Remote control system based on electrolysis workshop and fault prediction method

Abstract

The invention provides a remote control system and a fault prediction method based on an electrolysis workshop, the system comprises an edge gateway, a PLC cluster, a controlled end and the like, the PLC cluster transmits a remote control instruction to the PLC cluster through the edge gateway, and the PLC cluster transmits the remote control instruction to the controlled end through the edge gateway. Relatively dispersed personnel in a workshop can control equipment of a controlled end through a PLC cluster by a remote control system, automation is realized, and the production efficiency is improved to a certain extent; in the data interaction process, the edge gateway can obtain the information related to the controlled end of the PLC cluster and store the information in the database through the server, and the server can carry out fault prediction analysis and preventive maintenance suggestion according to the stored data, so that the equipment maintenance efficiency is improved, and the equipment maintenance cost is reduced; the design of the one-way gatekeeper and the two-way authentication of the one-way gatekeeper improves the data security.

Description

Technical Field

[0001] The present invention relates to the technical field of nonferrous metal processing automation, and in particular to a remote control system based on an electrolysis workshop and a method for predicting equipment failures at a controlled end. Background Art

[0002] Based on the working mode of the existing electrolysis plant, there are many problems that need to be solved urgently. First, the personnel of the electrolysis plant are scattered in various production areas and lack centralized management, which is not conducive to the development of unmanned and less-manned factories. The dispersion of employees means high coordination costs and slow response speed, which makes it difficult to achieve overall optimization. Secondly, although the electrolysis plant has adopted a large number of automated equipment, these equipment still require manual data statistical analysis to make decisions on maintenance or transformation. This decision-making process that relies on manual participation increases labor costs and also affects the level of automation in production, making it impossible to truly achieve unmanned operation. In general, the decentralized personnel management mode and manual equipment decision-making mechanism of the existing electrolysis plant have become bottleneck factors restricting the development of the electrolysis plant towards automation. If these problems cannot be solved in time, it will seriously affect the production efficiency and operating benefits of the electrolysis plant, which is not conducive to the long-term development of the electrolysis industry. Therefore, it is urgent to take effective measures to optimize the production and management mode of the electrolysis plant and promote the transformation and upgrading of the electrolysis industry.

[0003] For example, a remote control method patent disclosed on the China Patent Network has an application number of 201811611353.X. The remote control method of this patent is not suitable for the scenario where the personnel of the electrolysis plant are relatively dispersed. It requires the control end and the controlled end to be connected on the network. It has a high deployment cost and a low degree of automation, and cannot change the topology of the system well. Moreover, the system cannot predict the time when equipment failure will occur, so as to facilitate operators to formulate maintenance plans and reduce maintenance costs. Summary of the invention

[0004] The present invention mainly solves the problem of the efficiency of the existing electrolysis plant personnel in controlling the equipment in a relatively scattered manner and ensures, to a certain extent, the security of the data interaction process between the various modules of the system when it is working. In addition, through the design of the edge gateway and the PLC cluster, the operator can remotely control the equipment on the controlled end, which has solved the problem of the low level of automation in the existing electrolysis plant. Another purpose of the present invention is to solve the problem of high maintenance cost and complicated procedures of the workshop equipment in the existing electrolysis plant.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows: A remote control system, a server, a PLC cluster and a controlled end. The remote control system exchanges data with the equipment on the controlled end through the PLC cluster. The relevant data in the PLC cluster is stored in a database through the server. The server gives production suggestions based on the data in the database.

[0006] Preferably, the production recommendations include prediction of fault events on the controlled end and preventive maintenance on the controlled end; the fault event prediction calculates the time of occurrence of the next fault event based on the frequency difference between adjacent fault events and the time of occurrence of adjacent fault events; the preventive maintenance obtains relevant data of the controlled end equipment through the equipment in the remote control system, and gives preventive maintenance recommendations based on the relevant data.

[0007] Preferably, the equipment at the controlled end includes several anode shaping units, several residual pole washing units and several cathode stripping units; the PLC cluster includes the anode shaping unit PLC, the residual pole washing unit PLC and the cathode shaping unit PLC.

[0008] Preferably, the remote control system sends instructions to the PLC cluster via an edge gateway 1, and controls the equipment at the controlled end via the PLC; The control process of the anode shaping unit is as follows: the remote control system inputs the remote control instructions to the edge gateway one, and the edge gateway one sends the remote control instructions to the anode shaping unit PLC. Through the PLC, the anode shaping unit can be controlled to process the surface of mechanical parts.

[0009] The control process of the cathode stripping machine is as follows: the remote control system inputs the remote control command to the edge gateway one, and the edge gateway one sends the remote control command to the cathode stripping machine unit PLC. Through the PLC, the cathode stripping machine unit can be controlled to strip the metal sheets attached to the cathode plates.

[0010] The control process of the residual pole washing unit is as follows: the remote control system inputs the remote control command to the edge gateway 1, and the edge gateway 1 sends the remote control command to the PLC of the residual pole washing unit. The residual pole washing unit can be controlled by the PLC to clean the residual substances attached to the cathode plate.

[0011] Preferably, the relevant data in the PLC cluster is recorded by edge gateway 2, and edge gateway 2 sends the relevant data of the PLC to the server and stores it in the database through the server.

[0012] Preferably, a single-way network gate is provided between the server and the database. There are two servers, one server is responsible for storing data in the database, and the other server is responsible for making production maintenance recommendations based on the data in the database through relevant algorithms.

[0013] Preferably, two-way authentication is required during the data interaction between the server and the database, that is, mutual identity authentication between the server and the database.

[0014] Preferably, the remote control system includes a hardware control terminal and a monitoring terminal and control items, the hardware control terminal is used to input remote control instructions, and the monitoring terminal can be used for defect detection on the surface of related unit equipment.

[0015] Preferably, the remote control system based on the electrolysis workshop also includes a data management platform.

[0016] The steps of the fault event prediction method are as follows: S1: Input the fault events of the controlled device and the date of occurrence of the fault events into the server through the data management platform.

[0017] S2: The server searches for the daily working frequency of the anode shaping unit on the corresponding date according to the acquired fault event and its date. If the fault event has no adjacent fault event, the fault event is stored in the database. If so, step S3 is executed.

[0018] S3: The server searches for the daily work statistical frequency and time of the adjacent fault events based on the acquired fault event and its date.

[0019] S4: The server calculates the difference between the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events and the difference in occurrence time based on the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events.

[0020] S5: The server calculates the next occurrence of a fault event of the controlled device through a relevant algorithm based on the difference in daily working frequency between the fault event and the adjacent fault event and the difference in occurrence time, and records it in the database.

[0021] The beneficial effects of the present invention are as follows: centralized remote control of relevant equipment in the electrolysis workshop is realized with a high degree of automation; and production maintenance suggestions are given by collecting relevant data of the PLC cluster through relevant algorithm analysis, which makes the cost of maintaining the equipment of the system low; and the design of the one-way valve and the encryption authentication module makes the system more secure; the design of the PLC cluster and the edge gateway can meet the real-time nature of non-ferrous metal processing to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The figure is a schematic diagram of a remotely controlled controlled terminal of the present invention.

[0023] Figure 2 This is a fault event prediction flow chart of the present invention.

[0024] Figure 3 The present invention is a controlled terminal defect detection flow chart. DETAILED DESCRIPTION

[0025] In order to better illustrate the technical means used in the present invention, the technical means used in the present invention are described in detail below through some embodiments.

[0026] like Figure 1 As shown, the flowchart of the remote control of the controlled end of the present invention shows the entire remote control process in detail. First, the operator enters specific control instructions through the hardware port of the remote control system (such as a touch screen, keyboard, or dedicated console). These instructions may include various operating commands such as equipment start, stop, parameter adjustment, etc. Subsequently, the remote control system encodes and encrypts these control instructions to ensure transmission security, and then sends the instructions to the edge gateway through the network. As the data processing and forwarding center of the system, the edge gateway parses and verifies the instructions after receiving them, and then safely transmits the processed remote control instructions to the PLC (programmable logic controller) cluster.

[0027] As the core of the industrial control system, the PLC cluster immediately controls the controlled device or system according to the preset program logic after receiving the verified remote control command. This may involve multiple PLCs working together to achieve complex control functions. After the controlled end executes the relevant remote control command, the PLC cluster will immediately collect relevant information such as the operating status and parameter changes of the controlled end. This information may include key indicators such as equipment operating status, production parameters, and energy consumption data. The collected information is then packaged by the PLC cluster and transmitted to the edge gateway through a secure channel.

[0028] Edge Gateway 1, as the main data processing center, performs preliminary data analysis and collation after receiving this information, and then transmits the processed information back to the remote control system. This enables operators to monitor the execution and operating status of the controlled end in real time and achieve closed-loop control. At the same time, in order to ensure redundant backup of data and further analysis and utilization, Edge Gateway 2 will also obtain relevant information collected by the PLC cluster in parallel.

[0029] Edge Gateway 2 is connected to a dedicated server and is responsible for storing these important data in a designated database. This dual data flow design not only improves the reliability of the system, but also provides a basis for subsequent predictive maintenance.

[0030] When the operator controls the anode shaping unit through remote control instructions to perform ear pressing, shaping, weighing, rejecting, bottom milling, oblique lifting and other operations on the crude copper anode plate, the relevant operator needs to input relevant instructions for the anode shaping unit through the hardware control terminal in the remote control system, and the remote control system transmits the relevant instructions to the edge gateway one. After receiving the relevant remote control instructions, the edge gateway one transmits the instructions to the anode shaping unit PLC, and the anode shaping unit PLC converts the instructions into machine language according to the received remote control instructions to control the anode shaping unit, so that the anode shaping unit can process the crude copper anode plate according to the sent remote control instructions. In the anode shaping unit, the anode plate realizes automatic shaping, detection, spacing and transportation.

[0031] After the operation is completed, the anode shaping unit PLC obtains the relevant data of the anode shaping unit and uploads the relevant data to edge gateway one. At the same time, the relevant data of the anode shaping machine obtained by the anode shaping unit PLC will be obtained by edge gateway two, and edge gateway two will transmit the relevant data to the server. The server stores the data in a database to facilitate subsequent production suggestion related analysis; after the relevant data is uploaded to edge gateway one, edge gateway one uploads the relevant data to the remote control system, processes the data and displays it on the PC at the hardware control end of the remote control system, so as to facilitate the relevant operators to obtain the relevant information of the anode shaping unit to facilitate subsequent production operations.

[0032] The operator controls the cathode stripping unit through the remote control system. During the manufacturing process of electrolytic refined copper, the crane transports the shaped anode plate to the electrolytic cell. In the electrolytic cell, through the redox reaction, the copper in the anode plate dissolves and precipitates on the cathode plate to form high-purity cathode copper. After one cycle, the crane takes the cathode copper out of the electrolytic cell and puts it into the stripping unit for stripping, inspection, packaging, and weighing, and then puts the separated cathode plate back into the electrolytic cell. After two cycles, the anode plate becomes a residual electrode plate. The crane pulls the residual electrode plate to the residual electrode unit for washing, weighing, and packaging, and then sends it back for smelting. The operator inputs relevant instructions for the cathode stripping unit through the hardware end of the remote control system, and the remote control system transmits the relevant instructions to the edge gateway. After receiving the relevant remote control instructions, the edge gateway sends the instructions to the cathode stripping unit PLC. The cathode stripping unit PLC converts the instructions into machine language instructions to control the cathode stripping unit, so that the cathode stripping unit can strip off the high-purity copper deposited on the cathode plate, and at the same time remove the high-purity copper that does not meet the standards, and the qualified high-purity copper is weighed and packaged.

[0033] After the operation is completed, the cathode stripping unit PLC obtains the relevant data of the anode shaping unit and uploads the relevant data to the edge gateway one. At the same time, the relevant data of the anode shaping machine obtained by the anode shaping unit PLC will be obtained by the edge gateway two, and the edge gateway two will transmit the relevant data to the server. The server stores the data in the database to facilitate the subsequent production suggestion related analysis; after the relevant data is uploaded to the edge gateway one, the edge gateway one uploads the relevant data to the remote control system, processes the data and displays it on the PC of the hardware control end of the remote control system, so as to facilitate the relevant operators to obtain the relevant information of the cathode stripping unit to facilitate the subsequent production operations.

[0034] The operator controls the anode scrap washing unit through the remote control system to clean the anode mud on the scraps, and then re-produce the precious metal mixture on the anode scraps to reduce the residual anode mud on the anode scraps. In the manufacturing process of electrolytic refined copper, the crane transports the shaped anode plate to the electrolytic cell. In the electrolytic cell, through the redox reaction, the copper in the anode plate dissolves and precipitates on the cathode plate to form high-purity cathode copper. After one cycle, the crane takes the cathode copper out of the electrolytic cell and puts it into the stripping unit for stripping, inspection, packaging, weighing and other operations, and then puts the separated cathode plate back into the electrolytic cell. After two cycles, the anode plate becomes a scrap plate, and the crane pulls the scrap plate to the scrap unit for washing, weighing, packaging, and then returns it to the smelting. The operator inputs the relevant instructions of the cathode stripping unit through the hardware end of the remote control system, and the remote control system transmits the relevant instructions to the edge gateway. After receiving the relevant remote control instructions, the edge gateway sends the instructions to the PLC of the residual pole washing unit. The PLC of the residual pole washing unit converts the instructions into machine language instructions to control the residual pole washing unit, so that the residual pole unit can transport, rinse, flip, stack, weigh and pack the residual pole plates.

[0035] After the operation is completed, the PLC of the residual pole washing unit obtains the relevant data of the residual pole washing unit and uploads the relevant data to the edge gateway one. At the same time, the relevant data of the anode shaping machine obtained by the PLC of the anode shaping unit will be obtained by the edge gateway two, and the edge gateway two will transmit the relevant data to the server. The server stores the data in the database to facilitate the subsequent production suggestion related analysis; after the relevant data is uploaded to the edge gateway one, the edge gateway one uploads the relevant data to the remote control system, processes the data and displays it on the PC of the hardware control end of the remote control system, so as to facilitate the relevant operators to obtain the relevant information of the residual pole washing unit to facilitate the subsequent production operations.

[0036] like Figure 2As shown in the fault event prediction flow chart of the present invention, in the normal remote control process, the PLC cluster will also count the hourly working frequencies of several anode shaping units, several cathode stripping units and several residual pole washing units, and summarize the hourly working frequencies of the anode shaping units, cathode stripping units and residual pole washing units into the daily working frequencies of the anode shaping units, cathode stripping units and residual pole washing units, wherein the hourly working frequencies of the PLC cluster will be reset after being summarized into the daily working frequencies, so that the daily working frequencies counted by the PLC cluster will change for 24 hours every day.

[0037] For example, starting from 0:00 on a certain day, taking the anode shaping unit as an example, the working frequency of the anode shaping unit from 0:00 to 1:00 is 391 times. At this time, the initial values ​​of the hourly working statistical frequency and daily working statistical frequency of the PLC of the anode shaping unit are set to 0, and the hourly working frequency of the PLC of the anode shaping unit is 391 times. After the statistics are completed, the daily working frequency value of the anode shaping unit PLC is added with 391. After the operation is completed, the hourly working frequency value of the anode shaping unit is set to 0, and the working frequency of the anode shaping unit from 1:00 to 2:00 is continued to be obtained. After these operations, if the working frequency of the anode shaping machine from 1:00 to 2:00 is 261, the hourly working statistical frequency of the anode shaping unit of the anode shaping unit PLC is 261, and the daily working statistical frequency of the anode shaping unit PLC is added with 261. At this time, the anode shaping unit PLC The daily working statistical frequency of C is 261 plus 391, which is 652 times. After these operations are completed, the hourly working statistical frequency of the anode shaping unit PLC is set to 0, and the working frequency of the anode shaping unit from 2 o'clock to 3 o'clock is continued to be obtained, and so on to the working frequency of the anode shaping unit from 23 o'clock to 24 o'clock. After the anode shaping unit PLC obtains the working frequency of the anode shaping unit from 23 o'clock to 24 o'clock, the daily working statistical frequency of the anode shaping unit PLC is added to the hourly working statistical frequency of the anode shaping unit PLC corresponding to 23 o'clock to 24 o'clock. After the operation is completed, the hourly working statistical frequency of the anode shaping unit PLC is set to 0, and the daily working statistical frequency of the anode shaping unit PLC is collected by the edge gateway 2, and the edge gateway 2 records the date corresponding to the daily working statistical frequency of the anode shaping unit, and stores these data in the database.

[0038] Fault prediction is performed based on the daily working statistical frequency of the anode shaping unit PLC mentioned above. The steps of fault prediction are as follows (taking the anode shaping unit as an example): S1: The detected fault events of the anode shaping unit and the dates of occurrence of the fault events are input into the server through the data management platform.

[0039] S2: The server searches for the daily working frequency of the anode shaping unit on the corresponding date according to the acquired fault event and its date. If the fault event has no adjacent fault event, the fault event is stored in the database. If so, step S3 is executed.

[0040] S3: The server searches for the daily work statistical frequency and time of the adjacent fault events based on the acquired fault event and its date.

[0041] S4: The server calculates the difference between the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events and the difference in occurrence time based on the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events.

[0042] S5: The server calculates the next fault event of the anode shaping unit according to the difference in daily working frequency between the fault event and the adjacent fault event and the difference in occurrence time through a relevant algorithm, and records it in the database.

[0043] Through the prediction of fault events, the time of the next fault event of the anode shaping unit can be predicted. Taking into account that the existing anode shaping unit technology is relatively mature and not easy to damage, according to the daily work statistical frequency of the anode shaping unit PLC, it can be known that the time of the fault event is in units of daily units. In this way, the relevant operators can prepare for maintenance and inspection in this time period according to the time of the next fault event given by the server.

[0044] In the field of industrial equipment management and predictive maintenance, the accuracy of fault event prediction is crucial to the efficiency and economy of the entire production system. Accurate fault prediction time can not only optimize maintenance plans, but also minimize equipment downtime, thereby improving production efficiency. However, if the time error of fault event prediction is too large, it will bring a series of negative effects. First, premature prediction may lead to unnecessary downtime maintenance, waste valuable production time, and increase unnecessary manpower and material costs. On the contrary, if the prediction time is too late, the best maintenance window may be missed, resulting in sudden failure of equipment during production, causing greater losses and longer downtime. In addition, inaccurate prediction may also cause misconfiguration of maintenance resources, such as early deployment of personnel, ordering spare parts, etc., which not only increases inventory costs, but also may affect the normal maintenance plans of other equipment. In the long run, frequent prediction errors will reduce the credibility of the prediction system and affect decision makers' confidence in predictive maintenance strategies. Therefore, in order to ensure the efficient operation of the production system, minimize maintenance costs, optimize resource allocation and maximize equipment utilization, certain requirements are put forward for the accuracy of fault event prediction time.

[0045] In some embodiments, the method by which the server calculates the time between the next accident and the next accident based on the difference in daily work statistical frequency and the difference in occurrence time is as follows: the server obtains the time of occurrence of all previous fault events and the daily work statistical frequency corresponding to the fault events from the training data, obtains the time of occurrence of the fault event adjacent to each fault event and the corresponding daily work frequency based on these data, and subtracts them from them to obtain the difference in daily work statistical frequency between the fault event and the adjacent fault event, and further, the time difference is obtained by subtracting the time of occurrence of the adjacent fault event from the time of occurrence of the fault event. It should be noted that the obtained time is a positive value and cannot be a negative number. At the same time, the time interval between the next fault event and the fault event is obtained. It should also be noted that the time interval value between the next fault event and the fault event is positive and cannot be negative, otherwise the prediction loses its meaning; because the scale of its training data is large, the same time difference (with daily as the statistical unit) and the same daily work frequency difference may appear, and the corresponding next fault event and the fault event have different values ​​of difference. For this, an example is taken as an example: If the time difference is 2, and the daily working frequency difference is 70, and the corresponding next fault event intervals are 11, 13, 12, 9, and 8, respectively, when processing these data, remove the maximum value of the fault event interval 13 and the minimum value of the fault event interval 8. After removing these values, take the average value of these data as the next fault event interval when the time difference is 2 and the daily working frequency difference is 70, and its value is 10.67 (keep two decimal places); Therefore, for the same difference in the time of occurrence of a fault event and the difference in the daily working frequency of a fault event, the interval to the next fault event is the weighted average value after removing the maximum and minimum values.

[0046] The difference between the time when the fault event occurs and the difference between the daily working frequencies of the fault events are taken as feature data, and the time interval to the next fault event is taken as the function value. These data are fitted by kernel function regression.

[0047] Kernel function regression is mainly used to deal with the nonlinear relationship of these data. In most natural data, the distribution of many data sets in their feature space does not show a linear relationship, but more of a nonlinear relationship. The kernel function maps the feature data to a high-dimensional space and fits these data by finding the linear mapping of the data in the high-dimensional space. From the actual training data, if the difference between the daily work statistical frequency and the difference between the fault time and the fault occurrence time is used as the feature data, and the average value of the time interval between the fault events is used as the label value, these data mainly show a nonlinear relationship. At this time, the kernel function is mainly used for nonlinear regression. Here, the selection of the kernel function is also related to the performance of fault event prediction.

[0048] In the selection of kernel functions, commonly used kernel functions include: linear kernel function and Gaussian kernel function, among which: The linear kernel is a simple and efficient kernel function that is widely used in machine learning. Its calculation process is simple and clear, and only requires a dot product operation on the input vector, which makes the linear kernel model have a fast calculation speed and good interpretability. In addition, in many practical problems, the data presents the characteristics of linear separability, and the linear kernel function can effectively capture the key features of the data, thereby showing a strong generalization ability.

[0049] However, the linear kernel also has some limitations. Since it can only describe linear relationships, the linear kernel's ability to express complex nonlinear patterns is limited. At the same time, the linear kernel is also sensitive to noisy data, which may reduce the robustness of the model. In addition, in order to give full play to the advantages of the linear kernel, meticulous feature engineering is required, which brings certain challenges to the modeling process.

[0050] The Gaussian kernel function is a common nonlinear kernel function in machine learning, which has rich expression ability and good modeling performance. Compared with the linear kernel, the Gaussian kernel has many advantages: First, the Gaussian kernel can effectively capture the complex nonlinear relationship between input features. By introducing the exponential form of the Gaussian function, the Gaussian kernel can flexibly model various complex patterns, thereby greatly enhancing the expressive power of the kernel method. This enables the Gaussian kernel to often achieve excellent performance when facing problems with complex nonlinear structures, such as image classification, speech recognition and other application fields.

[0051] Secondly, the Gaussian kernel has strong robustness. Unlike the linear kernel, which is more sensitive to noisy data, the Gaussian kernel function can suppress the influence of outliers or noise to a certain extent. This is because the Gaussian kernel uses the similarity between input features. For outliers that are far apart, their kernel values ​​will quickly approach 0, greatly reducing their role in model training and prediction. This makes the Gaussian kernel method perform well in practical problems with noise interference.

[0052] In addition, the Gaussian kernel also has a certain regularization effect. The Gaussian kernel function implicitly constrains the model complexity to a certain extent, which helps to prevent overfitting and improve the generalization performance of the model. At the same time, the kernel matrix of the Gaussian kernel often has good numerical characteristics, which is conducive to numerical optimization solutions. These characteristics make the Gaussian kernel method often achieve satisfactory results in practical applications.

[0053] Of course, the Gaussian kernel function also has some limitations. First, the Gaussian kernel involves the adjustment of kernel function parameters, such as the selection of kernel width parameters, which brings certain difficulties to users. Improper parameter selection may seriously affect the performance of the model. In addition, the computational complexity of the Gaussian kernel increases rapidly with the increase in the number of samples, and the computational efficiency on large-scale data sets may be limited. This often requires the use of some approximation techniques, such as sparse approximation and low-rank approximation, to improve computational efficiency.

[0054] Another limitation is that the Gaussian kernel method is highly sensitive to feature scaling. The dimensionality difference of the input features will significantly affect the calculation results of the Gaussian kernel function, so careful feature preprocessing is usually required when using the Gaussian kernel. This makes the Gaussian kernel method more complex to model to a certain extent.

[0055] In this embodiment, a new kernel function is constructed to combine the characteristics of the polynomial kernel function that can capture local features and the Gaussian kernel function that can capture local features.

[0056] The reason why the Gaussian kernel is more sensitive to local features is mainly due to its computational mechanics and the characteristics of feature space mapping. The calculation of the Gaussian kernel function depends on the Euclidean distance between the input vectors. The closer the two input vectors are, the larger the kernel value is. This means that the Gaussian kernel pays more attention to the local similarity between adjacent samples and has less impact on sample points that are farther away.

[0057] This local sensitivity stems from the implicit feature space mapping of the Gaussian kernel. The Gaussian kernel maps the input space into an infinite-dimensional feature space, where each dimension corresponds to a certain local variation pattern of the input features. Therefore, the Gaussian kernel can effectively capture the local correlation and complex relationship between the input features, thereby better characterizing the local nonlinear patterns. At the same time, the Gaussian kernel also implicitly smoothes and regularizes the model, further enhancing its focus on local features.

[0058] In contrast, the polynomial kernel function pays more attention to global features than the Gaussian kernel function. The linear kernel only models the linear combination of the input vectors and cannot well characterize the local nonlinear relationship. Although the linear kernel can achieve good performance in some linearly separable problems, its expressive power is often limited for complex nonlinear problems. This is because the polynomial kernel function cannot capture the complex interactions between the input features compared to the Gaussian kernel function, but can only focus on the overall linear correlation.

[0059] This embodiment mainly constructs the kernel function by using the standard orthogonal basis and the Gaussian kernel function. This method fully combines the advantages of the orthogonal basis and the sensitivity of the Gaussian kernel function to local features, and can effectively enhance the modeling ability of the kernel method.

[0060] Specifically, the standard orthogonal basis can provide a complete and independent feature space representation, which helps to capture the global characteristics of the input data. The Gaussian kernel function is good at characterizing local similarities and can flexibly model complex nonlinear patterns. By combining these two techniques, we can construct a kernel function that takes into account both global and local characteristics, thereby greatly improving the performance of kernel methods in practical problems.

[0061] The key to the design of this hybrid kernel function lies in the reasonable integration of the advantages of the orthogonal basis and the Gaussian kernel. On the one hand, the orthogonal basis can provide a compact and numerically stable feature representation, laying the foundation for the subsequent Gaussian kernel calculation. On the other hand, the Gaussian kernel can assign different weights to the eigenvectors of the orthogonal basis expansion, effectively capturing local correlations and nonlinear patterns. In this way, we can construct a hybrid kernel function that has both global modeling capabilities and local sensitivity, which plays an important role in various complex problems.

[0062] The standard orthogonal polynomials are: Jacobi orthogonal polynomials, Jacobi orthogonal polynomials are defined on (-1,1) with respect to the weight function (1-x) a (1+x) b Orthogonal polynomials.

[0063] Legendre polynomials, Legendre polynomials are orthogonal polynomials defined on the interval (-1,1) about the weight function 1. In fact, Legendre orthogonal polynomials are functions (1-x) a (1+x) b , the special case of a=0 and b=0.

[0064] Chebyshev orthogonal polynomials are defined on the interval (-1,1) about the weight function Orthogonal polynomials.

[0065] Laguerre polynomial, Laguerre polynomial is defined on the interval (0, +∞) about the weight function e -x x a Orthogonal polynomials.

[0066] Hermitian orthogonal polynomials, Hermitian orthogonal polynomials are polynomials defined on the interval (-∞, +∞) that are orthogonal to the weight function.

[0067] According to the above orthogonal basis, the feature data x can be mapped to Its kernel function can be obtained The kernel function matrix K can be obtained, where the kernel function matrix K is: The matrix is: This matrix is ​​positive definite.

[0068] The above orthogonal basis can better capture the global features of feature data than the Gaussian kernel function. The orthogonal basis is a complete and independent feature space representation, which can provide a compact and stable global feature encoding for the input data. This global feature encoding can effectively summarize the overall statistical characteristics of the data and lay a solid foundation for subsequent modeling tasks.

[0069] In contrast, the Gaussian kernel function pays more attention to local features. The calculation of the Gaussian kernel depends on the Euclidean distance between the input vectors. The closer the two input vectors are, the larger the kernel value is. This means that the Gaussian kernel pays more attention to the local similarity between adjacent samples and has less impact on sample points that are farther away. This local sensitivity makes the Gaussian kernel advantageous in modeling local complex patterns, but its ability to capture global features may be limited to a certain extent.

[0070] However, in actual data modeling, we should not only pay attention to the integrity of the data, but also the locality of the data. These two aspects are the key to modeling and are indispensable.

[0071] Although it was mentioned above that the orthogonal basis can better capture the global features of the data, while the Gaussian kernel is good at describing local similarities. However, actual problems are often more complex, and a single global modeling or local modeling method is difficult to fully meet the needs. For example, in image recognition and other problems, global features such as overall texture, shape and other information are important, but local features such as edges, corners and other details are also indispensable. Only focusing on global features may ignore some key local patterns, while only focusing on local features cannot grasp the overall information.

[0072] Therefore, in actual modeling, we need to take into account both the overall and local characteristics of the data and seek a balance between the two. This can be achieved by designing a hybrid kernel function, that is, organically combining the global feature encoding of the orthogonal basis with the local sensitivity of the Gaussian kernel to give full play to the advantages of both technologies. At the same time, we also need to flexibly adjust the weights of global and local features according to the characteristics of the specific problem. Sometimes global features dominate, and sometimes local features are more critical. Only by fully considering the characteristics of actual data and balancing the global and local aspects can we build a kernel method model with excellent performance.

[0073] Therefore, combining the Gaussian kernel with the orthogonal polynomial kernel enables the regression model of this embodiment to capture both the overall characteristics of the data and the local characteristics of the data, thereby enabling the regression model to better predict the interval for the next fault event to occur.

[0074] The following properties of the kernel function are the key to constructing the kernel function in this embodiment: The new kernel function formed by adding two kernel functions, the kernel function matrix K formed by the new kernel function is a semi-positive definite matrix; the new kernel function formed by multiplying two kernel functions, the kernel function matrix K formed by the new kernel function is a semi-positive definite matrix; the kernel functions constructed by orthogonal polynomials of different orders are multiplied by the corresponding weights and multiplied with the Gaussian kernel function, and the constructed kernel functions are added; Its expression is where ω i >0, and In this embodiment, in order to simplify the calculation, the order of n is 3; The test results of different orthogonal polynomials for test data are shown in the following table (before optimizing the parameters): Kernel Function Average error of test data Maximum error of test data Jacobi-Gaussian kernel function 1.1376 1.3752 Legendre Gaussian kernel function 1.7631 2.1112 Chebyshev Gaussian kernel function 1.1321 1.4112 Laguerre-Gaussian kernel function 0.6789 0.9651 Hermitian Gaussian kernel function 0.5416 0.6771 The test results of different orthogonal polynomials for test data are shown in the following table (after optimizing the parameters): Kernel Function Average error of test data Maximum error of test data Jacobi-Gaussian kernel function 0.9267 1.2871 Legendre Gaussian kernel function 1.0012 1.5312 Chebyshev Gaussian kernel function 1.4312 1.7231 Laguerre-Gaussian kernel function 0.6001 0.9231 Hermitian Gaussian kernel function 0.4416 0.5187 It can be seen from the above experimental data that the kernel function matrix constructed by Hermite orthogonal polynomials and Gaussian kernel functions has better prediction performance, and the error is lower than that of Jacobi Gaussian kernel function, Legendre Gaussian kernel function, Chebyshev Gaussian kernel function and Laguerre Gaussian kernel function. After testing the test data, the kernel function used in this embodiment is the kernel function constructed by Hermite orthogonal polynomials and Gaussian kernel functions as the kernel regression function.

[0075] After the regression model is trained, the model is deployed in the server. Since the positive shaping unit, the cathode stripping unit and the residual anode washing unit will continuously generate new data during the working process, the database will generate new data after the unit works for one day. As the working date increases, the scale of the data will become larger and larger. If training is performed based on these data at this time, there is no doubt that the burden on the server will increase, and as the data scale increases, the data processing process will undoubtedly increase; In order to solve the above problem, in this embodiment, it is first set that when the size of the server data set increases to a certain extent compared with the previous data set, the model will be retrained according to the new data set, and a new model will be obtained to predict the relevant data. The new data set is larger than the previous data set, so the server will randomly sample the data in the database as the training data of the kernel function regression model, and the number of random samples is set according to the current data size: It should be noted here that n_old is the size of the last database training data, n_new is the size of the current database training data, and the number of random samples is set to:

[0076] In some embodiments, the method by which the server calculates the time between the next accident and the next accident based on the difference in daily work statistical frequency and the difference in occurrence time is as follows: the server obtains the time of occurrence of all previous fault events and the daily work statistical frequency corresponding to the fault events from the training data, obtains the time of occurrence of the fault event adjacent to each fault event and the corresponding daily work frequency based on these data, and subtracts them from them to obtain the difference in daily work statistical frequency between the fault event and the adjacent fault event, and further, the time difference is obtained by subtracting the time of occurrence of the adjacent fault event from the time of occurrence of the fault event. It should be noted that the obtained time is a positive value and cannot be a negative number. At the same time, the time interval between the next fault event and the fault event is obtained. It should also be noted that the time interval value between the next fault event and the fault event is positive and cannot be negative, otherwise the prediction loses its meaning; because the scale of its training data is large, the same time difference (with daily as the statistical unit) and the same daily work frequency difference may appear, and the corresponding next fault event and the fault event have different values ​​of difference. For this, an example is taken as an example: If the time difference is 13, and the daily working frequency difference is 200, and the corresponding next fault event intervals are 21, 24, 18, 15, and 19 respectively, when processing these data, first calculate the average value, the average value of these data is 19.4, and the difference between the data and the average value is (take the absolute value): 1.6, 4.6, 1.4, 4.4, 0.4, and define the weight function; The weight function expression is: where d i is the absolute value of the difference between the i-th data and the average value of the data it belongs to, and d_max is the largest data in the data set; The time interval until the next fault event corresponding to the characteristic data occurs is: The difference between the time when the fault event occurs and the difference between the daily working frequencies of the fault events are taken as characteristic data, and the time interval between the fault events is taken as the function value. These data are fitted through the BP neural network.

[0077] The advantages of using BP neural network for failure time prediction are as follows: First, BP neural network has powerful nonlinear modeling capabilities. The occurrence of fault events is often affected by a variety of complex factors, including the performance of the equipment itself, the use environment, maintenance, etc. This complex internal mechanism is difficult to accurately describe with a simple linear model. BP neural network can effectively capture these complex input-output mapping relationships through multi-layer nonlinear transformations, thereby better fitting the nonlinear problem of fault time prediction.

[0078] Secondly, the BP neural network has good adaptive learning capabilities. During the training process, the BP network can automatically adjust internal parameters, such as weights and biases, to minimize prediction errors and continuously approach the optimal prediction model. This adaptive learning feature enables the BP network to flexibly model different fault data sets without excessive manual intervention. For new fault modes or environmental changes, the BP network can also quickly adapt by continuing to learn.

[0079] In addition, BP neural network has strong generalization ability. By learning a large amount of fault data, BP network can capture the general law of fault occurrence, rather than being limited to specific fault samples. This means that the trained BP model has good transferability and can be applied to predict new fault events, thereby improving the wide applicability of fault prediction.

[0080] In summary, BP neural network has obvious advantages in the problem of fault time prediction due to its powerful nonlinear modeling ability, adaptive learning characteristics, good generalization and fault tolerance for complex data. Compared with traditional statistical analysis methods, BP network can better describe the complex mechanism of fault occurrence, thereby improving the accuracy and stability of prediction. Therefore, applying BP neural network to fault time prediction is undoubtedly a very effective modeling method.

[0081] The activation functions used by BP neural network are: The Sigmoid activation function has a domain from negative infinity to positive infinity, and a value range from 0 to 1, which is an open interval. As the domain value increases, the function graph gradually approaches 1. If the domain tends to negative infinity, the function tends to 0.

[0082] The Tanh activation function has a domain from negative infinity to positive infinity, a value range from -1 to 1, and as the domain value increases, the function graph gradually approaches 1. If the domain tends to negative infinity, the function tends to -1.

[0083] The ReLU activation function has a domain of negative infinity and a value range of 0 to positive infinity.

[0084] The Softsign activation function has a domain from negative infinity to positive infinity, a value range from -1 to +1, and as the domain value increases, the function graph gradually approaches 1. If the domain tends to negative infinity, the function tends to -1.

[0085] The PReLU activation function has a domain from negative infinity to positive infinity and a value range from positive infinity to negative infinity.

[0086] The ELU activation function has a domain from negative infinity to positive infinity and a range from -a to positive infinity. In particular, if the domain tends to negative infinity, the function tends to -a.

[0087] The Swish activation function has a domain from negative infinity to positive infinity, a range from 0 to positive infinity, and if the domain tends to negative infinity, the function tends to 0.

[0088] The Mish activation function has a domain from negative infinity to positive infinity, a range from 0 to positive infinity, and if the domain tends to negative infinity, the function tends to 0.

[0089] The Softmax activation function has a domain from negative infinity to positive infinity, a value range from 0 to 1, and as the domain value continues to increase, the function graph gradually approaches 1. If the domain tends to negative infinity, the function tends to 0.

[0090] The test results of different activation functions for test data are as follows: Activation Function Average error of test data Maximum error of test data Sigmoid activation function 1.9172 3.8177 Tanh activation function 2.3265 2.9001 ReLU activation function 0.7651 1.339 Softsign activation function 1.5161 3.7651 PReLU activation function 1.5976 2.6339 ELU activation function 0.7181 1.0089 Swish activation function 2.3991 4.6720 Mish activation function 0.6128 0.8543 According to the above test results, this embodiment selects the Mish activation function as the activation function of the BP neural network, and predicts the time when the next fault event will occur based on the trained BP neural network.

[0091] In some embodiments, the server will comprehensively consider the results of training different models to improve the accuracy and reliability of fault event time prediction. For example, the frequency difference between a fault event and an adjacent fault event is 61, and the daily work statistical frequency difference is 921. The server is provided with several trained fault event time interval prediction models, namely model 1, model 2 and model 3. After inputting these feature data, the output of model 1 is 42 days, the output of model 2 is 39 days, and the output of model 3 is 41 days. In this embodiment, the server will comprehensively consider the prediction results of the three models and take the weighted average of their outputs. If the weights of each model are equal, the time interval for the next fault event to occur is (42+39+41) / 3=40.67 days, which is the time interval for the next fault event to occur.

[0092] In addition to the simple average fusion method, the fusion weight can be dynamically adjusted according to the prediction performance of each model, or more complex fusion mechanisms such as voting and stacking can be adopted to further improve the accuracy and robustness of failure time prediction.

[0093] Specifically, different fusion weights can be dynamically assigned to each model based on its prediction effect on historical data. Models with better prediction performance will receive higher weights, while models with poorer performance will receive relatively lower weights. This not only fully utilizes the advantages of excellent models, but also reduces the impact of inferior models on the final prediction results.

[0094] The calculation of dynamic weights can be based on performance indicators such as mean absolute error (MAE) and root mean square error (RMSE). For a model m, its prediction performance indicator P on the validation set is m As the weight w of the model m The calculation is as follows: Its expression is: Where i represents the model index, is the sum of all model performance indicators, which plays a normalization role. Obviously, the better the prediction performance of the model, the higher its weight w m The bigger it is.

[0095] In addition, more complex multi-model fusion strategies such as voting fusion and stacked generalization are also attempted in some embodiments.

[0096] Voting fusion refers to voting on the prediction results of each model and taking the prediction value with the most votes as the final output. This method is more robust to abnormal values ​​or outliers and can suppress the extreme prediction of a single model to a certain extent.

[0097] Stacked generalization is a fusion mechanism based on meta-learning. First, several base learner models are trained, and then their prediction outputs are used as new features to train a meta-model to learn the final prediction. This method can explore the relationship between base learners, thereby obtaining a more sophisticated fusion result.

[0098] This prediction method based on multi-model fusion has many advantages. First, it can give full play to the advantages of different models and improve the accuracy and reliability of prediction. Each model has its own unique modeling characteristics and strengths. By averaging their prediction results, the limitations of a single model can be effectively eliminated, thus obtaining a more robust prediction.

[0099] For example, in the above example, Model 1 may be better at capturing the impact of the frequency difference feature on the failure time, while Model 2 may be more accurate in modeling the daily working frequency difference feature, and Model 3 may take into account the combined impact of both. By integrating the prediction results of the three models, we can fully absorb the advantages of each model and obtain a more comprehensive and reliable prediction.

[0100] Secondly, the multi-model fusion method can also improve the stability of prediction. The prediction of a single model is easily affected by factors such as specific data and parameter settings, and there is a certain degree of volatility. However, by fusing the outputs of multiple models, the fluctuation range of the prediction results can be greatly reduced, making the prediction more stable and reliable. This is crucial for fault warning of key equipment, helping to reduce false alarms and missed alarms and improve the effectiveness of warnings.

[0101] In some embodiments, the defect detection method of the monitoring end for the controlled end is: like Figure 3 A controlled terminal defect detection process of the present invention is shown as follows: S11: Acquire a texture image of the surface of the component at the controlled end through a high-definition camera at the monitoring end.

[0102] S12: Convert the acquired texture image of the component surface into a grayscale image, denoise the grayscale image, and calculate the contrast of the image. If the contrast of the image reaches the set threshold, execute step S14; if the contrast of the image is less than the set threshold, execute step S13, and recalculate the image contrast after executing step S13.

[0103] S13: performing gamma transformation on the grayscale image. The purpose of this transformation is to make the pixel change of the defective part in the image larger than that of the normal part of the controlled end.

[0104] S14: Based on the obtained image, a point on the obtained image is regarded as a node of the graph, a weight between two nodes is calculated based on pixels of the node and other nodes and a positional relationship between the two nodes, and a graph is constructed.

[0105] If a node of the graph calculates the weight between the two points with all the points, then taking an image with a pixel size of 1000*1000 as an example, if it is an image with a size of n*m, the time complexity of the constructed graph is: o(n*n*m*m); The time complexity of constructing a graph with an image size of 1000*1000 is too high. Therefore, in actual production, the status of the controlled end components cannot be quickly known to meet the real-time nature of production. If we calculate the weight of a node only with the nodes adjacent to it, for example, 4 nodes are adjacent to it; Then the time complexity of constructing a graph based on the image is: o(n*m), which is an order of magnitude faster than the above construction method. Therefore, in this embodiment, for a certain node, its weight value is mainly calculated with adjacent nodes that meet certain conditions.

[0106] S15: Segmenting the image according to the image constructed from the grayscale image to segment defective parts and non-defective parts in the grayscale image.

[0107] The following is a detailed description of a method for segmenting a graph in this embodiment: According to the constructed graph G<V,W> , based on a goal, in the two sub-graphs segmented, the cut value between the same sub-graph is relatively large, and the cut value between different sub-graphs is relatively small; Note that for two different graphs A and B, the expression for cut is: cut(A,B)=Σ u∈A，u∈B W(u,v); To achieve the above requirements, it is necessary to find the minimum value of the objective function; Its objective function is: The first part of the above formula is the sum of the weights from subgraph A to subgraph B, and the second part is the sum of the weights from subgraph B to subgraph A in the first part of the above formula; It should be noted here that V is the set of all points of the graph; In the process of minimizing the objective function f, not only the minimum cut value between different graphs is considered, but also the maximum cut value between the same graph is considered. The derivation process is as follows: From this formula we can get: f = 2-g; make From the above formula, we can see that while minimizing f, the g value is also maximized. From the formula of the g value, we can see that the function value of g mainly expresses the sum of the cut values ​​in Figure A and Figure B.

[0108] Therefore, by minimizing f, we can get f=2-g, and g is also maximized. Therefore, by obtaining the minimum value of the objective function f, the maximum value of the g function is also obtained at the same time.

[0109] However, since solving the minimum value of the objective function value f is an NP-hard problem, in actual solution, an approximate solution algorithm is mainly used to make the solved f value close to the minimum value of the objective function value f.

[0110] The objective function f can also be expressed as: Where D is a diagonal matrix with elements d(i), and W is a weight matrix; For d(i), the expression is d(i) = Σ j w i，j ; And DW becomes a Laplace matrix, which is a semi-positive definite matrix; So the f can also be expressed as: That is, find the matrix (ID -0.5 WD -0.5 ) can approximate the minimum value of the objective function, which can be expressed as:

[0111] On the basis of the above, according to the constructed graph, the constructed graph can be divided into K classes according to the K-means algorithm, and then greedy pruning is adopted; During the greedy pruning process, two blocks are merged each time, and this process is repeated until the number of blocks is divided to the preset expected value. In each merging process, the following definition must be satisfied at a minimum; The expression is: Wherein V is the set of all nodes in the graph, and their weight values ​​are updated after merging.

[0112] The defective parts of the equipment at the controlled end can be detected in the above manner, and the detected relevant data are transmitted to the database through the edge 1. The server can provide relevant production maintenance suggestions based on the data.

[0113] In some embodiments, the preventive maintenance monitors the mechanisms and equipment through time cycles, motion cycles and usage conditions, and performs adjustments, lubrication and regular inspections while they are still in normal condition. For equipment that requires regular maintenance, it is necessary to set a monitoring cycle or mechanical motion cycle or usage conditions to determine whether preventive maintenance reminders are needed.

[0114] The edge gateway device in the present invention can provide support for the operation and maintenance of the entire system through intelligent fault diagnosis and automated fault handling. Specifically, the edge gateway device can infer the possible cause of the unit failure by real-time acquisition of the PLC status signal, and automatically provide corresponding solutions. For example, during the stripping operation, if the position signal of a separation device or a stripping knife is detected to be abnormal, the edge gateway will automatically send a correction instruction to the corresponding PLC cluster. After receiving the instruction, the PLC will immediately perform a secondary stripping operation in order to restore the normal state. If the plate fork still fails to generate the expected signal after this operation, the gateway device will continue to send instructions until the plate fork works normally. The key is that if multiple automated attempts fail to solve the fault, the edge gateway will record the event and automatically remind the relevant maintenance personnel to intervene manually. At the same time, based on the analysis of the fault signal, the gateway device will also give specific maintenance suggestions, providing valuable reference for subsequent troubleshooting and repair. This intelligent operation and maintenance solution based on PLC status monitoring and automatic fault handling not only greatly improves the reliability and stability of the system, but also reduces the cost and difficulty of manual maintenance. Compared with the traditional passive maintenance mode, this active early warning and automatic repair design can greatly shorten the fault response time and improve the overall operating efficiency of the production line.

[0115] In some embodiments, the controlled end, the remote control system, the server and the database are in the same network environment, and a one-way network gate is provided between the network and the server and the database. In this embodiment, it is mainly composed of two servers, one server is responsible for storing the data related to the controlled end in the PLC cluster collected by the edge gateway 2 in the server, and the other server is responsible for obtaining the data in the database and performing calculations when performing fault prediction, and adding a one-way network gate on the transmission path from the data in the database to the server.

[0116] By setting up a one-way firewall between these subsystems, the data source and the target end can be effectively isolated, illegal access can be prevented, and the one-way nature of data transmission can be ensured. For example, by setting up a one-way firewall between the server and the database, the database can be prevented from being accessed or tampered with remotely, ensuring the accuracy and reliability of the fault prediction analysis results.

[0117] At the same time, the one-way network gate can also avoid data backflow caused by network failure or operational errors. Once the data flows back, it may have a serious impact on the entire fault prediction system. For example, if the PLC data collected by the edge gateway flows back to the server, it will cause server data confusion and affect the accuracy of fault prediction. The one-way network gate can effectively block data backflow, ensure that data always flows in the expected direction, and enhance the stability of the entire system.

[0118] It is important to emphasize that in the field of nonferrous metal processing, the safety and reliability of the fault prediction system are not only related to the normal operation of production, but also to the economic benefits and social responsibility of the entire enterprise. Once a major fault accident occurs, it will not only cause huge economic losses, but also may cause serious environmental pollution and even threaten the lives of employees and surrounding residents. Therefore, enterprises must attach great importance to this issue and take effective technical measures to ensure the safety of the production process.

[0119] The one-way network firewall technology is an effective solution that has emerged in this context. It can not only effectively block illegal access and prevent data tampering, but also ensure the unidirectionality of data transmission and prevent data backflow caused by network failures. These characteristics undoubtedly provide strong technical support for non-ferrous metal processing companies to build a safe and reliable fault prediction system.

[0120] Of course, the one-way network gate is only a part of the entire fault prediction system security system. Enterprises also need to take comprehensive security protection measures from multiple angles. For example, security factors should be fully considered at the beginning of system design, and reliable hardware and software components should be used; during operation, system vulnerability scans and security checks should be performed regularly to promptly discover and repair hidden dangers; at the same time, employee security awareness training should be strengthened to increase everyone's attention to system security. Only by organically combining various protection measures can a complete security protection system be truly built to ensure the stable and efficient operation of the fault prediction system.

[0121] In some embodiments, the controlled end, the remote control system, the server and the database are in the same network environment, and a one-way network gate is provided between the network, the server and the database. In this embodiment, the function of the server is to upload PLC data to the database and make production suggestions based on the data in the database, such as fault prediction and preventive maintenance suggestions, etc. Two-way authentication is adopted between the server and the database; That is, when the server obtains data from the database, not only does the identification module provided in the database need to identify the identity of the server, but the server also needs to identify the identity of the database.

[0122] When the server obtains data from the database, the detailed process of two-way identity authentication is as follows: Initial connection request: The server initiates a connection request to the database, which contains the server's initial identity identifier.

[0123] Database response: After the database receives the request, it generates a random challenge string and sends it back to the server.

[0124] Server Authentication: After receiving the challenge, the server signs the challenge string using its private key and sends the signature along with its digital certificate to the database.

[0125] Database authentication server identity: The database identification module uses the server's public key (obtained from the digital certificate) to verify the validity of the signature. At the same time, it checks the certificate's validity period, issuing authority and other information.

[0126] Database Authentication: After verifying the server's identity, the database generates its own random challenge string, signs it with its own private key, and sends it to the server along with the database's digital certificate.

[0127] Server authenticates the database: After receiving the authentication information from the database, the server uses the same method to verify the identity of the database, including the signature and certificate validity of the database.

[0128] Session key negotiation: After both parties have been authenticated, the server and database use an asymmetric encryption algorithm (such as Diffie-Hellman key exchange) to negotiate a temporary session key.

[0129] Secure channel establishment: An encrypted communication channel is established using the negotiated session key, and a symmetric encryption algorithm (such as AES) is usually used for subsequent data transmission.

[0130] Access control checks: The database further checks the server's access permissions to determine whether it has the authority to obtain the specific data requested.

[0131] Data transmission: If all authentication and authorization are successful, the database starts transmitting the requested data to the server through a secure channel.

[0132] Integrity Verification: Each packet transmitted is accompanied by a message authentication code (MAC), which allows the receiver to verify whether the data has been tampered with during transmission.

[0133] Session maintenance: Throughout the data transfer process, the two parties may periodically update the session key for added security.

[0134] Session Termination: After the data transmission is completed, the two parties negotiate to terminate the session, destroy the session key, and clear the relevant temporary authentication information.

[0135] This detailed two-way authentication process ensures mutual trust between the server and the database, prevents security threats such as identity forgery and man-in-the-middle attacks, and protects the confidentiality and integrity of data through encrypted communication. The entire process not only verifies the identities of both parties, but also establishes a secure communication channel, providing a solid security foundation for subsequent data exchange.

[0136] The present invention deeply explains its purpose, technical scheme and beneficial effects through specific embodiments, but these embodiments are only used as examples to show the application mode of the invention and do not constitute a limitation on the protection scope of the present invention. We explicitly point out that any reasonable modification, equivalent substitution or technical improvement under the guidance of the spirit and principles of the present invention should be included in the protection scope of the present invention. This means that as long as these changes do not deviate from the core idea and basic function of the invention, they should be protected by patent rights. The scope of protection of the present invention should be broad, including all direct and obvious variants and non-obvious innovations that technical experts can reasonably deduce based on the disclosure of the present invention. This broad protection is intended to promote further research and development based on the present invention, while ensuring that its innovation and practicality are fully protected by law.

Claims

1. A remote control system based on an electrolysis workshop, characterized in that: The system includes: a remote control system, a server, a PLC cluster and a controlled end. The remote control system exchanges data with the equipment on the controlled end through the PLC cluster. The relevant data in the PLC cluster is stored in a database through the server. The server gives production suggestions based on the data in the database.

2. A remote control system based on an electrolysis workshop according to claim 1, characterized in that: The production suggestion includes the prediction of fault events of the controlled end and the preventive maintenance of the controlled end; the fault event prediction calculates the occurrence time of the next fault event according to the frequency difference of adjacent fault events and the occurrence time of adjacent fault events; The preventive maintenance is to obtain relevant data of the controlled terminal equipment through the equipment in the remote control system, and give preventive maintenance suggestions according to the relevant data.

3. A remote control system based on an electrolysis workshop according to claim 1, characterized in that: The equipment at the controlled end includes several anode shaping units, several residual pole washing units and several cathode stripping units; the PLC cluster includes the anode shaping unit PLC, the residual pole washing unit PLC and the cathode shaping unit PLC.

4. A remote control system based on an electrolysis workshop according to claim 3, characterized in that: The remote control system sends instructions to the PLC cluster through the edge gateway 1, and controls the equipment on the controlled end through the PLC; The control process of the anode shaping unit is as follows: the remote control system inputs the remote control instructions to the edge gateway one, and the edge gateway one sends the remote control instructions to the anode shaping unit PLC. Through the PLC, the anode shaping unit can be controlled to process the surface of mechanical parts.

5. A remote control system based on an electrolysis workshop according to claim 3, characterized in that: The control process of the cathode stripping machine is as follows: the remote control system sends the input remote control instruction to the edge gateway one, and the edge gateway one sends the remote control instruction to the cathode stripping machine unit PLC, through which the cathode stripping machine unit can be controlled to strip the metal sheets attached to the cathode plates; after the cathode stripping machine completes its work, the control process of the residual pole washing unit is as follows: the remote control system sends the input remote control instruction to the edge gateway one, and the edge gateway one sends the remote control instruction to the residual pole washing unit PLC, through which the residual pole washing unit can be controlled to clean the metal residues attached to the cathode plates.

6. A remote control system based on an electrolysis workshop according to any one of claims 1 to 4, characterized in that: The relevant data in the PLC cluster is recorded by the edge gateway 2, and the edge gateway 2 sends the relevant data of the PLC to the server and stores it in the database through the server.

7. A remote control system based on an electrolysis workshop according to any one of claims 1 to 4, characterized in that: A single network gate is set between the server and the database. There are two servers. One server is responsible for storing data in the database, and the other server is responsible for making production maintenance recommendations based on the data in the database through relevant algorithms. Two-way authentication is required during the data interaction between the server and the database, that is, mutual identity authentication between the server and the database.

8. A remote control system based on an electrolysis workshop according to any one of claims 1 to 4, characterized in that: The remote control system includes a hardware control terminal, a monitoring terminal and control items. The hardware control terminal is used to input remote control instructions, and the monitoring terminal detects defects on the surface of related unit equipment. The remote control system based on the electrolysis workshop also includes a data management platform.

9. A fault prediction method, characterized in that: The fault event prediction method is as follows: the server searches for the time of occurrence of adjacent fault events related to the fault event based on the input relevant information and the daily working statistical frequency, and calculates the interval of the next fault event occurrence time based on the difference between the time of occurrence of the adjacent fault event and the daily working statistical frequency and the fault event.

10. A fault prediction method according to claim 9, characterized in that: The fault prediction method steps are: S1: Input the fault events of the controlled device and the date of occurrence of the fault events into the server through the data management platform; S2: The server searches for the daily working frequency of the device of the controlled end on the corresponding date according to the acquired fault event and its date. If the fault event has no adjacent fault event, the fault event is stored in the database. If so, step S3 is executed. S3: The server searches for the daily work statistical frequency and time of the adjacent fault events according to the acquired fault event and its date; S4: The server calculates the difference between the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events and the difference in occurrence time according to the daily working statistical frequency of the fault event and the daily working statistical frequency of the adjacent fault events; S5: The server calculates the next occurrence of a fault event of the controlled device through a relevant algorithm based on the difference in daily working frequency between the fault event and the adjacent fault event and the difference in occurrence time, and records it in the database.

Images