Zinc leaching multi-condition control method and system based on regression DeepRED rule extraction

Abstract

The present invention relates to the technical field of zinc leaching multi-condition control, and specifically discloses a zinc leaching multi-condition control method and system based on regression DeepRED rule extraction, including step 100, extracting spatiotemporal features and using them for condition classification; step 200, extracting neural network rules based on regression DeepRED; step 300, constructing a data model based on mechanism error compensation; step 400, constructing a pH value control framework based on the data model of rule extraction and mechanism error compensation to control the pH value of an intermediate leaching tank. The present invention solves the problem of diverse operating states of the leaching process and insufficient manual control precision, balances control precision and interpretability, and realizes intelligent control of the pH value of the intermediate leaching tank.

Description

Technical Field

[0001] This invention relates to the field of zinc leaching multi-condition control technology, specifically to a zinc leaching multi-condition control method based on DeepRED regression rule extraction, and also to a zinc leaching multi-condition control system based on DeepRED regression rule extraction. Background Technology

[0002] Pyrometallurgical and hydrometallurgical processes are the two main forms of zinc smelting currently, and nearly 85% of my country's zinc is obtained through hydrometallurgical processes. The hydrometallurgical zinc smelting process consists of batching, roasting, leaching, purification, electrolysis, and casting. Among these, the leaching process is a very important production link, which converts the zinc calcined sand from the upstream roasting process into zinc sulfate solution and removes impurity ions such as iron from the solution. The solution is then sent to the downstream purification process to remove impurity ions such as cobalt and copper.

[0003] In existing technologies, industrial production typically employs five cascaded leaching tanks arranged in a hierarchical manner. Tanks 1 through 3 are the main feeding tanks, aiming to achieve efficient leaching of zinc calcined sand in a relatively acidic solution. Tanks 4 and 5 primarily serve as impurity removal tanks. The introduction of oxygen oxidizes ferrous ions in the solution into ferric ions, which then undergo hydrolysis in the less acidic solution to form solid impurities, thus achieving impurity removal. Tank 3, as the intermediate leaching tank, simultaneously connects the zinc calcined sand dissolution in the pre-leaching stage with the impurity ion removal in the post-leaching stage. The neutral leaching process in hydrometallurgical zinc refining is generally abstracted into an industrial topology of cascaded leaching tanks for research, and current studies primarily utilize the 1-3 tank structure, indicating that this is indeed a conventional zinc leaching structure.

[0004] In neutral leaching, the pH value at the outlet of the intermediate leaching tank is a crucial parameter characterizing the reaction state of the process. It directly determines the effectiveness of the initial leaching and subsequent impurity removal in the neutral leaching process, and the process has strict requirements regarding its range. If the pH value is too high, it indicates that the leaching of zinc roasted sand in the initial stage is inhibited, resulting in a low leaching rate and wasting raw materials. If the pH value is too low, it will lead to Fe in the later stage of the process. 3+ Incomplete hydrolysis and precipitation affects the stable operation of subsequent purification and electrolysis processes.

[0005] Therefore, timely and accurate acquisition of the pH value at the outlet of the intermediate leaching tank and control of it within the ideal range are key to achieving stable and efficient operation of the neutral leaching process.

[0006] Currently, controlling the pH value of intermediate leaching tanks mainly faces the following challenges:

[0007] (1) On-site control is mainly manual. On-site workers manually adjust the waste acid flow rate based on the pH value measured by the physical measurement device. However, the sampling of physical measurement on-site is time-consuming, which makes the real-time performance of manual control poor. In addition, manual control consumes a lot of manpower and mostly relies on the operating experience of on-site workers, resulting in generally low control accuracy and unstable control effect.

[0008] (2) Utilizing various machine learning models, reinforcement learning, and deep learning methods, among other modern advanced tools, for intelligent control of pH in intermediate leaching tanks can ensure control accuracy. However, these methods often suffer from insufficient interpretability. In actual industrial settings, for safety reasons, workers may develop a crisis of trust in models with insufficient interpretability, believing that a lack of interpretability could lead to unpredictable and significant errors. This greatly limits the application of high-accuracy black-box and gray-box models in industrial settings.

[0009] (3) Rules, as a form that is easy for humans to understand, are an important part of the field of interpretability. However, simply applying rules to control problems often leads to insufficient control precision due to the contradiction between interpretability and control precision. Summary of the Invention

[0010] In view of this, the present invention provides a zinc leaching multi-condition control method and system based on DeepRED regression rule extraction, which can solve the problems of diverse operating states and insufficient precision of manual control in the neutral leaching process of hydrometallurgical zinc refining, balance control precision and interpretability, and realize intelligent control of pH value of intermediate leaching tank.

[0011] To achieve the above objectives, the basic solution of this invention provides a multi-condition control method for zinc leaching based on DeepRED regression rules, comprising the following steps:

[0012] Step 100: Extract spatiotemporal features and use them for working condition classification;

[0013] Step 200: Extract neural network rules based on DeepRED regression;

[0014] Step 300: Construct a data model based on mechanism error compensation;

[0015] Step 400: Construct a pH control framework based on a data model of rule extraction and mechanism error compensation to control the pH value of the intermediate leaching tank.

[0016] In one possible design, step 100 specifically includes: dividing the time window, selecting multivariate time series data within a certain period as a single sample for training and testing; taking into account the spatiotemporal characteristics of multivariate time series data, using the Slow Feature Analysis (SFA) method to extract temporal features, and then using the Graph Attention Network (GAT) method to extract spatial features; after extracting the spatiotemporal features, using them together as input to the Gaussian Mixture Model (GMM) for clustering, and using clustering metrics to measure the quality of clustering results under different working conditions, thereby dividing the neutral leaching dataset into different working conditions.

[0017] In one possible design, step 200 specifically involves calculating the mutual information value between the time series of each inlet variable and the time series of the outlet pH value, and determining the control time delay between each inlet variable and the outlet pH value based on the maximum mutual information value. After time-series registration of the data, the continuous input and output data are fuzzified, transforming the continuous input and output data into a fuzzy language subset. Then, the DeepRED regression rule extraction method is used to extract rules in the form of "IF...... THEN......" from the trained neural network. The antecedent of the rule is the amount of roasted sand added to the neutral leaching reaction tank, the flow rate of the mixed liquor, the acid leaching supernatant at the control time delay, and the outlet pH value at the current time. The consequent is the waste acid flow rate at the control time delay of tanks #1 and #2.

[0018] In one possible design, step 300 includes controlling the pH value of the intermediate leaching tank by using the waste acid flow rates of tanks 1 and 2 as controlled variables during the neutral leaching process, and using the deviation between the mechanistic model and the actual model as the learning target of the high-precision data model to construct a data model based on mechanistic error compensation. This data model is a model of all inlet variables with respect to the pH value, and all inlet variables include two controlled variables.

[0019] In one possible design, step 400 includes obtaining the controlled variables, namely the flow rates of waste acid in tank #1 and waste acid in tank #2, through rule extraction, and then sending the values ​​of the two controlled variables as inputs to the mechanism error compensation data model to obtain the predicted value of the mechanism error compensation data model.

[0020] In one possible design, in step 400, the mechanistic model is as follows:

[0021] For slot #1:

[0022] ;

[0023] For slot #2:

[0024] ;

[0025] For slot #3:

[0026] ;

[0027] Assuming the liquid flow rates at the inlet and outlet remain constant, then:

[0028] F 1#out =F 1#acid +F 1#mixed +F 1#thermal；

[0029] F 2#out =F 1#out +F 2#acid；

[0030] F 3#out =F 2#out；

[0031] Among them, c 1# c 2# c 3# F represents the outlet hydrogen ion concentration of tanks #1, #2, and #3, respectively. 1#acid F 1#mixed F 1#thermal These represent the flow rates of waste acid, mixed solution, and acid leaching supernatant added to tank #1, respectively. 1#acid c 1#mixed c 1#thermal These represent the concentrations of the waste acid, the mixed solution, and the acid leaching supernatant added to tank #1, respectively. 2#acid c 2#mixed c 2#thermal They represent joining respectively

[0032] The concentrations of waste acid, mixed solution, and acid leaching supernatant in tank #2, F 2#acid and c 2#acid These represent the flow rate and concentration of the waste acid added to tank #2, respectively, where k represents the reaction rate constant, and A... 1# A 2# A 3# F represents the effective surface area of ​​the zinc calcinate reaction solid added to tanks #1, #2, and #3, respectively. 1#out F 2#out F 3#out These represent the outlet flow rates of tanks #1, #2, and #3, respectively, where V is the volume of a single reaction tank (generally, the volumes of the three reaction tanks are considered equal). Furthermore, the waste acid added to tanks #1 and #2 originates from waste electrolyte generated in the downstream electrolysis process, thus resulting in c... 1#acid =c 2#acid .

[0033] The present invention also provides a zinc leaching multi-condition control system based on regression DeepRED rule extraction, including a memory, a control processor, and a computer program stored in the memory and executable on the control processor. The control processor executes the program to implement the zinc leaching multi-condition control method based on regression DeepRED rule extraction as described above.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] (1) Considering that the neutral leaching reaction tank has diverse operating states, large parameter fluctuations, and obvious spatiotemporal characteristics, this invention proposes a working condition classification method that integrates spatiotemporal feature extraction. Considering the spatiotemporal characteristics of the neutral leaching process data, the spatiotemporal features are extracted and used for working condition classification, thereby improving the accuracy of working condition classification and enabling better analysis of the operating characteristics inside the leaching reaction tank.

[0036] (2) Considering the balance requirements of control accuracy and interpretability in industrial control, this invention proposes a pH control framework based on a data model of rule extraction and mechanism error compensation. The rule extraction part adopts the DeepRED regression method, which can extract highly interpretable IF-THEN rules from a high-precision neural network. The output of the IF-THEN rules is added to the data model of mechanism error compensation as the controlled variable. The output of the data model of mechanism error compensation is used to feed back the control setpoint, so that the entire control framework can run completely and achieve stable control of pH value in intermediate leaching tank. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This invention relates to a working condition classification method that integrates spatiotemporal feature extraction.

[0039] Figure 2 This invention provides a pH control framework based on a data model using rule extraction and mechanism error compensation.

[0040] Figure 3 This is a schematic diagram of the neutral leaching process upon which the present invention is based;

[0041] Figure 4 A schematic diagram of clustering indices for working condition classification in an application example of the present invention;

[0042] Figure 5 This is a schematic diagram of the rule extraction algorithm in an application example of the present invention;

[0043] Figure 6 A comparison chart of the prediction accuracy of the data model for mechanism error compensation in an application example of the present invention;

[0044] Figure 7 The diagram shows the control effect under multiple operating conditions in an application example of the present invention, where (a), (b), (c), (d), and (e) represent operating conditions 1 to 5, respectively.

[0045] Figure 8 for Figure 1 Schematic diagram of mid-spatial-temporal feature extraction;

[0046] Figure 9 for Figure 2 Schematic diagram of control time delay calculation. Detailed Implementation

[0047] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0048] There are already cases of applying rule extraction to industrial control. Zhang B et al. (Article source: Zhang B, Yang C, Zhu H, et al. Controllable-domain-based fuzzy rule extraction for copper removal process control[J]. IEEE transactions on fuzzy systems, 2017,26(3): 1744-1756) proposed a fuzzy rule extraction strategy based on controllable domain for copper removal processes. First, the definition of representative control samples was introduced. Then, a positive label-free learning algorithm was used to divide the input variable space into several controllable domains. The support vector machine algorithm was used to extract fuzzy control rules for different controllable domains. The control effect proved that the strategy can improve the production qualification rate and process stability. Liang H et al. (Article source: Liang H, Zou J, Zuo K, et al. Animproved genetic algorithm optimization fuzzy controller applied to the wellhead back pressure control system[J]. Mechanical Systems and SignalProcessing, 2020, 142: (106708.) To address the nonlinearity and time-varying nature of throttle valve control in wellhead backpressure control systems, an improved genetic algorithm is proposed to optimize a fuzzy controller. First, a wellhead backpressure control model is established and its transfer function is calculated. Second, an improved genetic algorithm is designed to optimize the highly nonlinear fuzzy control rules between the input and response in a fuzzy PID controller. However, most of these cases only use simple machine learning algorithms for rule extraction, resulting in limited quality of the extracted rules.

[0049] Neural networks possess powerful nonlinear fitting capabilities, excellent generalization abilities, robustness, and parallel computing capabilities. Therefore, extracting rules from neural networks to enhance their interpretability has spurred numerous research efforts by experts and scholars (Article source: Zhang Y, Tiňo P, Leonardis A, et al. A survey on neural network interpretability[J]. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021, 5(5): 726-742). Currently, a mainstream classification method is based on the relationship between the extracted rules and the structure of the trained neural network, dividing neural network rule extraction algorithms into three categories: decomposition methods, teaching methods, and compromise methods. The CRED algorithm proposed by Sato M et al. (Source: Sato M, Tsukimoto H. Rule extraction from neural networks via decision tree induction[C] / / IJCNN'01. International JointConference on Neural Networks. Proceedings (Cat. No. 01CH37222). IEEE, 2001,3: 1870-1875) uses decision trees to decompose neural networks and merges the generated branches through the c / d-rule algorithm to produce different classification granularities. It can also consider the interpretation of continuous and discrete network inputs and outputs. Meanwhile, the CRED algorithm proposed by Zilke J et al. (Source: Zilke J, LozaMencía E, Janssen F. DeepRED—ruleextraction from deep neural networks[C] / / Discovery Science: 19thInternational Conference, DS 2016, Bari, Italy, October 19-21, 2016,Proceedings 19. Springer International Publishing, 2016:) The DeepRED algorithm proposed in 457-473 is a generalization of the CRED algorithm and can be used to extract rules from neural networks with multiple hidden layers; Bologna G (Source: Bologna G).A simple convolutional neural network with rule extraction[J]. Applied Sciences, 2019, 9(12): 2411.) proposes a rule extraction method for CNNs, which extracts rules on fully connected layers of CNNs with the help of Discretized Interpretable Multilayer Perceptron (DIMLP). Chakraborty M et al. (Article source: Chakraborty M, Biswas S, Purkayastha B. Rule extraction from neural network using input data ranges recursively[J]. NewGeneration Computing, 2019, 37: 67-96.) proposed a simple algorithm, ERENNR, for extracting compromise rules from neural networks. It analyzes hidden nodes based on the data range of input attributes relative to their outputs, and analyzes output nodes using the logical combination of the outputs of hidden nodes relative to the output classes. Finally, it generates a rule set by moving backward from the output layer. However, most neural network rule extraction methods are currently applied to classification, while research on industrial control is relatively scarce.

[0050] For Zhang Y's solution, the complexity of industrial sites makes it impossible to guarantee the quality of industrial data, which may contain some unreasonable data. This unreasonable data can affect the quality of the extracted rules. Therefore, before classifying the samples, unreasonable samples with similar control conditions to excellent samples but poor control effects must first be removed. Secondly, manual labeling is difficult and time-consuming, and many samples in industrial sites cannot be identified with corresponding labels, resulting in a limited number of labels needed for classification. Therefore, a positive unlabeled learning method is used to identify reliable positive and negative samples from the unlabeled dataset, and then the dataset is divided into four controllable domains. Finally, for each controllable domain, a multi-class SVM method is used to extract fuzzy rules, and the output of the fuzzy rules, i.e., support vectors, is converted into the values ​​of the controlled variables and applied to the final copper removal process control.

[0051] Zilke J's approach provides a neural network rule extraction algorithm called DeepRED, which effectively extracts rules from deep neural networks with multiple hidden layers. First, a neural network with multiple hidden layers is trained for classification problems. Then, for each class, the activation output value corresponding to the input of each hidden layer is represented in one-hot encoding. A decision tree is constructed between every two adjacent network layers, and the branches from the root node to the leaf node of each decision tree are transformed into rules. Intermediate variables are then eliminated, thus obtaining the rules from the neural network input to the output.

[0052] In Zhang B's scheme, fuzzy rules are extracted using a multi-class SVM. SVM, a commonly used machine learning algorithm, has the advantage of strong generalization ability and the capacity to handle nonlinear problems in high-dimensional feature spaces. However, it suffers from high computational complexity and difficulty in parameter selection. This actually limits the accuracy of the SVM model, which in turn affects the accuracy of the support vectors equivalent to the rules, ultimately impacting the overall control accuracy.

[0053] Zilke J's approach is good at extracting rules from neural networks used for classification problems, but it is less effective when it comes to regression problems that are often encountered in control.

[0054] like Figure 1 As shown, this embodiment of the invention discloses a multi-condition control method for zinc leaching based on DeepRED regression rules, including the following steps:

[0055] Includes the following steps:

[0056] Step 100: Extract spatiotemporal features and use them for working condition classification;

[0057] Step 200: Extract neural network rules based on DeepRED regression;

[0058] Step 300: Construct a data model based on mechanism error compensation;

[0059] Step 400: Construct a pH control framework based on a data model of rule extraction and mechanism error compensation to control the pH value of the intermediate leaching tank.

[0060] Specifically, step 100 includes dividing the time window, selecting multivariate time series data within a certain period as a single sample for training and testing; considering the spatiotemporal characteristics of multivariate time series data, the Slow Feature Analysis (SFA) method is used to extract temporal features, and then the Graph Attention Network (GAT) method is used to extract spatial features. After extracting the spatiotemporal features, they are used together as input to the Gaussian Mixture Model (GMM) for clustering. Clustering metrics are used to measure the quality of clustering results under different working conditions, thereby dividing the neutral leaching dataset into different working conditions.

[0061] The neutral leaching process is characterized by diverse operating states within the reaction tank, with numerous actual operating conditions and significant differences between them. This necessitates different control strategies for each condition. Therefore, it is essential to classify the operating conditions before implementing control measures.

[0062] The leaching process involves numerous and dynamically fluctuating inlet parameters, requiring data from only a specific period to fully describe the operating conditions. This results in a time series characteristic, exhibiting temporal properties. Simultaneously, the inlet variables are distributed across reaction tanks in different spatial locations, creating a close spatial connection and thus possessing spatial characteristics. Therefore, the data used for operating condition classification possesses both spatiotemporal properties.

[0063] To address the above problems, a working condition classification method integrating spatiotemporal feature extraction is proposed, such as... Figure 1 and Figure 8 As shown, considering the spatiotemporal characteristics of neutral leaching process data, spatiotemporal features are extracted and used for operating condition classification, thereby improving the accuracy of operating condition classification:

[0064] Since the operating conditions characterize the running status of the leaching tank over a period of time, multivariate time series data within a specific time window are selected as individual samples for training and testing by dividing the time window. Considering the spatiotemporal characteristics of multivariate time series data, Slow Feature Analysis (SFA) is used to extract temporal features, followed by Graph Attention Network (GAT) to extract spatial features. After extracting the spatiotemporal features, they are used together as input to a Gaussian Mixture Model (GMM) for clustering. Because GMM can output the probability of each sample belonging to each class, it can reduce the impact of input uncertainty to some extent.

[0065] Specifically, step 200 involves calculating the mutual information value between the time series of each inlet variable and the time series of the outlet pH value, and determining the control lag between each inlet variable and the outlet pH value based on the maximum mutual information value. After time series registration of the data, continuous input and output data are fuzzified, transforming them into a fuzzy language subset. Then, the DeepRED regression rule extraction method is used to extract rules in the form of "IF...... THEN......" from the trained neural network. The antecedent of the rule is the amount of roasted sand added to the neutral leaching reaction tank, the flow rate of the mixed liquor and the acid leaching supernatant at the control lag time, and the current outlet pH value. The consequent is...

[0066] The waste acid flow rate at the control time delay of tanks #1 and #2, and each variable in each tank has a different control time delay relative to the pH value of tank #3.

[0067] Before extracting rules, the issue of control lag needs to be considered. Control lag refers to the time from when an inlet variable enters the tank to when it finally affects the pH value of tank #3. Generally, the pH value at a point will be affected by an input over a period of time. For simplicity, we only consider the influence of an input at a point a certain time ago on the pH value at a single point. First, we calculate the mutual information value between the time series of each inlet variable and the time series of the outlet pH value, and determine the control lag between each inlet variable and the outlet pH value based on the maximum mutual information value. However, the magnitude of the influence of each inlet variable on the pH value is different. Therefore, we calculate the Spearman coefficient of each inlet variable with respect to the outlet pH value, normalize their Spearman coefficients according to the different tanks to which the inlet variables belong, and weight them with the control lag of each variable to obtain the overall control lag of all inlet variables in this tank.

[0068] After time-series registration of the data, in order to extract control rules, the continuous inputs and outputs are first fuzzified to transform the continuous data into a fuzzy language subset. Then, the DeepRED regression rule extraction method is used to extract rules in the form of "IF...... THEN......" from the trained neural network. The antecedent of the rule is the amount of roasted sand added to the three tanks, the flow rate of the mixed solution and the acid leaching supernatant (considering the control time delay, not the current time), and the outlet pH value at the current time. The consequent is the waste acid flow rate of tanks 1 and 2 (considering the control time delay, not the current time).

[0069] Step 300 includes the following: during the neutral leaching process, the waste acid flow rates of tanks 1 and 2 are used as controlled variables to control the pH value of the intermediate leaching tank. The deviation between the mechanism model and the actual model is used as the learning target of the high-precision data model to construct a data model based on mechanism error compensation. This data model is a model of all inlet variables with respect to the pH value. All inlet variables include two controlled variables.

[0070] In the neutral leaching process, the waste acid flow rates of tanks 1 and 2 are typically used as controlled variables to control the pH value of the intermediate leaching tank. Therefore, after obtaining the two controlled variables, it is necessary to construct a model of all inlet variables (including the two controlled variables) with respect to the pH value. Existing first-order reaction kinetic models can describe the reaction situation of the leaching process well, but they still cannot fully account for actual situations such as time delay, reactor volume changes, and changes in inlet liquid concentration such as waste acid. As a result, there is still a significant deviation between the mechanistic model and the actual model. Therefore, this invention considers using the deviation between the mechanistic model and the actual model as the learning target of the high-precision data model. Through error compensation, the interpretability of the mechanistic model is utilized while ensuring the overall high accuracy of the model.

[0071] Step 400 includes obtaining the controlled variables, namely the flow rates of waste acid #1 and waste acid #2, through rule extraction, and then sending the values ​​of the controlled variables as two inputs to the mechanism error compensation data model to obtain the predicted values ​​of the mechanism error compensation data model.

[0072] The pH value at the outlet of the intermediate leaching tank is between 3.0 and 3.5. Here, we set the control setpoint to 3.25 and use the error between the predicted value and the setpoint from the mechanistic error compensation data model as input to the rule extraction module, thus enabling continuous online control. Figure 2 and Figure 9 As shown.

[0073] In step 400, the mechanism model is as follows:

[0074] For slot #1:

[0075] ;

[0076] For slot #2:

[0077] ;

[0078] For slot #3:

[0079] ;

[0080] Assuming the liquid flow rates at the inlet and outlet remain constant, then:

[0081] F 1#out =F 1#acid +F 1#mixed +F 1#thermal；

[0082] F 2#out =F 1#out +F 2#acid；

[0083] F 3#out =F 2#out；

[0084] Among them, c 1# c 2# c 3# F represents the outlet hydrogen ion concentration of tanks #1, #2, and #3, respectively. 1#acid F 1#mixed F 1#thermal These represent the flow rates of waste acid, mixed solution, and acid leaching supernatant added to tank #1, respectively. 1#acid c 1#mixed c 1#thermal These represent the concentrations of the waste acid, the mixed solution, and the acid leaching supernatant added to tank #1, respectively. 2#acid c 2#mixed c 2#thermal They represent joining respectively

[0085] The concentrations of waste acid, mixed solution, and acid leaching supernatant in tank #2, F 2#acid and c 2#acid These represent the flow rate and concentration of the waste acid added to tank #2, respectively, where k represents the reaction rate constant, and A... 1# A 2# A 3# F represents the effective surface area of ​​the zinc calcinate reaction solid added to tanks #1, #2, and #3, respectively. 1#out F 2#out F 3#out These represent the outlet flow rates of tanks #1, #2, and #3, respectively, where V is the volume of a single reaction tank (generally, the volumes of the three reaction tanks are considered equal). Furthermore, the waste acid added to tanks #1 and #2 originates from waste electrolyte generated in the downstream electrolysis process, thus resulting in c... 1#acid =c 2#acid .

[0086] Specifically, this invention takes the neutral leaching process of a zinc smelter as the research object, such as... Figure 3As shown, the plant employs five cascaded leaching tanks arranged in a high-low order for industrial production. Tanks 1 through 3 are the main feeding tanks, aiming to achieve efficient leaching of zinc calcined sand in a relatively acidic solution. Tanks 4 and 5 primarily serve as impurity removal tanks. The introduction of oxygen oxidizes ferrous ions in the solution into ferric ions, which then hydrolyze in the lower acidity solution to form solid impurities, thus achieving impurity removal. Tank 3, as the intermediate leaching tank, is responsible for both the zinc calcined sand dissolution in the pre-leaching stage and the impurity ion removal in the post-leaching stage. Generally, as long as the pH value of tank 3 meets the requirements (stable within 3.0-3.5), a high zinc calcined sand leaching rate in the pre-leaching stage can be guaranteed, providing conditions for efficient impurity ion removal in the post-leaching stage.

[0087] This invention first performs a fusion of spatiotemporal feature extraction on a neutral leaching dataset for work condition classification. Slow feature analysis and graph attention networks are used to extract temporal and spatial features, respectively. The adjacency matrix (structure of the graph network) remains unchanged; only the feature representation of each node after passing through the network is modified. Next, the two features are merged according to variable dimensions and fed into a Gaussian mixture model. Since the Gaussian mixture model cannot automatically infer the number of sub-Gaussian models, i.e., the number of work condition classifications, we use two commonly used clustering metrics, the Calinski-Harabasz index (CH index) and the Davies-Bouldin index (DB index), to measure the clustering results under different numbers of work conditions. The results are as follows: Figure 5 As shown. Since the CH index is essentially the ratio of inter-cluster distance to intra-cluster distance, and the DB index is essentially the sum of the average intra-cluster distances divided by the maximum distance between the two class centers, and smaller intra-class covariance is better, while larger inter-class covariance is better, a larger CH index and a smaller DB index indicate better clustering results. Therefore, through... Figure 4 In this invention, the neutral leaching dataset was divided into five categories of operating conditions, which is equivalent to five leaching datasets for different operating conditions. By analyzing these five categories of operating conditions, we can find that they correspond to low acid, high acid leaching of supernatant, stable operating conditions with medium acid, unstable operating conditions with high acid, and stable operating conditions with high mixed solution.

[0088] After the operating conditions are divided, pH control needs to be performed separately for each category. First, rules are extracted using DeepRED regression to obtain the operating setpoints for the waste acid flow rates of tank #1 and tank #2. The DeepRED regression algorithm is represented as follows: Figure 5As shown, the extracted rules are all transformed from the branches of the CART regression tree into IF-THEN rules. After the rules are extracted, the consequents of the rules actually correspond to the flow rates of waste acid in tanks #1 and #2. These two controlled variables are then fed into the data model for mechanism error compensation, where the mechanism model is as follows:

[0089] Assuming the reactor volume remains constant during the neutral leaching reaction, the following first-order kinetic equation can be written for this liquid-solid heterogeneous reaction:

[0090] ;

[0091] Where: r represents the reaction rate, c represents the concentration of the substance of interest, v represents the stoichiometric coefficient of the substance of interest, k represents the reaction rate constant, A represents the effective surface area of ​​the reaction solid, α represents the reaction order, and V is the initial volume of the solution.

[0092] The neutral leaching process satisfies the first-order kinetic equation, therefore:

[0093] v=-1, α=1;

[0094] Substituting into the first-order dynamic equation, we get:

[0095] ;

[0096] From the perspective of material balance, the kinetic equation for each reactor can be obtained as follows:

[0097] ;

[0098] For slot #1:

[0099] ;

[0100] For slot #2:

[0101] ;

[0102] For slot #3:

[0103] ;

[0104] Assuming the liquid flow rates at the inlet and outlet remain constant, then:

[0105] F 1#out =F 1#acid +F 1#mixed +F 1#thermal；

[0106] F 2#out =F 1#out +F 2#acid；

[0107] F 3#out =F 2#out；

[0108] This allows us to obtain the hydrogen ion concentration c in tank #3. 3# The calculation formula still has some unknown parameters, which need to be obtained through parameter identification. The goal of parameter identification is to make the pH value of tank #3 in the mechanism model as close as possible to the actual value on site. The method used is the Grey Wolf Optimization Algorithm proposed by Mirjalili et al. in 2014. After obtaining the output of the mechanism model, the deviation between the output value and the true value is used as the supervision value of the LSTM network. After training, a better data model output can be obtained. The output of the mechanism model under one working condition, and the comparison between the mechanism-data model output and the true value are shown in the figure. Figure 6 As shown, the output of the mechanism-data model is closer to the true value.

[0109] Then, pH control frameworks based on data models of rule extraction and mechanism error compensation were built under five operating conditions, and the control results are as follows: Figure 7 As shown, (a) represents a stable low-acid condition, (b) a stable high-acid leaching supernatant condition, (c) a stable medium-acid condition, (d) a unstable high-acid condition, and (e) a stable high-mixture condition. It can be seen that the control framework proposed in this invention outperforms manual control and PID control in terms of control effectiveness.

[0110] Effect verification:

[0111] The outcome of control problems is often represented by a control curve, and the quality of the control curve is generally judged by a combination of dynamic and steady-state indicators. Here, when evaluating control effectiveness, we use three indicators: the average control error, the standard deviation of the control error, and the steady-state error. These three indicators comprehensively reflect the overall control effectiveness; the smaller the average control error, the standard deviation of the control error, and the steady-state error, the more accurate and stable the control effect. The definitions of the three indicators are as follows:

[0112] (1) Mean Control Error (MEA):

[0113] ;

[0114] Among them, e i e0 represents the pH control value output by the data model for mechanism error compensation at time i, n represents the set value of the pH value to be controlled, and n represents the number of allowed control times.

[0115] (2) Standard Deviation of Control Error (STD):

[0116] ;

[0117] in, This represents the average value of the control error.

[0118] (3) Controlling Steady-state Error (ESS):

[0119] e sta =|e end -e0|;

[0120] Among them, e end This represents the pH control value output by the data model for mechanism error compensation at the last moment.

[0121] In addition to the three general control indicators mentioned above, two other indicators, Fluctuation Range (FR) and Compliance Rate (CR), are defined for the neutral leaching process. Fluctuation Range represents the range of all pH control values ​​of the proposed method from the lower limit to the upper limit, while Compliance Rate represents the range of pH control values ​​of the proposed method that are in the range of 3.0 to 3.5.

[0122] In this case study, the proposed method was applied to five different operating conditions. Simultaneously, manual control and PID control, widely used in industrial settings, were compared with the proposed method, and the resulting control curves are shown below. Figure 7 As shown in Table 1, the calculated five control indicators are represented by MEA, STD, ESS, FR, and QR.

[0123] Table 1 Comparison of the control effects of the proposed control framework with manual control and PID control

[0124]

[0125] As can be seen from the table above, the proposed method outperforms manual control and PID control in most control indices under different operating conditions. This indicates that the proposed method, while having high interpretability, also outperforms manual control and PID control, which are currently widely used in industrial settings, in terms of control accuracy. This demonstrates the effectiveness of the proposed method for KPI control in process industry production processes.

[0126] The present invention also provides a zinc leaching multi-condition control system based on regression DeepRED rule extraction, including a memory, a control processor, and a computer program stored in the memory and executable on the control processor. The control processor executes the program to implement the zinc leaching multi-condition control method based on regression DeepRED rule extraction as described above.

[0127] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments. Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized form in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each particular application, but such implementation decisions should not be construed as departing from the scope of the invention. The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, battery compartment control board, micro battery compartment control board, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal. In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted therefrom as one or more instructions or code on a computer-readable medium.Computer-readable media includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. Storage media can be any available medium accessible to a computer. By way of example and not limitation, such computer-readable media may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as computer-readable media. For example, if software is transmitted from a website, a central control computer, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then that coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of media. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multi-condition control method for zinc leaching based on regression DeepRED rule extraction, characterized in that: The following steps are involved: Step 100, extracting spatiotemporal features and using them for working condition classification; Step 200, extracting neural network rules based on regression DeepRED; Step 300, constructing a data model based on mechanism error compensation; Step 300 includes, in the neutral leaching process, using the waste acid flow of tank 1# and tank 2# as the controlled quantity, controlling the pH value of the intermediate leaching tank, and using the deviation between the mechanism model and the actual model as the learning target of the high-precision data model to construct a data model based on mechanism error compensation, the data model is a model of all input variables with respect to pH value, and all input variables include two controlled quantities; Step 400, constructing a pH value control framework based on a data model of rule extraction and mechanism error compensation to control the pH value of the intermediate leaching tank, said step 400 includes obtaining the controlled quantity, i.e., the flow value of the waste acid in tank 1# and the waste acid in tank 2#, through rule extraction, and then sending the values ​​of the two controlled quantities as input to the data model of mechanism error compensation to obtain the predicted value of the data model of mechanism error compensation.

2. The zinc leaching multi-condition control method based on regression DeepRED rule extraction according to claim 1, characterized in that: Step 100 specifically includes, by dividing the time window, selecting the multivariate time series data within a period of time as a single sample for training and testing; in view of the spatiotemporal characteristics of the multivariate time series data, the slow feature analysis SFA method is used to extract the time features, and then the graph attention network GAT method is used to extract the spatial features. After the spatiotemporal features are extracted, they are clustered together as the input of the Gaussian mixture model GMM, and the clustering index is used to measure the quality of the clustering results under different numbers of working conditions, so as to divide the neutral leaching data set into different working conditions.

3. The zinc leaching multi-condition control method based on regression DeepRED rule extraction according to claim 1 or 2, characterized in that: Step 200 specifically includes calculating the mutual information value between each inlet variable time series and the outlet pH value time series, and determining the control time lag between each inlet variable and the outlet pH value according to the maximum mutual information value. After the data is time-series aligned, the continuous input data and output data are fuzzified, and the continuous input data and output data are converted into fuzzy language subsets. Then, the rule extraction method of regression DeepRED is used to extract rules in the form of "IF... THEN..." from the trained neural network, wherein the antecedent part of the rule is the amount of roasted sand added to the neutral leaching reaction tank after the control time lag alignment, the flow rate of the mixed liquor and the acid leaching supernatant, and the outlet pH value at the current moment, and the successor part is the waste acid flow rate of the 1# tank and the 2# tank at the control time lag moment.

4. The zinc leaching multi-condition control method based on regression DeepRED rule extraction according to claim 1 or 2, characterized in that: In step 400, the mechanism model is: For 1# groove: ； For 2# groove: ； For 3# groove: ； Assuming that the liquid flow rate at the inlet and outlet remains unchanged, then: F 1#out =F 1#acid +F 1#mixed +F 1#thermal； F 2#out =F 1#out +F 2#acid； F 3#out =F 2#out； Among them, c 1# , c 2# , c 3# Respectively represent the outlet hydrogen ion concentration of 1# tank, 2# tank, and 3# tank, F 1#acid , F 1#mixed , F 1#thermal They represent the flow rates of the waste acid, mixed liquor and acid leaching supernatant added to tank 1#, c 1#acid , c 1#mixed , c 1#thermal They represent the concentrations of the waste acid added to tank 1#, the mixed solution and the acid leaching supernatant respectively, c 2#acid , c 2#mixed , c 2#thermal They represent the concentrations of the waste acid added to tank 2#, the mixed solution and the acid leaching supernatant respectively, 2#acid and c 2#acid They represent the flow rate and concentration of the waste acid added to tank 2#, k represents the reaction rate constant, and A 1# , A 2# , A 3# F represents the effective surface area of ​​the zinc roasted sand reaction solid added to tank 1#, tank 2#, and tank 3#, respectively. 1#out , F 2#out , F 3#out They represent the outlet flow rates of No. 1, No. 2, and No. 3 tanks, respectively. V is the volume of a single reaction tank (it is generally believed that the volumes of the three reaction tanks are equal). The waste acid added to No. 1 and No. 2 tanks comes from the waste electrolyte generated in the downstream electrolysis process, so there is c 1#acid =c 2#acid .

5. A zinc leaching multi-condition control system based on regression DeepRED rule extraction, characterized in that: It comprises a memory, a control processor and a computer program stored in the memory and executable on the control processor, wherein the control processor executes the program to implement the zinc leaching multi-condition control method based on regression DeepRED rule extraction as described in any one of claims 1 to 4.

Images