encoding, node i is represented as: ; Where V i represents the initial input data of node i, E i represents the set of neighbor nodes of node i, obtained through the spatial structure expression, Indicates Layer encoder hidden layer input to hidden weight matrix, represents the initial node i before encoding, represents the sigmoid activation function, It represents the attention coefficient between node i and node j, indicating the importance of neighboring nodes to the target node, that is, the attention mechanism.

[0036] The attention coefficient between node i and node j The calculation formula is as follows: ; Where k = 1, 2, 3, ... l, Indicates that node i and node j are in The attention coefficient of the layer is expressed as follows: ; In the formula, , , ,and , It is Trainable parameters of the layer.

[0037] Step C2: Fit the distribution parameters through variational inference combined with graph attention encoding to construct the latent variable matrix.

[0038] In this embodiment, the fitted distribution parameters are and , the expression is as follows: ; ; In the formula, represents the mean vector matrix, represents the variance matrix, and Shared training parameter matrix. Through the The graph attention encoding of the layer is obtained. Obtained through calculation of spatial structure expression.

[0039] Hidden variable vector matrix The generation follows a Gaussian distribution, expressed as follows: ; In the formula, is a latent variable vector matrix, whose posterior distribution , and the variational posterior distribution obtained by variational inference Similar, will Assume that it is a Gaussian distribution, and its diagonal covariance matrix is ​​as follows: ; ; Step C3: Decoding, obtaining the reconstructed spatial relationship and distribution parameters, the expression is as follows: ; ; In the formula, represents the sigmoid activation function, and Shared weights, the above adjacency matrix The reconstruction is generated by 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 , ensuring the consistency of its reconstruction.

[0040] Step C4: Calculate the reconstruction probability and obtain the probability distribution parameters of the original variables through the latent variables to evaluate the model's ability to generate data based on the approximate posterior distribution given the latent variables.

[0041] Specifically, the probability density of the input is defined as: ; Then the Monte Carlo method is used for sampling, and the reconstruction probability is obtained as: ; 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 selectively respond to the variability of different variables, thereby flexibly responding to different degrees of variability exhibited by different variables.

[0042] Step C5: Establish a training loss function to balance the reasoning accuracy, effective graph structure reconstruction, and accurate node feature reconstruction of the VGATEPi model.

[0043] The expression of the training loss function is as follows: ; In the formula, Represents the loss of the variational inference process, expressed in the form of KL divergence term, Represents the loss of spatial structure reconstruction, calculated by the binary cross entropy function, Represents the loss of node reconstruction.

[0044] ; ; ; To measure the accuracy of the node features generated by the model during the learning process, the 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 real features. This integrated loss function enables the entire model to balance various objectives and effectively capture the dynamic changes of complex industrial processes during the learning stage.

[0045] Step C6: Calculate the adaptive threshold based on the incremental variational Bayesian inference method.

[0046] Based on the established training loss function, a decision indicator set D={P1,P2,…,P N}, an adaptive threshold for cascade reaction process monitoring is estimated based on the generated decision indicators.

[0047] Based on the incremental variational Bayesian inference method, within the Bayesian framework, the prior distribution p(T) and likelihood and the prior p(D). By applying Bayes’ theorem, the posterior distribution is expressed as follows: ; In the incremental scenario, the decision indicator set D is input sequentially, so the posterior distribution is specifically expressed as: ; Where D k-1 represents the previously observed indicator, D k Represents an upcoming metric that will be observed.

[0048] In the present invention, the normal distribution is adopted To approximate the posterior distribution, it is expressed as the following expression: ; In the formula, and represent the mean and variance of the distribution respectively.

[0049] By minimizing and KL divergence can be obtained for The approximate expression is as follows: ; In the formula, is a constant that is independent of the parameters. Minimizing the KL divergence by c>1 is transformed into identifying the optimal variational parameter, which can be expressed as follows: ; In the formula, , represents the log-likelihood of the current data, from the posterior distribution of the previous indicator set. In the example, the posterior of the old set Substitute the priors for the set of new observations. This process combines the priors and sample data in a reasonable way, and can produce more comprehensive results than using either prior information or sample data alone.

[0050] 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.

[0051] In order to adapt to the changes in data distribution and correlation, the present invention adopts sliding window technology and uses variational Bayesian reasoning to perform adaptive threshold f k (T) Update. This method improves adaptability by dynamically adjusting the adaptive threshold using the latest real-time data, and its expression is as follows: ; ; 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 mean calculated within the sliding window.

[0052] In step A3, the anomaly-aware localization mechanism integrates the attention mechanism and the node reconstruction attribute. The anomaly-aware localization mechanism calculates the fault score of each variable node by aggregating the edge paths, thereby determining the faulty variable node. The expression is as follows: ; 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.

[0053] In step A4, the real-time data obtained by 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. When the real-time data does not meet the trigger condition, the hybrid model is not updated, that is, the hybrid model before the real-time data is input is still used.

[0054] In all cases, fault detection is based on the current adaptive threshold.

[0055] When an abnormality occurs, the abnormality perception and positioning mechanism can be used to perform causal reasoning and root cause positioning of the fault, that is, the positioning of the variable node.

[0056] The present invention achieves interpretability and improved feature learning through prior empowerment, adopts a learning paradigm of reconstruction probability, provides intuitive quantification of "credibility", and performs instance-level self-explanatory causal reasoning based on the inherent transparency of the model. The positioning mechanism fully utilizes the inherent transparency of the model, does not require additional proxy models, reduces model complexity, and makes it more interpretable and easy to deploy.

[0057] The present invention can adaptively and continuously learn from historical normal operation data, prior information and mutual information under the condition of actual state and environmental conditions change, so as to realize real-time and accurate monitoring of the system. And it can dynamically adjust the learning strategy according to the 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, overcoming the limitations of traditional methods in dealing with environmental changes.

[0058] Aiming at the problem of abnormality identification and root cause analysis in cascade nonferrous metallurgical processes, the method of the present invention provides a high-level understanding of complex fault conditions from the perspective of spatial structure, and can identify the root cause of the fault. By integrating prior information and probabilistic modeling methods, while ensuring model transparency, it provides accurate and reliable 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 abnormal maintenance decisions, helping to reduce the risk of misdiagnosis and missed diagnosis.

[0059] The training process of the present invention only requires normal sample data and can effectively process unlabeled data, thereby solving the common unlabeled data problem in the practical application of cascade nonferrous metallurgy. In this way, the present invention can still perform effective model training and monitoring in the absence of fault samples, thereby improving the applicability and flexibility of the system in actual operation. This feature enables the framework to achieve fault detection and prediction in scenarios without a large amount of labeled data, providing a solution to the situation of data scarcity in the cascade nonferrous metallurgy process.

[0060] The following is an example of the application of an adaptive and interpretable monitoring method for a cascade metallurgical process provided by the present invention in a smelter purification process. Figure 3As shown, the smelter purification process is a classic cascade nonferrous metallurgical process. The main purpose of this process is to remove impure metal ions in the nonferrous metal mineral leachate by displacement precipitation to provide high-purity metal electrolyte for subsequent processes. It consists of two chemical reaction-related processes: copper removal process and cobalt removal process. Figure 3 B is the silo, R is the reactor, W is the conveyor, H is the heat exchanger, F represents the feed condition variable, TH is the thickener, ST is the storage tank, PF is the pressure filter, V is the operating variable, and Q is the outlet variable. Usually, factors such as fluctuations in the physical and chemical properties of the feed and external disturbances will lead to time-varying problems in the dynamic characteristics of the process and will have associated effects.

[0061] In this example, 16 variables were measured from the factory, and their specific physical meanings are shown in Table 1: Table 1

[0062] In this embodiment, a total of 600 normal state samples are used for training, and 1200 mixed samples are collected as test data sets at a ratio of 1:2. In the mixed samples, the first 900 are normal samples and the last 300 are fault samples. Normal operation data contains various types of time-varying characteristics, such as feed changes and equipment aging, while fault data mainly exhibits step types.

[0063] The effectiveness of the method is verified by comparing the detection performance of different methods. The comparison methods include the classic 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 of the evaluation method. The results are as follows Figure 4 As shown, the proposed method integrates adaptation and spatial topology learning and achieves excellent performance with a FAR of 0.44% and a FDR of 100%. Compared with VAERP, the FAR is reduced by 94.26% and performs well in continuous non-stationary processes, which shows that the proposed method is able to continuously learn and adapt to new patterns.

[0064] The present invention also illustrates interpretability through visual and quantitative evaluation, such as Figure 5 , Figure 6 , Figure 7 and Figure 8 shown.

[0065] 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 ratio, followed by the redox potential V11 of the copper removal process of reactor R11. Next is the total inlet flow rate F21 of the cobalt removal process and the amount of zinc powder added V14 of the copper removal process. This shows that there are potential abnormal locations in both interrelated processes. In addition, Figure 6 It shows that the transmission relationship related to the total inlet flow of the copper removal process is very complex and has a wide range of effects on other variables. Based on the relationship path, the present invention visualizes the first four key variables and their transmission relationships according to their LC value ratios, such as Figure 8 As shown. It can be inferred that the total inlet flow rate F11 of the copper removal process is the main cause of the fault. This transfer relationship is due to the abnormal total flow rate of the copper removal process, which leads to abnormal redox potential of the copper removal process. In addition, the operating variables are not adjusted in time, resulting in a decrease in the overall performance of the copper removal process. Since this is a continuous process, the abnormal flow rate of the copper removal process also leads to abnormal total flow rate of the cobalt removal process, which in turn affects the operating conditions of cobalt removal, such as abnormal redox potential, and ultimately has a negative impact on the cobalt removal performance. These observations emphasize the ability of the present invention to consider the relationship between all variables from a global perspective. These observations emphasize that the method of the present invention can consider the relationship between variables from a global perspective, demonstrating its ability to handle indirect abnormal propagation relationships in complex industrial processes, especially those involving common influencing factors and indirect transfer paths.

[0066] In order to better evaluate the quality of the hybrid model explanation, this paper uses the fidelity (the larger the better) and stability indicators (the smaller the better) for quantitative evaluation. For fair comparison, all evaluations are performed under the condition of sparsity of 0.75. The experimental results are shown in Appendix 2: Table 2

[0067] The proposed method achieved the highest fidelity score, indicating that it can accurately identify key features that affect the system state, thereby enhancing the operator's trust in the monitoring model. In addition, its stability value is 0.0034, indicating that the interpretation results are highly consistent. This confirms that the proposed method is superior to other methods in terms of interpretation validity and interpretation stability.

[0068] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modifications to equivalent embodiments of equivalent changes by using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution 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 a cascade metallurgical process, characterized in that: The steps include: Step A1: Construct a BSOM model based on a Bayesian self-organizing map network and a VGATEPi model with a variational graph attention autoencoder combined with probabilistic reasoning; Step A2: Collect historical normal operation data, prior information and mutual information of the cascade 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; The incremental learning trigger and the trained VGATEPi model are used as a hybrid model, and an abnormality perception and location mechanism for fault location and explanation is constructed; Step A4: Collect 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 adaptive threshold is exceeded, it is judged as an abnormal state, otherwise it is a normal state. When the input data does not trigger the incremental learning trigger, the difference with the adaptive threshold is calculated and compared based on the non-updated VGATEPi model. If it exceeds the adaptive threshold, it is considered to be an abnormal state, otherwise it is a normal state; When an abnormal state is determined, the abnormal perception and positioning mechanism combines prior information to perform an explainable analysis of the abnormal state.

2. The method for adaptive and interpretable monitoring of a cascade metallurgical process according to claim 1, characterized in that: In step A2, the graph structure data uses variables as nodes, including a priori information matrix expressed by an adjacency matrix, an information matrix expressed by an adjacency matrix, and a spatial structure combining the priori information matrix and the information matrix.

3. The method for adaptive and interpretable monitoring of a cascade metallurgical process according to claim 1, characterized in that: In step A3, the historical normal operation data is input 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, use the Euclidean distance to calculate the best matching neuron with the input data, calculate the similarity between each neuron and the input data through the Euclidean distance, select the best matching unit among all neurons, and thus determine the winner neuron; Step B3: Weight update based on Bayesian regularization: After selecting the best matching unit among the neurons found, the weights of all neurons in the neighborhood of the winner neuron are updated to align the spatial distribution of the connection weight vector with the probability distribution of the input data; Step B4: Generate an incremental trigger strategy using the quantization error as an incremental decision criterion, wherein the quantization error is a mapping of the data distribution of the input data.

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

5. The method for adaptive and interpretable monitoring of a cascade 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: receiving graph structure data, encoding based on the graph attention network, and obtaining graph attention encoding; Step C2: Fit the distribution parameters through variational inference combined with graph attention encoding to construct the latent variable matrix; Step C3: decoding to obtain the reconstructed spatial relationship and distribution parameters; Step C4: Calculate the probability of re-modification and obtain the parameters of the original variable distribution through the latent variables; Step C5: Establish training loss function; Step C6: Generate a decision indicator set based on the established training loss function, and estimate an adaptive threshold for monitoring the cascade reaction process based on the generated decision indicators.

6. The method for adaptive and interpretable monitoring of a cascade 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 method for adaptive and interpretable monitoring of a cascade metallurgical process according to claim 5, characterized in that: In step C6, the adaptive threshold f k (T) Adaptive update based on variational Bayesian inference, the expression is as follows: ; ; In the formula, is the weighting coefficient, , S k represents the moving mean calculated in the sliding window, k=1,2,3,…l, l is the number of encoder layers of the VGATEPi model; Where, T k represents the threshold of the input data, Represents a normal distribution The mean of Represents a normal distribution The variance of , Represents the quantile of the t-distribution and is used to build confidence intervals for thresholds.

8. The method for adaptive and interpretable monitoring of a cascade metallurgical process according to claim 6, characterized in that: In step A3, the anomaly-aware localization mechanism integrates the attention mechanism and the node reconstruction attribute. The anomaly-aware localization mechanism uses the following formula to aggregate the edge paths to calculate the fault score of each variable node: ； 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.

Images