A material feeding control method, system, device and medium for a cadmium oxide production line

Abstract

The present invention belongs to the field of industrial control technology, and discloses a material feeding control method, system, device and medium for a cadmium oxide production line, including: collecting sensor data corresponding to various indicators in the production process, where the sensor data is first sensor data collected at multiple time scales or second sensor data collected at a single time scale; fusing the first sensor data to obtain a multi-scale fusion vector of each indicator collected at multiple time scales; inputting all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain time series features output by the LSTM network; inputting the time series features into a BP neural network to obtain optimized PID parameters output by the BP neural network; and inputting the optimized PID parameters into a PID controller to optimize and adjust the feeding ratio of the material. It can improve the accuracy and stability of the feeding control of the cadmium oxide production line.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial control, and particularly relates to a material feeding control method, system, device and medium for a cadmium oxide production line. Background Art

[0002] Cadmium oxide (CdO), as an important inorganic compound, is widely used in fields such as semiconductor materials, optoelectronic devices, pigments, and transparent conductive films. Its production process usually involves the oxidation reaction of cadmium metal or cadmium compounds in oxygen or air. The production process of cadmium oxide is relatively complex, especially in industrial-scale production, which involves precise ratio control of various raw materials, including cadmium metal, oxygen, catalysts, and other additives.

[0003] In modern industrial production, PID automatic control based on BP neural network has become a widely used intelligent control method, especially in complex process industries such as chemical engineering and materials. The BP neural network uses its powerful non-linear mapping ability to adjust the parameters of the PID controller in real time, helping the system maintain stability in an uncertain environment.

[0004] However, with the complication and refinement of production processes, especially in application scenarios with high-precision requirements such as feeding control, the BP neural network has exposed some deficiencies in processing real-time data and maintaining control accuracy: the dynamic relationships between various data in real-time data are constantly changing in the production environment, and it is difficult for the BP neural network to adapt in real time, often resulting in a lag in control response; it lacks sensitivity to time-series and multi-scale data and cannot accurately capture and predict mutation signals or trend changes, affecting control accuracy.

[0005] Therefore, the accuracy and stability of current feeding control need to be improved. Summary of the Invention

[0006] The purpose of the present invention is to provide a material feeding control method, system, device and storage medium for a cadmium oxide production line, which can improve the accuracy and stability of feeding control.

[0007] The first aspect of the present invention discloses a material feeding control method for a cadmium oxide production line, including:

[0008] Collecting sensor data corresponding to various indicators in the production process, where the sensor data is first sensor data collected at multiple time scales or second sensor data collected at a single time scale;

[0009] Fusing the first sensor data to obtain a multi-scale fusion vector of each indicator collected at multiple time scales;

[0010] Input all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain the time series features output by the LSTM network;

[0011] Input the time series features into a BP neural network to obtain the optimized PID parameters output by the BP neural network;

[0012] Input the optimized PID parameters into a PID controller to optimize and adjust the feeding ratio of the material.

[0013] In some embodiments, after collecting the sensor data corresponding to various indicators in the production process, it further includes:

[0014] Calculate the correlation between the sensor data and the control target, and adjust the acquisition frequency of the sensor and / or adjust the weight corresponding to the sensor data according to the correlation;

[0015] and / or,

[0016] Calculate the covariance between the sensor data, and perform data dimensionality reduction and / or adjust the weight corresponding to the sensor data according to the covariance.

[0017] In some embodiments, after collecting the sensor data corresponding to various indicators in the production process, it further includes:

[0018] Eliminate the noise and outliers in the sensor data to obtain the processed sensor data;

[0019] Input the processed sensor data into a PID controller to perform a preliminary adjustment on the feeding ratio of the material.

[0020] In some embodiments, after collecting the sensor data corresponding to various indicators in the production process, the sensor data is also sampled multiple times at different time scales, and the first sensor data is generated according to the sensor data.

[0021] In some embodiments, the time series features include the features extracted from the sensor data corresponding to the feeding rates of cadmium metal, oxygen, and catalyst, and the features extracted from the sensor data corresponding to the reactor temperature, reactor pressure, and oxygen concentration; the optimized PID parameters include the proportional parameters, integral parameters, and derivative parameters corresponding to cadmium metal, oxygen, and catalyst respectively.

[0022] In some embodiments, inputting all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain the time series features output by the LSTM network includes:

[0023] In the LSTM network, the forgetting gate is used to forget the interference of sudden fluctuations on the control parameters during the stable production stage. The input gate is used to screen and add the data from the multi-scale fusion vector and the second sensor data to the memory unit of the LSTM network. The output gate is used to output the key features of multiple time steps according to the memory unit of the LSTM network, forming the time series features.

[0024] The second aspect of the present invention discloses a material feeding control system for a cadmium oxide production line, including:

[0025] A sensing layer for collecting sensor data corresponding to multiple indicators during the production process. The sensor data is the first sensor data collected at multiple time scales or the second sensor data collected at a single time scale;

[0026] A processing layer for fusing the first sensor data to obtain a multi-scale fusion vector for each indicator collected at multiple time scales;

[0027] An intelligent optimization layer inputs all the multi-scale fusion vectors and the second sensor data into the LSTM network to obtain the time series features output by the LSTM network; inputs the time series features into the BP neural network to obtain the optimized PID parameters output by the BP neural network, and inputs the optimized PID parameters into the control layer;

[0028] A control layer for optimizing and adjusting the feeding ratio of the material according to the optimized PID parameters;

[0029] An execution layer for controlling the valve, pump, and material conveying device according to the adjustment instruction of the control layer to control the feeding ratio of the material.

[0030] In some embodiments, the processing layer is further configured to calculate the correlation between the sensor data and the control target, and adjust the acquisition frequency of the sensor and / or adjust the weight corresponding to the sensor data according to the correlation;

[0031] And / or,

[0032] Calculate the covariance between the sensor data, and perform data dimensionality reduction and / or adjust the weight corresponding to the sensor data according to the covariance.

[0033] The third aspect of the present invention discloses an electronic device, including a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the material feeding control method for the cadmium oxide production line disclosed in the first aspect.

[0034] The fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the material feeding control method of the cadmium oxide production line disclosed in the first aspect.

[0035] The beneficial effects of the present invention are as follows: by collecting sensor data on multiple time scales, richer information can be extracted in terms of short-term and long-term changes. By using the LSTM network to capture time series features, mutation signals or trend changes can be accurately captured and predicted. Then, by using the BP neural network to convert the time series features into optimized PID parameters, the dynamic relationships among various data changing constantly in the production environment can be addressed, improving the accuracy and stability of the feeding control of the cadmium oxide production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings herein show specific examples of the technical solutions of the present invention and, together with the specific embodiments, form part of the description for explaining the technical solutions, principles and effects of the present invention.

[0037] Unless otherwise specifically stated or defined, in different drawings, the same reference numerals represent the same or similar technical features. For the same or similar technical features, different reference numerals may also be used to represent them.

[0038] Figure 1 is an architecture diagram of a material feeding control system for a cadmium oxide production line disclosed in an embodiment of the present invention;

[0039] Figure 2 is a flowchart of a material feeding control method for a cadmium oxide production line disclosed in an embodiment of the present invention;

[0040] Figure 3 is a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Unless otherwise specifically stated or defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. In the case of combining the technical solutions of the present invention with real scenarios, all technical and scientific terms used herein may also have meanings corresponding to the purpose of implementing the technical solutions of the present invention. The "first, second..." used herein are only for distinguishing names and do not represent specific quantities or orders. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0042] It should be noted that when an element is considered to be "fixed to" another element, it can be directly fixed to the other element or there can be an intermediate element; when an element is considered to be "connected to" another element, it can be directly connected to the other element or there can be an intermediate element at the same time; when an element is considered to be "mounted on" another element, it can be directly mounted on the other element or there can be an intermediate element at the same time. When an element is considered to be "provided in" another element, it can be directly provided in the other element or there can be an intermediate element at the same time.

[0043] Unless otherwise specified or defined, the "said" and "the" used in this article refer to the technical features or technical content mentioned or described before the corresponding position. The technical features or technical content may be the same as or similar to the technical features or technical content it mentions. In addition, the terms "including" and "having" and any variations thereof used in this article are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0044] At present, the stability and accuracy of the feed control in the production line are still not ideal enough, which limits the application effect of the traditional BP neural network in high-precision and multi-variable industrial control.

[0045] To solve the above problems, the present invention uses methods such as dynamic correlation monitoring, covariance analysis, multi-scale data fusion, and incremental data cleaning for data processing to improve data correlation, data quality, and the time dependence of the model. Then, a hybrid control model formed by combining the LSTM (Long Short-Term Memory Network) network and the BP (Backpropagation Neural Network) neural network is used for feed control to form an intelligent, high-precision, and robust material feed control system. It can respond in a timely manner to changes in the dynamic relationship between data, accurately capture and predict mutation signals or trend changes according to time series and multi-scale data, and improve control accuracy and stability.

[0046] For the more difficult-to-control situation of the precise proportioning of raw materials in the cadmium oxide production line, such as Figure 1 shown, this embodiment provides a material feed control system. The overall architecture of the control system can be divided into a sensing layer, a processing layer, an intelligent optimization layer, a control layer, and an execution layer.

[0047] The sensing layer collects sensor data corresponding to various indicators (i.e., key parameters in the production process) during production. Among them, some indicators are collected at multiple time scales (such as 1 second and 5 seconds) to obtain the first sensor data, and some indicators are collected at a single time scale to obtain the second sensor data. These real-time collected sensor data provide basic data input for control and optimization, and are also used for PID control and the learning and optimization of the hybrid control model. The hardware included in the sensing layer of this embodiment is: flow sensors, temperature sensors, pressure sensors, gas sensors, level sensors, edge processing gateways, and PLCs (programmable logic controllers). Among them, the flow sensors are used to monitor the flow rate and feeding rate of materials (such as cadmium metal, oxygen, catalysts, etc.) to ensure that the input of each material is precisely controllable; the temperature sensors are used to monitor the temperature inside the reactor to ensure that the reaction environment meets the requirements; the pressure sensors are used to monitor the pressure inside the reactor to avoid the process conditions exceeding the safe range; the gas sensors are used to detect the concentration of oxygen or other gases during the reaction to ensure that the supply ratio of the gas meets the reaction requirements; the level sensors are used to monitor the inventory of materials to avoid the reaction being affected by insufficient or excessive materials; the edge processing gateways are used to collect and transmit the data feedback by the sensors to ensure that the real-time data of the sensors are transmitted to the PLC; the PLC is used to preliminarily process the sensor data, and feedback the processing results to the PID controller and the processing layer, and can also control the basic production process.

[0048] The processing layer preprocesses and optimizes the collected sensor data, performs multi-level and multi-scale processing, and provides high-quality, strongly relevant, and noise-filtered data input for the hybrid control model; for the first sensor data, fusion is also performed to obtain a multi-scale fusion vector of each indicator collected at multiple time scales. In this embodiment, the correlation between the sensor data and the control target is calculated in the processing layer, and the acquisition frequency of the sensors and / or the weights corresponding to the sensor data are dynamically adjusted according to the correlation; and / or, the covariance between the sensor data is calculated, and data dimensionality reduction and / or the weights corresponding to the sensor data are adjusted according to the covariance; in the processing layer, the noise and outliers in the sensor data can also be eliminated to obtain the processed sensor data, and at the same time, the processed sensor data is input into the control layer.

[0049] The intelligent optimization layer inputs all the multi-scale fusion vectors and the second sensor data into the LSTM network to obtain the time-series features output by the LSTM network; inputs the time-series features into the BP neural network to obtain the optimized PID parameters output by the BP neural network, and inputs the optimized PID parameters into the control layer to dynamically adjust the parameters of the PID controller through the BP neural network. The intelligent optimization layer is deployed on a high-performance industrial server or industrial control computer to ensure the computing power of the LSTM network and the BP neural network. The BP neural network will continuously learn the nonlinear behavior of the system based on historical data and real-time production data, and optimize the PID parameters to cope with the changes in process conditions and nonlinear responses.

[0050] The control layer includes: a PID controller and a valve controller. Among them, the PID controller is the main controller responsible for regulating the material input in the whole production process, and optimally adjusts the feeding ratio of the material according to the optimized PID parameters. It is the core control unit of the whole system and is responsible for real-time adjustment of the feeding ratio of the material. The valve controller is connected to the PID controller to adjust the valve opening in the material conveying system and control the flow of materials and gases.

[0051] The PID controller can also use the processed sensor data to preliminarily adjust the ratios of various materials, that is, according to the processed sensor data, through the difference between the set target values (desired feeding ratio, temperature or pressure) and the processed sensor data, perform proportional, integral and derivative adjustments, and output control instructions to the valve controller.

[0052] The execution layer is the actual implementation unit of the control instructions. According to the control instructions of the control layer, it controls valves, pumps and material conveying devices to adjust the feeding ratio of the material. The execution layer includes: pneumatic valves, conveying pumps and batching scales, etc. Among them, the pneumatic valve is used to control the flow of fluids or gases, such as the feeding control of materials such as cadmium metal, oxygen, and additives; the conveying pump is used to accurately control the feeding amount of solid or liquid materials; the batching scale is used to accurately dispense powder or granular materials. The devices in the execution layer will act according to the instructions issued by the PID controller, and send the execution results back to the sensing layer again through the feedback loop to form a closed-loop control.

[0053] When the above material feeding control system operates, it executes a material feeding control method for a cadmium oxide production line in an embodiment of the present invention to control the material feeding of the cadmium oxide production line. To facilitate the understanding of the present invention, the specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings of the specification.

[0054] As Figure 2 shown, the method includes the following steps:

[0055] Step S100: Collect sensor data corresponding to various indicators in the production process, where the sensor data is the first sensor data collected at multiple time scales or the second sensor data collected at a single time scale;

[0056] Among them, the indicators are key process parameters in production, such as: the feeding rates of cadmium metal, oxygen, and catalyst, reactor temperature, reactor pressure, and oxygen concentration, etc. Through various sensors (such as flow, temperature, and pressure sensors), sensor data corresponding to various indicators is collected.

[0057] In this embodiment, when collecting sensor data, for indicators with significant dynamic changes in the production process, such as fluctuations in the feeding rate or changes in oxygen concentration, multiple time scales are used for collection, that is, the same indicator is collected at two time scales, short time and long time, respectively, to obtain the first sensor data. For indicators with slow changes in the production process, it is collected at a single time scale to obtain the second sensor data. The short time scale can capture the rapid fluctuations of the indicator; the long time scale can capture the long-term trend of the indicator. For example: for the feeding rate, the short time scale is to collect once per second, used to detect sudden changes (such as sudden adjustment of the valve opening), and the long time scale is to take an average value every 5 seconds, used to analyze the long-term change trend (such as the gradual increase or decrease of the feeding rate over a period of time). Through the combination of long-time and short-time collection, the multi-level change characteristics of the indicator can be captured. Especially in the dynamic control scenario, it can not only respond to short-term fluctuations in a timely manner but also adjust the control strategy according to the long-term trend. Moreover, the first sensor data contains rich information, which is beneficial for the LSTM network to learn the dependency relationship of sensor data at different time sequences.

[0058] In some embodiments, for indicators with significant dynamic changes in the production process, after collecting the sensor data, the sensor data is sampled multiple times at different time scales. For example, it is resampled in the sensor data at time scales of 1 second or 5 seconds respectively, so as to generate the first sensor data according to the sensor data.

[0059] Step S200: Fuse the first sensor data to obtain a multi-scale fusion vector for each indicator collected at multiple time scales;

[0060] This step is called multi-scale data fusion, which fuses the short-time and long-time data in the first sensor data into a multi-scale fusion vector for subsequent LSTM network. Assume the short-time data vector is , and the long-time data vector is , then the multi-scale fusion vector is: . Among them, represents the short-time data vector, such as the rapid change value of the feeding rate; represents the long-time data vector, such as the long-term change trend of the feeding rate; represents a multi-scale fusion vector, providing information on the changes of each index at different time scales.

[0061] In cadmium oxide production, the first sensor data containing data at different time scales is fused and then input into the LSTM network. This can not only quickly respond to feed fluctuations but also track the long-term trend of the feed, enhancing the accuracy of the control system.

[0062] Step S300: Input all the multi-scale fusion vectors and the second sensor data into the LSTM network to obtain the time series features output by the LSTM network;

[0063] Industrial feed control usually involves data changes at different time scales. For example, the change in temperature is slow, but the changes in oxygen concentration and feed rate are fast. The BP neural network is weak in processing data at different time scales and lacks sensitivity to the characteristics of time series and multi-scale data, resulting in the control system being unable to accurately capture and predict mutation signals or trend changes, affecting the control accuracy.

[0064] Therefore, in this embodiment, all the multi-scale fusion vectors and the second sensor data are input into the LSTM network, and the LSTM network can learn the time series characteristics of the input data. The LSTM network contains an input gate, a forget gate, and an output gate, enabling it to selectively retain or forget historical information, solving the problem of gradient disappearance in traditional neural networks and being suitable for processing sequence data with long-term dependencies. Then, the time series features output by the LSTM network are used as the input to the BP neural network.

[0065] In the LSTM network, the input at each time step is processed by three key gates, namely the forget gate, the input gate, and the output gate. The forget gate is used to forget the interference of sudden fluctuations on the control parameters during the stable production stage. The input gate is used to screen and add data to the memory unit of the LSTM network from the multi-scale fusion vector and the second sensor data. The output gate outputs the key features of multiple time steps according to the memory unit of the LSTM network to form time series features.

[0066] Specifically, the forget gate determines how much past information the LSTM network needs to discard at the current time step. In feed control, the forget gate decides whether to forget some previous historical features, such as forgetting the interference of sudden fluctuations on the control parameters during the stable production stage. The calculation formula of the forget gate is: , where is the output of the forget gate, and the value ranges from 0 to 1, representing the proportion of forgetting. 0 means complete forgetting, and 1 means complete retention; is the weight matrix of the forget gate; is the hidden state of the previous time step. is the input at the current time step (i.e., time series data such as temperature, feed rate, etc.); is the bias of the forget gate; is the Sigmoid activation function that maps values between 0 and 1.

[0067] The input gate determines how much of the new information at the current time step (i.e., the current sensor data) needs to be added to the memory cell. This step is particularly crucial for feed control with high real-time requirements, enabling the LSTM network to respond quickly when changes occur in the control system (such as a sudden increase in the feed rate). The calculation process of the input gate is as follows: 1. Control the "joining strength" of the new information through the input gate: , where is the output of the input gate, is the weight matrix of the input gate, is the hidden state of the previous time step, is the input at the current time step (i.e., time series data such as temperature, feed rate, etc.), is the bias of the input gate. 2. Generate a candidate memory content , which is the information that can be added to the memory at the current time step: , is the weight matrix of the memory gate, is the bias of the memory gate; 3. Obtain the calculation result of the input gate: . is the state of the memory cell at the current time step; is the memory state of the previous time step; is the part of the historical information controlled by the forget gate, indicating how much memory from the previous moment is retained; is the part of the current information controlled by the input gate, indicating how much information of the current input is introduced.

[0068] The output gate determines which information will be extracted and output from the current memory cell as the input for the next time step. This output will serve as the hidden state at the current time step and be passed to the next time step. For feed control, the output gate can help the LSTM network output the key features at each time step based on a series of time series data for generating PID control parameters. The calculation formula of the output gate is: , and then, based on the activation value of the memory cell state and the weights of the output gate, the final hidden state can be calculated as: . Among them, is the weight matrix of the output gate, is the bias of the output gate; is the hidden state at the current time step, that is, the feature extraction result at the current time step, which contains the dynamic time series information of the input data; is the activation value of the memory cell, which compresses the state data to between -1 and 1; is the output control ratio of the output gate.

[0069] At each time step the hidden state of the LSTM network will contain the key features of the previous few time steps. These hidden states form a time series feature vector, that is, the time series feature for the input of the subsequent BP neural network.

[0070] In this embodiment, the hidden state output by the LSTM network is used as the input of the BP neural network, including 6 key features closely related to the feed control of 3 materials, the features extracted from the sensor data corresponding to the feed rates of cadmium metal, oxygen, and catalyst, and the features extracted from the sensor data corresponding to the reactor temperature, reactor pressure, and oxygen concentration.

[0071] Step S400: Input the time series features into the BP neural network to obtain the optimized PID parameters output by the BP neural network;

[0072] The BP neural network is mainly used to dynamically adjust the PID parameters of the PID controller (including: P, proportional parameter, I, integral parameter, D, derivative parameter), that is, to further convert the time series features learned from the time series features into PID parameters to achieve adaptive optimization control of the multi-material feed ratio in cadmium oxide production.

[0073] The BP neural network includes an input layer, a hidden layer, and an output layer. Among the time series feature vectors output by the LSTM layer include 6 key features, so the input layer has 6 neurons, corresponding to 6 input variables: , is the feed rate of cadmium metal, is the delivery rate of oxygen, is the feed rate of the catalyst, is the reactor temperature, is the reactor pressure, is the oxygen concentration. The hidden layer is responsible for extracting the features of the input information and learning the complex relationship between the input and output through non-linear mapping. In this embodiment, the hidden layer has 10 neurons, and the output of the hidden layer is calculated as: , represents the weight from the $i$-th neuron in the input layer to the $j$-th neuron in the hidden layer. The weight reflects the degree of influence of the $i$-th input variable in the input layer on the output of the $j$-th neuron in the hidden layer. represents the $i$-th input value in the input layer. The input variables include the feed rates of cadmium metal, oxygen, and catalyst, reactor temperature, pressure, and oxygen concentration. Therefore may correspond to one of these variables respectively; represents the bias term of the $j$-th neuron in the hidden layer. represents the weighted sum of all input variables in the input layer, which is the weighted input value of the $j$-th neuron in the hidden layer. The output vector of the hidden layer is: . The output layer is responsible for generating the PID parameters (P, I, D) of the PID controller for each material, which are used to adjust the feed rates of each material. Since 3 materials (cadmium metal, oxygen, catalyst) are controlled, the output layer contains 9 neurons, and the optimized PID parameters are . Among them, is the PID control parameter for cadmium metal, is the PID control parameter for oxygen, is the PID control parameter for catalyst. The calculation formula of the output layer is , where is the weight from the hidden layer to the output layer, is the bias term of the output layer.

[0074] Step S500: Input the optimized PID parameters into the PID controller to optimize and adjust the feed ratio of the materials.

[0075] After receiving the optimized PID parameters output by the BP neural network, the PID controller outputs a control signal to devices such as pneumatic valves, conveyor pumps, and metering devices through proportional (P), integral (I), and derivative (D) algorithms, according to the deviation between the optimized PID parameter value and the actual PID parameter value, to adjust the opening of the material feed valve or the flow rate of the pump, optimize and adjust the feed ratio of the materials, and achieve precise control of the material ratio. The action results of devices such as pneumatic valves, conveyor pumps, and metering devices (such as changes in material flow rate, temperature, or pressure in the reactor) are collected and fed back through sensors, forming a closed-loop control.

[0076] In this embodiment, continuous adjustment is also made according to real-time data and feedback to ensure that the production process is always in the optimal state. The BP neural network continues to learn online, that is, during the generation process, after obtaining a new set of sensor data each time, a forward propagation and a backward propagation will be performed to update the weights and biases. This step-by-step adjustment method enables the network to adapt to continuously changing data patterns, thereby coping with complex production conditions. As production continues, the control effect becomes more and more precise and intelligent.

[0077] In summary, by collecting sensor data using multiple time scales, richer information can be extracted for short-term and long-term changes. Then, the LSTM network is used to capture time series features, and the BP neural network is used to convert the time series features into optimized PID parameters, which can improve the accuracy and stability of the feed control in the cadmium oxide production line.

[0078] Traditional BP neural networks mainly rely on historical data for training. Although they have a certain learning ability, when faced with real-time data changes in complex processes, they cannot quickly and adaptively capture the dynamic correlations between key variables. For example, in the cadmium oxide production process, the mutual influence between the feed rate, temperature, pressure, and oxygen concentration is significant, and this dynamic relationship changes constantly in the production environment. Traditional BP neural networks are difficult to adapt in real time, often resulting in a lag in control response. Moreover, in a multivariable control system, the influence degrees of each data source on the control target are different. Traditional BP neural networks treat input data equally and cannot adjust the priority of data sources according to the changes in actual working conditions, which may lead to key data being ignored while secondary data being overemphasized. This causes the control of key parameters such as the feed rate by the BP neural network to fail, thereby affecting production efficiency and product quality.

[0079] Therefore, after collecting the sensor data corresponding to each index in the production process in this embodiment, dynamic correlation detection and / or covariance analysis are also performed. Dynamic correlation detection is used to calculate the correlation between the sensor data and the control target, and the acquisition frequency of the sensor and / or the weight corresponding to the sensor data are dynamically adjusted according to the correlation; covariance analysis is used to calculate the covariance between the sensor data, and data dimensionality reduction and / or the weight corresponding to the sensor data are adjusted according to the covariance.

[0080] Specifically, dynamic correlation detection identifies which indicators have a more significant impact on the control target (such as the feed ratio, reaction rate) under the current production conditions. The data acquisition frequency and priority are dynamically adjusted according to the monitored correlation to ensure that key data sources are given priority for processing.

[0081] First, the correlation calculation is performed. The Pearson Correlation Coefficient is used to calculate the linear relationship between the indicator and the control target. The calculation formula for the correlation coefficient is:

[0082] ,

[0083] where and respectively represent the sensor data of the indicator and the observed value of the control target, and respectively represent the mean of the sensor data of the indicators and the control target, The value range of is between [-1, 1]. The closer the value is to 1 or -1, the stronger the correlation between the two variables. When it is close to 0, the correlation is weak.

[0084] Then, priority adjustment is performed. Every other sampling period, the correlation of each indicator with the control target is calculated. If the correlation of some indicators exceeds the preset threshold (such as 0.7), the data acquisition frequency of that indicator is increased. For example, at a certain moment, it is found that the correlation between the oxygen concentration and the feed rate reaches 0.8, then the oxygen concentration acquisition frequency is increased to obtain more critical information.

[0085] Covariance analysis is used to identify the co-variation relationship between indicators in the feed control system, that is, two or more indicators have similar trends in numerical changes. For example, temperature and oxygen concentration may show a strong co-variation relationship under certain working conditions. When some indicators show a strong co-variation relationship, dimensionality reduction processing (such as principal component analysis, PCA) can be performed on them, or smoothing filtering can be performed to reduce the noise interference of data input. And / or, the weight corresponding to the sensor data is adjusted by adjusting the weight matrix in the LSTM network.

[0086] The real-time data collected in industrial production often contains various noises and sudden interferences. The traditional BP neural network does not have special processing for industrial noises, resulting in the noises and interferences in the input data affecting the control effect. Especially in the case of large data volatility, the network output is prone to instability, making the feed ratio control inaccurate and even causing the instability of the system.

[0087] Therefore, in this embodiment, incremental data cleaning is also performed to eliminate the noises and outliers in the sensor data. For example: after detecting the rationality of the sensor data, data correction or smoothing processing is performed according to the preset rules to obtain the processed sensor data. The processed sensor data can be directly input into the PID controller to preliminarily adjust the feed ratio of the material; and, the processed sensor data is fused and input into the LSTM network to ensure that the data entering the LSTM network maintains high quality.

[0088] Exemplarily, the process of outlier detection is: set the reasonable threshold for each indicator based on historical data. Assume the reasonable range of temperature , for the temperature data , if then it is judged as an outlier. The process of data correction or smoothing processing is: replace the outlier with the mean of the previous moment and the next moment ; or, use linear interpolation to fill the outlier.

[0089] In this embodiment, the feeding control system in cadmium oxide production collects the feeding rate once per second. Suppose that during a production conversion, the system records an abnormal increase in the fluctuation of the feeding rate. The incremental cleaning will detect each new piece of data per second according to the current fluctuation situation, and clean the data beyond the reasonable range by using methods such as interpolation or mean smoothing, so as to ensure that the sensor data is always stable and will not affect the production efficiency due to abnormal fluctuations.

[0090] Through the above methods such as dynamic correlation detection, covariance analysis, and incremental data cleaning, the quality of sensor data can be improved, ensuring the accuracy and effectiveness of sensor data, and improving the accuracy and stability of the feeding control of the cadmium oxide production line.

[0091] In this embodiment, a hybrid control model is composed of an LSTM network and a BP neural network. It is easy to understand that a data processing module can also be added to the hybrid control model, and operations such as dynamic correlation detection, covariance analysis, incremental data cleaning, and multi-scale data fusion can be performed in the data processing module.

[0092] Before using the hybrid control model, the LSTM network and the BP neural network need to be jointly trained. The specific training process is as follows: After performing preprocessing operations such as dynamic correlation detection, covariance analysis, incremental data cleaning, and multi-scale data fusion on the sensor data such as the real-time collected feeding rate, temperature, humidity, and oxygen concentration, the data is input into the LSTM network. The LSTM network processes the input data, extracts time series features, and generates feature vectors . The feature vectors output by the LSTM network are input into the BP network. The BP network generates PID parameters (P, I, D), compares the generated PID parameters with the target values, and calculates the MSE loss. Propagate backward layer by layer from the output layer of the BP network to the LSTM layer, calculate the gradient of the loss with respect to the weights of each layer, and update the weights of the LSTM and BP networks according to the gradient to complete one training iteration. Continuously repeat the above process until the loss function converges and the control effect of the hybrid control model meets the expected requirements.

[0093] Among them, the loss function reflects the deviation between the PID parameters generated by the hybrid control model and the expected control parameters, so that the hybrid control model can be adjusted in the optimal control direction. Suppose the target PID parameter is , and the PID parameter output by the model is , then the MSE loss function can be expressed as:

[0094] ,

[0095] Among them, N is the number of PID parameters (such as 9 for P, I, D of 3 kinds of materials), and They respectively represent the k-th PID parameter output by the hybrid control model and the target value.

[0096] To ensure that the hybrid control model has strong robustness to important parameters such as feed rate and temperature, a regularization term is added to prevent the hybrid control model from overfitting to noise data. The regularized loss function is:

[0097] ,

[0098] where M is the total number of all weight parameters in the hybrid control model, represents the i-th weight parameter.

[0099] During the training process, the addition of the regularized loss function affects the gradient calculation of the model, thereby automatically adjusting the weights in backpropagation. Regularization reduces the values of high-weight parameters by imposing a smoothing constraint, making the model more stable and less sensitive to noise. For the multi-variable PID control problem, the regularized loss function can improve the generalization performance and robustness of the model, making it more reliable in dynamic feed control and environmental changes.

[0100] In the joint training, the gradients between the LSTM and the BP neural network are updated through the backpropagation algorithm. The gradients of the loss function are passed layer by layer, from the BP layer back to the LSTM layer. The specific gradient calculation steps are as follows:

[0101] 1. Calculate the gradient of the output layer of the BP neural network. At the output layer of the BP network, calculate the partial derivative of the loss function with respect to the PID parameter. Assuming that the output layer is a fully connected layer, the gradient calculation formula is:

[0102] ,

[0103] where, represents the gradient of the loss function with respect to the k-th output neuron of the BP neural network. It is the partial derivative of the error between the current network output and the target value, and represents the key value used to update the network parameters during backpropagation. By calculating this gradient, the model can adjust the weights and biases of the output layer to gradually reduce the loss function, thereby optimizing the output. represents the partial derivative of the loss function with respect to the k-th output , indicating the contribution of the k-th output to the overall loss. It reflects the gap between the current output and the expected value and is used to guide the gradient update in backpropagation. This term is used to adjust the parameters related to the output to make the model output gradually approach the target value .

[0104] 2. Transfer the output layer gradient to the hidden layer of the BP network and calculate the gradient of each neuron in the hidden layer.

[0105] 3. Pass the gradient to the LSTM layer. The output feature vector of the LSTM layer is the input of the BP network. Therefore, the partial derivative of the loss function with respect to is:

[0106] ,

[0107] where represents the weight from the feature vector output by the LSTM to the output layer of the BP network.

[0108] After obtaining the partial derivative of the loss with respect to the LSTM output, continue to backpropagate the gradient through each gate (input gate, forget gate, output gate) of the LSTM structure.

[0109] The gradient calculation process of the LSTM is as follows:

[0110] 1). Calculate the forget gate gradient. The calculation formula is: , where represents the memory cell state at the previous time step. The task of the forget gate is to determine how much information in the memory state at the previous time step is passed to the current time step. Therefore will directly affect the gradient of the forget gate; represents the output value of the forget gate. This is the retention ratio calculated by the forget gate based on the input data at the current time step (value between 0 and 1), indicating how much of the memory at the previous time step needs to be retained; is the derivative of the Sigmoid activation function output by the forget gate. The Sigmoid function outputs a value between 0 and 1, and the derivative part determines the rate of change of the forget gate output, which is very important for the smoothness and stability of gradient propagation, is the partial derivative of the loss function with respect to the LSTM memory cell.

[0111] 2). Calculate the input gate gradient. The calculation formula is: , where represents the candidate memory cell state, which is the candidate memory content determined by the input data at the current time step and the hidden state at the previous time step. The input gate controls how much new information is added to the memory cell. Therefore, the candidate memory state will affect the gradient of the input gate; represents the output value of the input gate. This is the ratio calculated by the input gate, with a value between 0 and 1, indicating how much of the current input is added to the memory cell; The derivative of the Sigmoid activation function for the input gate output. This derivative represents the sensitivity of the input gate to changes in the input data and is used to smooth the increment of the input data. The partial derivative of the loss function with respect to the LSTM memory cell.

[0112] 3) Calculate the output gate gradient, and the calculation formula is: where represents the memory cell state after passing through the output of the activation function. The output gate determines how much information of the memory cell is passed to the hidden state. Therefore, the activation value of the current memory state will affect the gradient of the output gate; represents the output value of the output gate, and the value is between 0 and 1. The output gate controls how much information flows from the memory cell to the hidden state. Therefore, will directly affect the gradient; : This is the derivative of the Sigmoid activation function of the output gate, which is used to smooth the increment of the output of the output gate and ensure the stability of the model during parameter update. is the partial derivative of the loss function with respect to the hidden state.

[0113] After the hybrid control model is trained, the BP neural network learns the non-linear relationship between the material feed ratio and process parameters under different working conditions. After training, the BP neural network will be able to continuously adjust its own weights and biases according to real-time data to optimize the PID controller parameters. Specifically, after the hybrid control model is deployed on the cadmium oxide production line, when the cadmium oxide production line starts to run, the operator first determines the target values according to the production requirements, such as the feed ratio of the material, reaction temperature, pressure, etc., and sets the initial PID parameters according to the standard process conditions. During the operation, the LSTM network extracts the time series characteristics of the sensor data, and the BP neural network continuously receives the time series characteristics output by the LSTM network. The BP neural network performs online learning through these data, identifies the non-linear relationship of the system under different working conditions, and outputs the optimized PID parameters to achieve adaptive dynamic control.

[0114] Such as Figure 3 shown, an embodiment of the present invention discloses an electronic device, including a memory 401 storing executable program code and a processor 402 coupled to the memory 401;

[0115] Among them, the processor 402 calls the executable program code stored in the memory 401 to execute the material feed control method of the cadmium oxide production line described in the above embodiments.

[0116] An embodiment of the present invention also discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the material feeding control method for a cadmium oxide production line described in the above embodiments.

[0117] The purpose of the above embodiments is to exemplarily reproduce and deduce the technical solutions of the present invention, and to completely describe the technical solutions, purposes and effects of the present invention. The purpose is to enable the public to understand the disclosed content of the present invention more thoroughly and comprehensively, and not to limit the protection scope of the present invention thereby.

[0118] The above embodiments are not exhaustive enumerations based on the present invention. In addition, there may be multiple other embodiments not listed. Any replacement and improvement made without violating the concept of the present invention fall within the protection scope of the present invention.

Claims

1. A material feeding control method for a cadmium oxide production line, characterized in that, Including: Collecting sensor data corresponding to various indicators in the production process, where the sensor data is the first sensor data collected at multiple time scales or the second sensor data collected at a single time scale; Fusing the first sensor data to obtain a multi-scale fusion vector for each indicator collected at multiple time scales; Inputting all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain time series features output by the LSTM network; Inputting the time series features into a BP neural network to obtain optimized PID parameters output by the BP neural network; Inputting the optimized PID parameters into a PID controller to optimize and adjust the feeding ratio of the material; The collecting of sensor data corresponding to various indicators in the production process includes: Collecting sensor data corresponding to various indicators in the production process. For each indicator with significant dynamic changes in the production process, the corresponding sensor data is sampled multiple times using short-term and long-term time scales to generate the first sensor data.

2. The material feeding control method of the cadmium oxide production line according to claim 1, wherein After collecting the sensor data corresponding to various indicators in the production process, it further includes: Calculating the correlation between the sensor data and the control target, and adjusting the collection frequency of the sensor and / or adjusting the weight corresponding to the sensor data according to the correlation; And / or Calculating the covariance between the sensor data, and performing data dimensionality reduction and / or adjusting the weight corresponding to the sensor data according to the covariance.

3. The material feeding control method of the cadmium oxide production line according to claim 1, wherein, After collecting the sensor data corresponding to various indicators in the production process, it further includes: Eliminating noise and outliers in the sensor data to obtain processed sensor data; Inputting the processed sensor data into a PID controller to perform a preliminary adjustment on the feeding ratio of the material.

4. The material feeding control method of the cadmium oxide production line according to claim 1, wherein The time series features include features extracted from sensor data corresponding to the feeding rates of cadmium metal, oxygen, and catalyst, as well as features extracted from sensor data corresponding to reactor temperature, reactor pressure, and oxygen concentration; the optimized PID parameters include proportional parameters, integral parameters, and derivative parameters corresponding to cadmium metal, oxygen, and catalyst respectively.

5. The material feeding control method of the cadmium oxide production line according to claim 1, wherein Inputting all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain time series features output by the LSTM network, including: In the LSTM network, using a forget gate to forget the interference of sudden fluctuations on the control parameters during the stable production stage, using an input gate to screen and add data to the memory unit of the LSTM network from the multi-scale fusion vectors and the second sensor data, and using an output gate to output key features at multiple time steps according to the memory unit of the LSTM network to form the time series features.

6. A material feeding control system for a cadmium oxide production line, characterized in that, Including: A sensing layer for collecting sensor data corresponding to multiple indicators in the production process, where the sensor data is the first sensor data collected at multiple time scales or the second sensor data collected at a single time scale; A processing layer for fusing the first sensor data to obtain a multi-scale fusion vector for each indicator collected at multiple time scales; The intelligent optimization layer inputs all the multi-scale fusion vectors and the second sensor data into an LSTM network to obtain the time-series features output by the LSTM network; The time-series features are input into a BP neural network to obtain the optimized PID parameters output by the BP neural network, and the optimized PID parameters are input into the control layer; The control layer is used to optimize and adjust the feeding ratio of the material according to the optimized PID parameters; The execution layer is used to control valves, pumps, and material conveying devices to adjust the feeding ratio of the material according to the control instructions of the control layer; Collect the sensor data corresponding to various indicators in the production process, including: Collect the sensor data corresponding to various indicators in the production process. For each indicator with significant dynamic changes in the production process, the corresponding sensor data is sampled multiple times using short-term and long-term scales to generate the first sensor data.

7. The material feeding control system of the cadmium oxide production line according to claim 6, characterized in that, The processing layer is further used to calculate the correlation between the sensor data and the control target, and adjust the acquisition frequency of the sensor and / or adjust the weight corresponding to the sensor data according to the correlation; And / or Calculate the covariance between the sensor data, and perform data dimensionality reduction and / or adjust the weight corresponding to the sensor data according to the covariance.

8. An electronic device, characterized in that, It includes a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the material feeding control method of the cadmium oxide production line according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program causes a computer to execute the material feeding control method of the cadmium oxide production line according to any one of claims 1 to 5.

Images