Abnormality analysis method and system applied to hydrogel production control system

Abstract

The invention provides an anomaly analysis method and system applied to a hydrogel production control system, and the method comprises the steps: obtaining an initialized basic anomaly diagnosis neural network and X prior anomaly diagnosis neural networks with different anomaly diagnosis capabilities, and carrying out the abnormal state chain diagnosis prediction of system operation data, and performing guided learning on the initialized basic network by using a first prior anomaly diagnosis neural network with the weakest capability in the X prior anomaly diagnosis neural networks to generate a basic network after the first round of guided learning. And then utilizing an ith prior anomaly diagnosis neural network with a stronger anomaly diagnosis capability to continue guided learning on the basic network after the (i-1) th round of guided learning (i is an incremental integer greater than 1 and not greater than X), and finally outputting the basic network after the Xth round of guided learning as a target anomaly diagnosis neural network. The method is used for carrying out abnormal state chain diagnosis prediction on input system operation data, and the accuracy and reliability of abnormity analysis of the hydrogel production control system can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an abnormality analysis method and system applied to an aqueous gel production control system. Background Art

[0002] In the water-based gel production industry, the stable operation of the production control system is crucial to product quality and production efficiency. The water-based gel production process involves multiple complex links and numerous parameter variables. The production control system is prone to various abnormalities during operation. If these abnormalities are not discovered and handled in time, they may lead to serious consequences such as product quality degradation, production stagnation, and even equipment damage.

[0003] At present, the existing technologies mostly use a single fixed model or a simple rule matching method to perform abnormal analysis on the aqueous gel production control system. The use of a single fixed model is difficult to adapt to the complex and variable system operation data caused by changes in raw material characteristics, production process adjustments, and environmental factor fluctuations during the aqueous gel production process, because its structure and parameters are basically fixed after training. Faced with abnormal situations of different types and degrees, this single model often shows problems of inaccurate diagnosis and a high rate of missed diagnosis.

[0004] Although the anomaly analysis method based on simple rule matching has a certain efficiency in identifying certain specific and known anomaly patterns, the formulation of rules often relies on experience. For complex and unknown abnormal state chains, it is difficult to make effective diagnosis because it is impossible to exhaust all possible situations in advance. In addition, when there are subtle changes or new abnormal situations in the production system, it is necessary to manually re-sort and update the rules, which is not only labor-intensive, but also difficult to ensure timeliness and accuracy. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides an abnormality analysis method applied to an aqueous gel production control system, the method comprising: Obtaining an initialized basic abnormality diagnosis neural network and X a priori abnormality diagnosis neural networks, wherein the a priori abnormality diagnosis neural networks have different abnormality diagnosis capabilities, wherein the basic abnormality diagnosis neural network and the X a priori abnormality diagnosis neural networks are used to diagnose and predict abnormal state chains of system operation data of the aqueous gel production control system; Using a first a priori abnormality diagnosis neural network to guide learning of the initialized basic abnormality diagnosis neural network, to generate a basic abnormality diagnosis neural network after a first round of guided learning, wherein the first a priori abnormality diagnosis neural network is a priori abnormality diagnosis neural network with the weakest abnormality diagnosis ability among the X a priori abnormality diagnosis neural networks; Using the i-th prior abnormality diagnosis neural network to guide the learning of the basic abnormality diagnosis neural network after the i-1th round of guided learning, to generate the basic abnormality diagnosis neural network after the i-1th round of guided learning, wherein i is an increasing integer greater than 1 and not greater than X, the basic abnormality diagnosis neural network after the i-1th round of guided learning is generated by using the i-1th prior abnormality diagnosis neural network to guide the learning, the i-th prior abnormality diagnosis neural network and the i-1th prior abnormality diagnosis neural network are any one of the X prior abnormality diagnosis neural networks, and the abnormality diagnosis ability of the i-th prior abnormality diagnosis neural network is stronger than the abnormality diagnosis ability of the i-1th prior abnormality diagnosis neural network; The basic abnormality diagnosis neural network after the Xth round of guided learning is output as a target abnormality diagnosis neural network, so as to perform diagnosis and prediction of abnormal state chains for any input system operation data through the target abnormality diagnosis neural network.

[0006] On the other hand, an embodiment of the present application also provides an intelligent production service system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiment of the present application obtains multiple prior abnormal diagnosis neural networks with different abnormal diagnosis capabilities and an initialized basic abnormal diagnosis neural network. This combination breaks through the limitations of traditional single neural network diagnosis. Traditional methods often rely only on a fixed neural network model, which is difficult to fully adapt to the complex and changeable operation of the aqueous gel production control system. The present method uses multiple prior networks with different capabilities, which can capture abnormal features in the system operation data from different angles, providing a rich and diverse source of knowledge for subsequent guided learning.

[0008] In the guided learning process, the prior anomaly diagnosis neural networks from weak to strong are used to guide the learning of the basic anomaly diagnosis neural network in turn. For example, starting the guided learning with the first prior anomaly diagnosis neural network with the weakest anomaly diagnosis ability can enable the basic anomaly diagnosis neural network to first master the relatively simple and basic anomaly diagnosis mode, avoiding the learning difficulties or overfitting caused by facing complex diagnostic tasks at the beginning. As the learning rounds progress, the prior anomaly diagnosis neural network with gradually enhanced ability is used for guidance, so that the basic anomaly diagnosis neural network can gradually absorb more advanced and complex anomaly diagnosis knowledge, continuously improve its diagnostic ability, and realize a step-by-step and hierarchical learning process. This unique learning method can effectively mine abnormal features at different levels, greatly improving the accuracy and comprehensiveness of the neural network's diagnosis and prediction of abnormal state chains, which is incomparable to traditional single learning methods or disordered learning methods.

[0009] The target abnormality diagnosis neural network, which is the final output, has a powerful and comprehensive abnormality diagnosis capability due to multiple rounds of guided learning from weak to strong. It can not only accurately identify common abnormal conditions in the water-based gel production control system, but also accurately diagnose and predict some hidden and complex abnormal state chains. This helps to timely discover potential problems in the production process, take corresponding measures in advance, avoid production failures, improve the stability and reliability of the water-based gel production control system, effectively reduce production costs, and improve production efficiency and product quality.

[0010] In addition, when the aqueous gel production control system changes or new types of abnormalities appear, it is only necessary to appropriately adjust the parameters of the prior abnormality diagnosis neural network or add a new prior network to quickly optimize the target abnormality diagnosis neural network to adapt it to the new production environment and diagnostic needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the abnormality analysis method applied to the aqueous gel production control system provided in the embodiment of the present application.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of the intelligent production service system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 1 is a flow chart of an abnormality analysis method applied to an aqueous gel production control system provided by an embodiment of the present application. The abnormality analysis method applied to an aqueous gel production control system is introduced in detail below.

[0014] Step S110, obtaining an initialized basic abnormality diagnosis neural network and X a priori abnormality diagnosis neural networks, wherein the abnormality diagnosis capabilities of the a priori abnormality diagnosis neural networks are different, wherein the basic abnormality diagnosis neural network and the X a priori abnormality diagnosis neural networks are used to diagnose and predict the abnormal state chain of the system operation data of the aqueous gel production control system.

[0015] In this embodiment, during the production process of the aqueous gel, the production control system will generate a large amount of system operation data. For example, in a factory that produces aqueous gels used in medical dressings, its production control system covers data from raw material mixing, reactor temperature control, stirring speed adjustment to finished product packaging. When the basic abnormality diagnosis neural network is initialized, it may be set to have some basic network structures and initial weights, which are set based on a general understanding of the aqueous gel production process.

[0016] Assume that X is equal to 3, that is, there are 3 prior anomaly diagnosis neural networks. These 3 prior anomaly diagnosis neural networks have different anomaly diagnosis capabilities. The first prior anomaly diagnosis neural network may be built based on a simpler algorithm and can diagnose some common and obvious abnormal conditions. For example, it can detect when the temperature of the reactor deviates from the set value beyond a certain range, but it has weaker diagnostic capabilities for complex abnormal state chains (such as changes in gel consistency caused by temperature abnormalities, which in turn affect the sealing of subsequent packaging links). The second prior anomaly diagnosis neural network may be an improvement on the first one. In addition to being able to detect temperature abnormalities, it can also conduct a preliminary analysis of the relationship between stirring speed abnormalities and temperature abnormalities. Its abnormality diagnosis capability is at a medium level. The third prior anomaly diagnosis neural network is a more complex and accurate network that can comprehensively consider the complex interrelationships between multiple factors such as the composition ratio of raw materials, pressure, temperature, stirring speed, gel consistency, and operating status of the packaging equipment, and accurately diagnose abnormal state chains caused by the interaction of multiple factors. It has the strongest abnormality diagnosis capability. These prior anomaly diagnosis neural networks and basic anomaly diagnosis neural networks are prepared for the purpose of diagnosing and predicting the abnormal state chain of the system operation data of the aqueous gel production control system.

[0017] Step S120, using the first a priori abnormality diagnosis neural network to guide learning of the initialized basic abnormality diagnosis neural network to generate a basic abnormality diagnosis neural network after a first round of guided learning, wherein the first a priori abnormality diagnosis neural network is the a priori abnormality diagnosis neural network with the weakest abnormality diagnosis ability among the X a priori abnormality diagnosis neural networks.

[0018] Continuing with the example of the above aqueous gel production, the initialized basic abnormality diagnosis neural network does not have a good understanding and judgment ability of the data in the production process. The first prior abnormality diagnosis neural network, that is, the network that can only detect obvious abnormalities such as simple temperature deviation, begins to guide the learning of the basic abnormality diagnosis neural network.

[0019] Assume that during the production process, the basic abnormality diagnosis neural network receives a series of temperature data from the reactor temperature sensor, stirring speed data from the stirring speed sensor, and gel consistency detection data. The first prior abnormality diagnosis neural network has learned that when the temperature data suddenly rises beyond the normal fluctuation range, this may be an abnormal situation. Therefore, the method and basis of this judgment can be passed to the basic abnormality diagnosis neural network. For example, the first prior abnormality diagnosis neural network will show the basic abnormality diagnosis neural network how to identify abnormal patterns in temperature data, including how to determine the normal temperature fluctuation range, and how to mark it as a possible abnormality when the temperature exceeds this range.

[0020] In this guided learning process, the basic abnormality diagnosis neural network adjusts its own neuron connection weights and internal parameters according to the information provided by the first prior abnormality diagnosis neural network. For example, if the temperature data rises by 5 degrees Celsius in a short period of time and lasts for a period of time, the first prior abnormality diagnosis neural network will tell the basic abnormality diagnosis neural network that this is a situation that needs attention. The basic abnormality diagnosis neural network will adjust its parameters accordingly so that it can make similar judgments when encountering similar temperature change patterns in the future, thereby generating the basic abnormality diagnosis neural network after the first round of guided learning.

[0021] Step S130, using the i-th prior abnormality diagnosis neural network to guide the learning of the basic abnormality diagnosis neural network after the i-1th round of guided learning, to generate the basic abnormality diagnosis neural network after the i-1th round of guided learning, wherein i is an increasing integer greater than 1 and not greater than X, the basic abnormality diagnosis neural network after the i-1th round of guided learning is generated by guiding the learning using the i-1th prior abnormality diagnosis neural network, the i-th prior abnormality diagnosis neural network and the i-1th prior abnormality diagnosis neural network are any one of the X prior abnormality diagnosis neural networks, and the abnormality diagnosis ability of the i-th prior abnormality diagnosis neural network is stronger than the abnormality diagnosis ability of the i-1th prior abnormality diagnosis neural network.

[0022] Assume i=2, then the i-1th prior anomaly diagnosis neural network is the first prior anomaly diagnosis neural network, and the i-th prior anomaly diagnosis neural network is the one with medium anomaly diagnosis ability. The basic anomaly diagnosis neural network after the i-1th round of guided learning has learned some judgment ability about simple abnormal situations from the first prior anomaly diagnosis neural network.

[0023] Now, the i-th prior anomaly diagnosis neural network starts to guide the learning of the basic anomaly diagnosis neural network after the i-1th round of guided learning. In the production of aqueous gels, in addition to temperature anomalies, the stirring speed also has an important impact on product quality. The i-th prior anomaly diagnosis neural network has learned that when the stirring speed is unstable and the temperature fluctuates at the same time, it may cause uneven gel consistency.

[0024] After receiving the system operation data including temperature, stirring speed and gel consistency, the i-th prior anomaly diagnosis neural network can pass the analysis method of the complex relationship between these data to the basic anomaly diagnosis neural network after the i-1th round of guided learning. For example, when the stirring speed fluctuates by more than 10 revolutions per minute at 100 revolutions per minute, and the temperature fluctuates by more than 3 degrees Celsius within 10 minutes, the gel consistency may change beyond the normal range in the subsequent production links.

[0025] The basic abnormality diagnosis neural network after the i-1th round of guided learning further adjusts its own neuron weights and parameters based on the information provided by the i-th prior abnormality diagnosis neural network. It learns to not only focus on the abnormality of a single data, but also consider the relationship between multiple related data. For example, it can adjust its own algorithm so that when receiving similar combined abnormal data of temperature and stirring speed, it can predict the possible abnormality of gel consistency, thereby generating the basic abnormality diagnosis neural network after the i-th round (here is the second round) of guided learning.

[0026] When i=3, the i-1th prior anomaly diagnosis neural network is the network with medium anomaly diagnosis ability, and the i-th prior anomaly diagnosis neural network is the network with the strongest anomaly diagnosis ability. The basic anomaly diagnosis neural network after the i-1th round (the second round) of guided learning is already able to handle some simple associated anomalies.

[0027] This strongest i-th prior anomaly diagnosis neural network is able to analyze the complex relationship between all data from the raw material input ratio to the final packaging link. For example, it knows that too high a certain ingredient in the raw material may cause the pressure in the reactor to increase, which in turn affects the temperature control and ultimately affects the consistency of the gel and the sealing of the packaging. It imparts this comprehensive and complex analysis method of the abnormal state chain to the basic anomaly diagnosis neural network after the i-1th round (2nd round) guided learning.

[0028] After receiving this information, the basic abnormality diagnosis neural network after the guided learning of the i-1th round (the second round) adjusts its internal structure and parameters again. It begins to learn how to comprehensively consider the data of all production links, from the initial state of the raw materials to the packaging integrity of the final product. For example, when the raw material component ratio data is abnormal, it can predict the abnormal conditions that may occur in the subsequent links based on the previously learned knowledge, including changes in the pressure, temperature, stirring speed, gel consistency and operating status of the packaging equipment in the reactor, thereby generating the basic abnormality diagnosis neural network after the i-th round (the third round) of guided learning.

[0029] Step S140, outputting the basic abnormality diagnosis neural network after the Xth round of guided learning as a target abnormality diagnosis neural network, so as to perform diagnosis and prediction of abnormal state chains for any input system operation data through the target abnormality diagnosis neural network.

[0030] After the previous guided learning process, when X = 3, the basic abnormality diagnosis neural network after the third round of guided learning has become a very powerful target abnormality diagnosis neural network. In the production process of hydrogel, this target abnormality diagnosis neural network can diagnose and predict the abnormal state chain of any input system operation data.

[0031] For example, when a new batch of raw materials is put into production, the target abnormality diagnosis neural network receives a series of initial data such as the raw material composition ratio, the initial temperature of the reactor, the pressure, the stirring speed, etc. It can predict the abnormal conditions that may occur in the entire production process based on the knowledge learned before. If a key component in the raw material composition ratio is slightly higher than the normal range, it can predict that the pressure in the reactor may increase within 15 minutes after the start of the reaction, which may cause the temperature control system to require additional adjustments. If the temperature is not adjusted in time, the stirring speed may be affected, resulting in uneven gel consistency, which may eventually affect the sealing of the packaging.

[0032] This target abnormality diagnosis neural network can accurately diagnose and predict various abnormal state chains in the entire production process, thereby helping manufacturers to promptly identify potential problems and take corresponding measures to ensure the quality of aqueous gel products and the smooth progress of production.

[0033] Based on the above steps, the embodiment of the present application obtains multiple prior abnormal diagnosis neural networks with different abnormal diagnosis capabilities and an initialized basic abnormal diagnosis neural network. This combination breaks through the limitations of traditional single neural network diagnosis. Traditional methods often rely only on a fixed neural network model, which is difficult to fully adapt to the complex and changeable operation of the aqueous gel production control system. The present method uses multiple prior networks with different capabilities, which can capture abnormal features in the system operation data from different angles, providing a rich and diverse source of knowledge for subsequent guided learning.

[0034] In the guided learning process, the prior anomaly diagnosis neural networks from weak to strong are used to guide the learning of the basic anomaly diagnosis neural network in turn. For example, starting the guided learning with the first prior anomaly diagnosis neural network with the weakest anomaly diagnosis ability can enable the basic anomaly diagnosis neural network to first master the relatively simple and basic anomaly diagnosis mode, avoiding the learning difficulties or overfitting caused by facing complex diagnostic tasks at the beginning. As the learning rounds progress, the prior anomaly diagnosis neural network with gradually enhanced ability is used for guidance, so that the basic anomaly diagnosis neural network can gradually absorb more advanced and complex anomaly diagnosis knowledge, continuously improve its diagnostic ability, and realize a step-by-step and hierarchical learning process. This unique learning method can effectively mine abnormal features at different levels, greatly improving the accuracy and comprehensiveness of the neural network's diagnosis and prediction of abnormal state chains, which is incomparable to traditional single learning methods or disordered learning methods.

[0035] The target abnormality diagnosis neural network, which is the final output, has a powerful and comprehensive abnormality diagnosis capability due to multiple rounds of guided learning from weak to strong. It can not only accurately identify common abnormal conditions in the water-based gel production control system, but also accurately diagnose and predict some hidden and complex abnormal state chains. This helps to timely discover potential problems in the production process, take corresponding measures in advance, avoid production failures, improve the stability and reliability of the water-based gel production control system, effectively reduce production costs, and improve production efficiency and product quality.

[0036] In addition, when the aqueous gel production control system changes or new types of abnormalities appear, it is only necessary to appropriately adjust the parameters of the prior abnormality diagnosis neural network or add a new prior network to quickly optimize the target abnormality diagnosis neural network to adapt it to the new production environment and diagnostic needs.

[0037] In a possible implementation, step S130 includes: Step S131, using the i-th prior abnormality diagnosis neural network to perform graph convolution conversion on the first sample system operation data, to generate a graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

[0038] Step S132, using the basic abnormality diagnosis neural network after the i-1th round of guided learning to perform graph convolution conversion on the first sample system operation data, to generate a graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning.

[0039] Step S133, determining a global training error according to the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

[0040] Step S134, optimizing the neuron weight information of the basic abnormality diagnosis neural network after the i-1th round of guided learning according to the global training error, and generating the basic abnormality diagnosis neural network after the i-th round of guided learning.

[0041] In a possible implementation, the first sample system operation data includes a first baseline system operation sample, a first associated system operation sample and a first non-associated system operation sample, and the i-1th prior anomaly diagnosis neural network and the i-th prior anomaly diagnosis neural network are both dual graph convolution modules.

[0042] Step S131 includes: Step S1311, normalize the first benchmark system operation sample, the first associated system operation sample and the first non-associated system operation sample respectively to generate network learning data of the first benchmark system operation sample, network learning data of the first associated system operation sample and network learning data of the first non-associated system operation sample.

[0043] Step S1312, using the i-th prior abnormality diagnosis neural network to perform graph convolution transformation on the network learning data of the first benchmark system operation sample, the network learning data of the first associated system operation sample, and the network learning data of the first non-associated system operation sample, to generate a graph expression vector of the first benchmark system operation sample, a graph expression vector of the first associated system operation sample, and a graph expression vector of the first non-associated system operation sample.

[0044] Step S1313: Use the graph expression vector of the first benchmark system operation sample, the graph expression vector of the first associated system operation sample, and the graph expression vector of the first non-associated system operation sample as the graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

[0045] In this embodiment, in the production scenario of aqueous gel, the first example system operation data includes a variety of important operation examples. The first benchmark system operation example can be the operation data of a typical batch in the normal production process. For example, in this batch, the data of the entire process from the precise ratio of raw materials, the temperature in the reactor changes according to the predetermined curve, the stirring speed is stable in the appropriate range, to the consistency of the gel gradually reaching the standard requirements, and then to the normal parameters of the packaging link. The first associated system operation example is the data situation that is directly related to the operation of the benchmark system, such as when the temperature control system of the reactor fluctuates slightly, the changes in the stirring speed, gel consistency and sealing pressure of the subsequent packaging link related to the temperature. The first non-associated system operation example is the data situation that is not directly related to the operation of the benchmark system, such as the operation data of some unrelated equipment in the equipment maintenance work carried out in another area of ​​the production workshop, but these data are also collected as part of the overall system operation data.

[0046] For the case where both the i-th prior anomaly diagnosis neural network and the i-1-th prior anomaly diagnosis neural network are dual graph convolution modules, the first sample system operation data needs to be normalized before the graph convolution conversion. Taking the first benchmark system operation sample as an example, in the production of aqueous gel, there may be a variety of process parameters, such as the temperature of the reactor, the stirring speed, the flow rate of raw materials, etc. The purpose of normalizing these data is to enable the neural network to better process the data. For example, for temperature data, since its numerical range may be large and its change law is different from other parameters, specific processing is required. Assuming that the temperature range is between 20-80 degrees Celsius during normal production, it may be linearly mapped to a specific interval, such as between 0-1. This is the process of generating network learning data for the first benchmark system operation sample. Similar processing is also performed for the first associated system operation sample and the first non-associated system operation sample. For example, for the stirring speed data, if its original unit is revolutions per minute, it may be converted into a dimensionless numerical form according to the characteristics of the production equipment and the requirements of the neural network, thereby generating their respective network learning data.

[0047] Then, the network learning data of the first benchmark system operation sample, the network learning data of the first associated system operation sample, and the network learning data of the first non-associated system operation sample are respectively converted by graph convolution using the i-th prior anomaly diagnosis neural network. Graph convolution conversion is an operation that extracts and converts features of graph structure data. In the aqueous gel production control system, the system operation data can be regarded as a graph structure, in which each production link and equipment can be regarded as a node, and the association between links or equipment can be regarded as an edge. For the network learning data of the first benchmark system operation sample, the graph convolution conversion process of the i-th prior anomaly diagnosis neural network may be as follows: Assuming that each data point in the network learning data corresponds to a node in the graph, the graph convolution operation will extract new features based on the connection relationship (edge) between nodes and the characteristics of the nodes themselves (such as temperature, stirring speed and other values). After a series of calculations, a graph expression vector of the first benchmark system operation sample is generated. This vector contains the characteristic information about the operation of the benchmark system after graph convolution conversion, for example, it may contain the relationship characteristics between the temperature and stirring speed in different areas of the reactor, the coupling characteristics of the raw materials with the temperature and stirring speed at different reaction stages, etc. Similarly, after performing graph convolution transformation on the first associated system operation sample and the first non-associated system operation sample, their respective graph expression vectors are obtained. These three graph expression vectors are used as the graph convolution transformation results of the i-th prior anomaly diagnosis neural network.

[0048] At the same time, the same graph convolution conversion operation is performed on the first sample system operation data using the basic abnormal diagnosis neural network after the i-1th round of guided learning. For example, it can normalize the first benchmark system operation sample, the first associated system operation sample, and the first non-associated system operation sample, and then perform graph convolution conversion to generate the graph convolution conversion result of the basic abnormal diagnosis neural network after the i-1th round of guided learning. This result also contains the graph expression vectors of each sample, reflecting the feature extraction and conversion of the basic abnormal diagnosis neural network on these sample data.

[0049] Next, the global training error is determined based on the graph convolution conversion result of the basic abnormal diagnosis neural network after the i-1th round of guided learning and the graph convolution conversion result of the i-th prior abnormal diagnosis neural network. The determination of the global training error is a key step in measuring the difference in the processing results of the two neural networks. For example, the difference between the graph expression vector of the first benchmark system operation sample of the basic abnormal diagnosis neural network after the i-1th round of guided learning and the graph expression vector of the first benchmark system operation sample of the i-th prior abnormal diagnosis neural network is compared. Assume that in the graph expression vector of the i-th prior abnormal diagnosis neural network, the coupling characteristics of the temperature and stirring speed of a certain area in the reactor are represented as a specific numerical relationship, while in the graph expression vector of the basic abnormal diagnosis neural network after the i-1th round of guided learning, this numerical relationship is different, and this difference is calculated by a specific algorithm. Similar difference calculations are also performed on the graph expression vectors of the first associated system operation sample and the first non-associated system operation sample, and these differences are combined to determine the global training error.

[0050] Finally, based on this global training error, the neuron weight information of the basic abnormal diagnosis neural network after the i-1th round of guided learning is optimized. In the aqueous gel production control system, the neuron weight information determines the way the neural network processes the input data. If the global training error is large, it means that the processing result of the basic abnormal diagnosis neural network after the i-1th round of guided learning is quite different from the processing result of the i-th prior abnormal diagnosis neural network, and the neuron weight needs to be adjusted. For example, if in a certain production link, the relationship between temperature and gel consistency has a large difference in the processing results of the two neural networks, then the neuron weight related to this relationship will be adjusted. By continuously adjusting the neuron weight, the processing result of the basic abnormal diagnosis neural network after the i-1th round of guided learning is closer to the processing result of the i-th prior abnormal diagnosis neural network, thereby generating the basic abnormal diagnosis neural network after the i-1th round of guided learning. This new neural network can more accurately diagnose and predict the abnormal state chain when processing the system operation data of the aqueous gel production control system.

[0051] In a possible implementation, the graph convolution transformation of the network learning data of the first benchmark system running sample is implemented by using Y sequentially connected graph convolution transformation layers included in the i-th prior anomaly diagnosis neural network.

[0052] Using the first graph convolution transformation layer among the Y sequentially connected graph convolution transformation layers, graph convolution transformation is performed on the network learning data of the first benchmark system running sample to generate a transformation result of the first graph convolution transformation layer.

[0053] Using the j-th graph convolution transformation layer, perform graph convolution transformation on the conversion result of the j-1-th graph convolution transformation layer to generate the conversion result of the j-th graph convolution transformation layer, and determine the graph expression vector of the first benchmark system running sample based on the conversion result of the Y-th graph convolution transformation layer, where j is an increasing integer greater than 1 and not greater than Y.

[0054] The step of using the first graph convolution conversion layer among the Y sequentially connected graph convolution conversion layers to perform graph convolution conversion on the network learning data of the first benchmark system running sample to generate a conversion result of the first graph convolution conversion layer includes: Step A110, obtaining a process parameter timing sequence corresponding to the first benchmark system operation sample, wherein the process parameter timing sequence includes continuously collected data of temperature control parameters, pressure feedback parameters, fluid flow rate monitoring parameters and gel consistency detection parameters.

[0055] In this embodiment, in the production process of aqueous gel, temperature control parameters are extremely critical. For example, in the reactor, precise temperature control is required at different stages to ensure the normal progress of the chemical reaction, and the continuously collected data reflects the change of temperature over time. The pressure feedback parameters are related to the sealing of the reactor and the reaction state of the internal substances. For example, gas may be generated during the reaction, causing pressure changes. The collection of these pressure data can reflect the stability of the production process. The fluid flow rate monitoring parameters involve the input of raw materials and the flow of various fluids during the reaction process. For example, the raw materials are injected into the reactor at a certain flow rate, and the gel flows in the pipeline to the next production link. The flow rate. The gel consistency detection parameters reflect the key indicators of the quality of aqueous gel products. As the production process progresses, the gel consistency will gradually change and need to be maintained within an appropriate range.

[0056] Step A120, normalizing the process parameter timing sequence to generate a normalized process parameter timing sequence, wherein the normalization includes dynamic range compression of temperature control parameters, noise filtering of pressure feedback parameters, and time alignment of fluid flow rate monitoring parameters.

[0057] For dynamic range compression of temperature control parameters, in the production of aqueous gels, it is assumed that the temperature range during normal production is between 20-80 degrees Celsius, but the actual collected data may have a certain fluctuation range. Through dynamic range compression, this range is mapped to an interval that is more suitable for neural network processing, such as between 0-1, which helps to improve the accuracy and efficiency of data processing. Noise filtering is performed on pressure feedback parameters. Since there may be various interference factors in the production environment, such as slight vibration of equipment, etc., which will cause noise in pressure data, the use of appropriate filtering algorithms (such as Kalman filtering algorithms) can remove these noises, so that the pressure feedback parameters can more accurately reflect the actual production status. Time alignment operations are performed on fluid flow rate monitoring parameters. On the aqueous gel production line, flow rate monitoring devices at different locations may have different sampling frequencies. The data with different sampling frequencies are unified into a preset timestamp sequence through interpolation algorithms and other methods for subsequent accurate analysis.

[0058] Step A130, based on the process correlation indicators between the process parameters in the normalized process parameter timing sequence, construct an adjacency matrix corresponding to the fluid flow topology structure of the first benchmark system operation sample, wherein the nodes of the adjacency matrix represent production control units, and the edge weights represent the coupling strength of the process parameters between adjacent production control units.

[0059] In the production of aqueous gels, the production control unit may include equipment such as reactors, agitators, and pipelines. Taking the reactor and agitator as an example, if the temperature control parameters are closely related to the operating status of the agitator (reflected in aspects such as pressure feedback parameters), then the edge weights between the nodes representing the two production control units, the reactor and the agitator, in the adjacency matrix (indicating the coupling strength of process parameters between adjacent production control units) will be large. By analyzing the correlation between different process parameters, such as the relationship between temperature and pressure, flow rate and gel consistency, etc., the adjacency matrix can be accurately constructed.

[0060] Step A140, based on the multidimensional parameter set of each production control unit in the normalized process parameter timing sequence, generate a feature matrix corresponding to the adjacency matrix, wherein the row dimension of the feature matrix corresponds to the production control unit identifier, and the column dimension corresponds to the normalized eigenvalues ​​of the temperature control parameters, pressure feedback parameters, fluid flow rate monitoring parameters and gel consistency detection parameters.

[0061] For each production control unit, such as a reactor, its temperature control parameters are statistically analyzed in time series, for example, the maximum temperature fluctuation value, the average heating rate, and the steady-state temperature holding time are extracted as the first eigenvector, which reflects the characteristics of the reactor in temperature control. The pressure feedback parameters are subjected to frequency domain energy analysis, and the low-frequency pressure oscillation amplitude, the high-frequency noise energy ratio, and the pressure mutation event count are extracted as the second eigenvector, which helps to analyze the characteristics of pressure changes. The fluid flow rate monitoring parameters are subjected to dynamic mode decomposition, and the laminar state duration, turbulence intensity index, and the periodic change frequency of the flow rate are extracted as the third eigenvector, which can reflect the state of fluid flow. The gel consistency detection parameters are subjected to outlier detection, and the number of abnormal viscosity fluctuations, the steady-state consistency variance, and the transient consistency gradient are extracted as the fourth eigenvector, which is an important representation of gel quality. These eigenvectors are spliced ​​vertically according to the production control unit identification to generate a feature matrix, in which the row dimension corresponds to the production control unit identification, and the column dimension corresponds to the normalized eigenvalues ​​of the temperature control parameters, pressure feedback parameters, fluid flow rate monitoring parameters, and gel consistency detection parameters.

[0062] Step A150, performing multi-channel data combination of the adjacency matrix and the characteristic matrix to generate multi-channel graph data of the first benchmark system operation sample, wherein the multi-channel graph data includes a fluid flow topology channel, a parameter timing correlation channel, and a dynamic process state channel.

[0063] The multi-channel graph data includes the fluid flow topology channel, which reflects the connection relationship and fluid flow path between production equipment; the parameter timing association channel, which reflects the relationship between different process parameters in the time series; and the dynamic process state channel, which integrates the dynamic state information of each production control unit.

[0064] Step A160, inputting the multi-channel graph data into the first graph convolution transformation layer among the Y sequentially connected graph convolution transformation layers, wherein the first graph convolution transformation layer includes a graph convolution kernel, a node feature aggregation module, and a nonlinear activation module.

[0065] Step A170, extracting process topology features from the adjacency matrix and the feature matrix through the graph convolution kernel to generate an initial convolution feature graph, wherein the weight matrix of the graph convolution kernel dynamically adjusts the aggregation weight based on the fluid flow path between the production control units.

[0066] Step A180, based on the node connection relationship in the initial convolution feature graph, use the node feature aggregation module to perform cross-node aggregation on the process parameter features of adjacent production control units to generate an aggregated feature matrix, wherein the cross-node aggregation includes gradient weighted fusion of temperature control parameters and pressure feedback parameters, and dynamic correlation superposition of fluid flow rate monitoring parameters and gel consistency detection parameters.

[0067] In this embodiment, the graph convolution kernel, node feature aggregation module and nonlinear activation module in the first graph convolution conversion layer begin to play a role. The graph convolution kernel extracts the process topology features of the adjacency matrix and the feature matrix to generate an initial convolution feature graph. In this process, the weight matrix of the graph convolution kernel dynamically adjusts the aggregation weight based on the fluid flow path between the production control units. For example, in the production of aqueous gels, if the fluid flow between the reactor and the pipeline has a greater impact on a certain process parameter, then when performing the convolution operation, the weight corresponding to this path will be adjusted accordingly to highlight this influence relationship. Based on the node connection relationship in the initial convolution feature graph, the node feature aggregation module is used to perform cross-node aggregation on the process parameter features of adjacent production control units to generate an aggregate feature matrix. In the aqueous gel production scenario, for adjacent reactors and agitators, the temperature control parameters and pressure feedback parameters are gradient-weighted fused, for example, the weight is determined according to the degree of influence of temperature on pressure, and the two are fused into a new feature; the fluid flow rate monitoring parameters and the gel consistency detection parameters are dynamically correlated and superimposed, and considering that the fluid flow rate may affect the gel consistency, the superposition operation is performed according to the dynamic relationship between them. The aggregate feature matrix is ​​input into the nonlinear activation module for feature nonlinear mapping to generate the conversion result of the first graph convolutional transformation layer. The output dimension of the nonlinear activation module is consistent with the column dimension of the feature matrix, and the activation function parameters are dynamically calibrated based on the steady-state threshold of the aqueous gel production process. For example, in the detection of gel consistency, if the steady-state threshold is a certain value, when the gel consistency-related features in the aggregate feature matrix are close to this steady-state threshold, the parameters of the activation function will be adjusted accordingly to ensure that the feature mapping in this area is more accurate.

[0068] Step A190, inputting the aggregated feature matrix into the nonlinear activation module for feature nonlinear mapping to generate the conversion result of the first graph convolutional transformation layer, wherein the output dimension of the nonlinear activation module is consistent with the column dimension of the feature matrix, and the activation function parameters are dynamically calibrated based on the steady-state threshold of the aqueous gel production process.

[0069] Step A200: Use the conversion result as the input of the next graph convolutional transformation layer to generate a graph expression vector of the first benchmark system running sample through cascade processing of Y graph convolutional transformation layers.

[0070] In this process, each layer of graph convolution transformation will further extract and transform features, gradually digging deeper into the information in the data. Finally, based on the conversion results of the Yth graph convolution transformation layer, the graph expression vector of the first benchmark system operation sample is determined. This graph expression vector integrates the feature extraction and conversion results of the process parameters, production control unit relationships, and other aspects of the first benchmark system operation sample during the entire graph convolution transformation process, and can provide effective data representation for subsequent abnormal diagnosis, thereby accurately diagnosing and predicting the abnormal state chain in the aqueous gel production control system.

[0071] For example, the step of normalizing the process parameter timing sequence to generate a normalized process parameter timing sequence includes: The temperature control parameters are subjected to dynamic range compression to generate a compressed temperature control parameter sequence, wherein the dynamic range compression linearly scales the original temperature data based on the aqueous gel phase transition temperature interval.

[0072] The pressure feedback parameters are subjected to noise filtering to generate a smoothed pressure feedback parameter sequence, wherein the noise filtering adopts an adaptive Kalman filter algorithm based on the vibration frequency of the process equipment.

[0073] A time alignment operation is performed on the fluid flow rate monitoring parameters to generate a synchronous fluid flow rate parameter sequence, wherein the time alignment operation unifies flow rate data of different sampling frequencies into a preset timestamp sequence through an interpolation algorithm.

[0074] In this embodiment, the temperature control parameters are crucial to the quality and production process of the product. The production of aqueous gels may involve a specific phase change process. For example, at a certain reaction stage, the gel changes from a liquid state to a solid state. This process corresponds to a specific phase change temperature range, assuming it is 30-60 degrees Celsius. The original collected temperature control parameter data may have a wide range, and its numerical distribution may be unfavorable for subsequent data analysis and neural network processing. Dynamic range compression is performed by linearly scaling the original temperature data based on the aqueous gel phase change temperature range. For example, the minimum value of the original temperature data is 10 degrees Celsius and the maximum value is 80 degrees Celsius. It is linearly mapped to the interval of 0-1, so that the temperature control parameters are represented in a more appropriate range, and a compressed temperature control parameter sequence is generated.

[0075] The pressure feedback parameters are affected by factors such as the vibration of the process equipment, which will generate noise. In the aqueous gel production workshop, equipment such as reactors and agitators will vibrate during operation, and this vibration will interfere with the accuracy of the pressure feedback parameters. An adaptive Kalman filter algorithm based on the vibration frequency of the process equipment is used for noise filtering. The algorithm automatically adjusts the filter parameters according to the characteristics of the equipment vibration frequency. For example, if the vibration frequency of the reactor is 50Hz and the vibration frequency of the agitator is 60Hz, the algorithm will comprehensively consider these frequency information, filter the collected pressure feedback parameters, remove the noise caused by factors such as vibration, and thus generate a smooth pressure feedback parameter sequence.

[0076] Fluid flow rate monitoring parameters may have different sampling frequencies because the monitoring equipment at different locations may have different sampling frequencies, which will cause data inconsistency in time. For example, the sampling frequency of the flow rate monitoring equipment at the raw material input pipeline is once every 5 seconds, while the sampling frequency of the flow rate monitoring equipment at the reactor outlet is once every 10 seconds. The flow rate data with these different sampling frequencies are unified into a preset timestamp sequence through an interpolation algorithm. Assuming that the preset timestamp sequence is a time point every 2 seconds, the flow rate value corresponding to each time point is calculated through an interpolation algorithm to generate a synchronous fluid flow rate parameter sequence.

[0077] For example, the adjacency matrix corresponding to the fluid flow topology structure of the first benchmark system operation example is constructed, including: A cross-parameter covariance matrix between the temperature control parameter and the pressure feedback parameter in the process parameter time series is extracted to generate a first process correlation index.

[0078] The dynamic mutual information between the fluid flow rate monitoring parameter and the gel consistency detection parameter is extracted to generate a second process correlation index.

[0079] The first process correlation index and the second process correlation index are integrated to generate a comprehensive coupling strength coefficient between production control units.

[0080] The edge weights of the adjacency matrix are constructed according to the comprehensive coupling strength coefficient, wherein node pairs with edge weights greater than a preset threshold are retained as strongly coupled connections, and node pairs with edge weights less than the preset threshold are sparsely processed.

[0081] In this embodiment, from the process parameter time series, there is a certain correlation between the temperature control parameter and the pressure feedback parameter. For example, in the reaction process of aqueous gel production, the change in temperature will affect the pressure in the reactor, which is then reflected in the pressure feedback parameter. The first process correlation index is generated by extracting the cross-parameter covariance matrix between the temperature control parameter and the pressure feedback parameter. If the element value in the covariance matrix is ​​large, it means that the correlation between the two parameters is strong.

[0082] At the same time, there is a dynamic relationship between the fluid flow rate monitoring parameter and the gel consistency detection parameter. During the gel formation process, the flow rate of the fluid will affect the formation of the gel, and then affect the gel consistency. The second process correlation index is generated by calculating the dynamic mutual information between the fluid flow rate monitoring parameter and the gel consistency detection parameter. The dynamic mutual information can reflect the degree of mutual dependence between the two parameters.

[0083] Then, the first process relevance index and the second process relevance index are fused to generate the comprehensive coupling strength coefficient between the production control units. For example, if the value of the first process relevance index is 0.6 and the value of the second process relevance index is 0.4, the comprehensive coupling strength coefficient is obtained according to a certain fusion rule (such as weighted average). The edge weights of the adjacency matrix are constructed based on this comprehensive coupling strength coefficient. In the aqueous gel production system, if the comprehensive coupling strength coefficient between two production control units (such as a reactor and an agitator) is greater than a preset threshold (assuming it is 0.5), the two node pairs are retained as strongly coupled connections, indicating that the process parameters between them are closely related; if it is less than the preset threshold, the node pair is sparsely processed, which means that the connection relationship between them is weakened in the topological structure.

[0084] For example, generating a feature matrix corresponding to the adjacency matrix includes: The temperature control parameters of each production control unit are statistically analyzed in time series, and the maximum temperature fluctuation value, average heating rate and steady-state temperature holding time are extracted as the first eigenvector.

[0085] The pressure feedback parameters are subjected to frequency domain energy analysis, and the low-frequency pressure oscillation amplitude, high-frequency noise energy proportion, and pressure mutation event count are extracted as the second eigenvector.

[0086] The dynamic mode decomposition of the fluid velocity monitoring parameters is performed, and the laminar state duration, turbulence intensity index and flow velocity periodic change frequency are extracted as the third eigenvector.

[0087] The outlier detection is performed on the gel consistency detection parameters, and the number of abnormal viscosity fluctuations, the steady-state viscosity variance and the transient viscosity gradient are extracted as the fourth eigenvector.

[0088] The first eigenvector, the second eigenvector, the third eigenvector and the fourth eigenvector are vertically spliced ​​according to the production control unit identifier to generate the characteristic matrix.

[0089] During the production process of aqueous gel, the temperature control parameters of each production control unit are statistically analyzed in time series. Taking the reactor as an example, it may be divided into multiple time periods for analysis during the entire production cycle. By analyzing the temperature data in each time period, the maximum temperature fluctuation value, the average heating rate, and the steady-state temperature holding time are extracted as the first eigenvector. For example, in a certain time period, the maximum temperature fluctuation value is 5 degrees Celsius, the average heating rate is 1 degree Celsius / minute, and the steady-state temperature holding time is 10 minutes. These values ​​constitute the first eigenvector.

[0090] The pressure feedback parameters are subjected to frequency domain energy analysis. In the pressure feedback data of the reactor, the low-frequency pressure oscillation amplitude may reflect the slow change process of the material in the reactor, the high-frequency noise energy ratio can reflect the degree of external interference, and the pressure mutation event count can represent the sudden pressure change during the reaction process. The low-frequency pressure oscillation amplitude, high-frequency noise energy ratio, and pressure mutation event count are extracted as the second eigenvector through frequency domain analysis.

[0091] Dynamic mode decomposition is performed on the fluid flow rate monitoring parameters. During the flow of aqueous gel from raw material mixing to the final product, laminar and turbulent states may occur in the pipeline. By analyzing the fluid flow rate monitoring parameters, the laminar state duration, turbulence intensity index, and flow rate periodic change frequency are extracted as the third eigenvector. For example, the laminar state duration is 20 minutes, the turbulence intensity index is 0.3, and the flow rate periodic change frequency is 0.1Hz, which constitute the third eigenvector.

[0092] Outlier detection is performed on the gel consistency detection parameters. During the production process of the gel, the gel consistency should be kept within a certain range. If outliers appear, it may indicate that the production process is abnormal. Through the outlier detection algorithm, the number of abnormal consistency fluctuations, the steady-state consistency variance, and the transient consistency gradient are extracted as the fourth eigenvector. For example, the number of abnormal consistency fluctuations is 3 times, the steady-state consistency variance is 0.2, and the transient consistency gradient is 0.5, which constitute the fourth eigenvector. Finally, the first eigenvector, the second eigenvector, the third eigenvector, and the fourth eigenvector are vertically spliced ​​according to the production control unit identifier to generate a feature matrix.

[0093] For example, the process topology feature extraction is performed on the adjacency matrix and the feature matrix by using the graph convolution kernel to generate an initial convolution feature graph, including: The edge weights of the adjacency matrix are multiplied element by element with the column vectors of the feature matrix to generate a weighted feature matrix.

[0094] A sliding window convolution operation is performed on the weighted feature matrix in the row direction, wherein the convolution window size is consistent with the maximum number of directly coupled nodes of the production control unit.

[0095] Independent feature extraction is performed on different process parameter channels through multi-scale convolution kernels to generate temperature control channel characteristic maps, pressure feedback channel characteristic maps, fluid flow rate channel characteristic maps and gel consistency channel characteristic maps.

[0096] Each channel feature map is concatenated according to the column dimension to generate the initial convolution feature map.

[0097] In this embodiment, the edge weights of the adjacency matrix are first element-by-element multiplied with the column vectors of the feature matrix to generate a weighted feature matrix. In the aqueous gel production control system, it is assumed that the edge weights of the adjacency matrix represent the strength of association between the production control units, and the column vectors of the feature matrix contain the different process parameter characteristics of each production control unit. This strength of association is incorporated into the feature matrix through element-by-element multiplication. For example, if the edge weight is 0.8 and the column vector elements of the feature matrix are [1, 2, 3], the weighted column vector is [0.8, 1.6, 2.4], thereby generating a weighted feature matrix.

[0098] The weighted feature matrix is ​​subjected to a sliding window convolution operation in the row direction, where the convolution window size is consistent with the maximum number of directly coupled nodes of the production control unit. In the aqueous gel production system, if a production control unit (such as a reactor) is directly coupled with at most three other production control units, then the convolution window size is 3. The local process topology features can be extracted through the sliding window convolution operation.

[0099] Through multi-scale convolution kernels, independent feature extraction is performed on different process parameter channels to generate temperature control channel feature maps, pressure feedback channel feature maps, fluid flow rate channel feature maps, and gel consistency channel feature maps. In the production process of aqueous gels, different process parameter channels have different effects on the production process. For example, the temperature control channel feature map may reflect the distribution and change characteristics of temperature during the production process, and the pressure feedback channel feature map reflects the relevant characteristics of pressure. Finally, the feature maps of each channel are spliced ​​according to the column dimension to generate the initial convolution feature map.

[0100] For example, the cross-node aggregation of process parameter features of adjacent production control units to generate an aggregate feature matrix includes: According to the edge weights between nodes in the adjacency matrix, the process coupling contribution between the target node and each neighbor node is calculated.

[0101] Based on the process coupling contribution, the feature vectors of the neighboring nodes are weighted summed to generate the primary aggregation feature of the target node.

[0102] The original feature vector of the target node is fused with the primary aggregated feature through a gating mechanism, wherein the gating weight is determined by the dynamic relationship between the temperature control parameter and the fluid flow rate monitoring parameter.

[0103] The fused feature vectors are standardized to generate the aggregated feature matrix.

[0104] In this embodiment, the process coupling contribution of the target node and each neighbor node is calculated according to the edge weights between nodes in the adjacency matrix. In the aqueous gel production system, the reactor is the target node and the agitator is the neighbor node. If the edge weight between them is 0.6, then the process coupling contribution of the agitator to the reactor may be 0.6 (the specific calculation may be based on the actual coupling relationship model).

[0105] Based on the process coupling contribution, the feature vectors of neighboring nodes are weighted and summed to generate the primary aggregate feature of the target node. For example, the feature vector of the agitator is [1,2,3], and the process coupling contribution is 0.6, then the primary aggregate feature after weighted summation is [0.6,1.2,1.8].

[0106] The original feature vector of the target node is fused with the primary aggregate feature by a gating mechanism, where the gating weight is determined by the dynamic relationship between the temperature control parameter and the fluid flow rate monitoring parameter. Assuming that there is a certain functional relationship between the temperature control parameter and the fluid flow rate monitoring parameter during the production of the aqueous gel, the gating weight is determined to be 0.4 based on this relationship. The original feature vector of the target node [2, 3, 4] is fused with the primary aggregate feature [0.6, 1.2, 1.8] according to the gating weight.

[0107] The fused feature vectors are standardized to generate an aggregate feature matrix. The standardized processing makes the value of the feature vector within an appropriate range, which is convenient for subsequent calculation and analysis.

[0108] For example, the step of inputting the aggregated feature matrix into the nonlinear activation module for feature nonlinear mapping includes: For each feature dimension in the polymerization feature matrix, its relative offset from the steady-state threshold of aqueous gel production is calculated.

[0109] The slope parameter of the activation function is dynamically adjusted based on the relative offset, wherein a high-sensitivity activation function is used for a feature dimension whose offset exceeds a safe range, and a smooth saturation activation function is used for a feature dimension whose offset is within a safe range.

[0110] The adjusted activation function is applied to each element of the aggregated feature matrix to generate a conversion result after nonlinear mapping.

[0111] In this embodiment, the relative offset of each characteristic dimension in the polymerized characteristic matrix from the steady-state threshold of aqueous gel production is calculated. For example, in the characteristic dimension of gel consistency, assuming that the steady-state consistency threshold of aqueous gel production is 0.8 and the characteristic value of gel consistency in the polymerized characteristic matrix is ​​0.9, then the relative offset is (0.9-0.8) / 0.8=0.125.

[0112] The slope parameter of the activation function is dynamically adjusted based on the relative offset, where the feature dimension whose offset exceeds the safety range uses a high-sensitivity activation function, and the feature dimension whose offset is within the safety range uses a smooth saturation activation function. If the relative offset of the gel consistency exceeds the set safety range (assuming it is 0.1), a high-sensitivity activation function is used for the feature dimension corresponding to the gel consistency to better capture the changes in this feature; if it is within the safety range, a smooth saturation activation function is used. The adjusted activation function is applied to each element of the aggregated feature matrix to generate a conversion result after nonlinear mapping.

[0113] For example, before using the conversion result as the input of the next graph convolution conversion layer, the method further includes: The conversion results are evaluated for feature importance to generate a feature contribution score for each production control unit.

[0114] The edge weights of the adjacency matrix are dynamically modified according to the feature contribution score, wherein the edge weights between nodes whose feature contribution is lower than a threshold are downgraded.

[0115] The graph convolution kernel parameters of the next graph convolution transformation layer are updated based on the corrected adjacency matrix to adapt to the dynamically changing fluid flow topology.

[0116] In this embodiment, for example, the cosine similarity between the feature vector of a certain production control unit and the global feature mean is 0.8, which is used as part of the first importance index. At the same time, the coefficient of variation of the feature vectors of each production control unit in historical abnormal events is counted to generate a second importance index. Assume that the coefficient of variation of the feature vector of a certain production control unit in historical abnormal events is 0.3. Then the first importance index and the second importance index are fused to generate a feature contribution score, in which the fusion weight is dynamically adjusted according to the safety level of the current production stage. For example, if the current production stage is at a high safety level, the fusion weight of the first importance index is 0.7, and the fusion weight of the second importance index is 0.3. The feature contribution score is calculated based on this weight.

[0117] The edge weights of the adjacency matrix are dynamically modified according to the feature contribution score, and the edge weights between nodes with feature contribution below the threshold are downgraded. If the feature contribution score between a node is lower than the preset threshold (assuming it is 0.3), the edge weight between them is reduced from the original value (for example, 0.5) to 0.2.

[0118] The graph convolution kernel parameters of the next graph convolution transformation layer are updated based on the corrected adjacency matrix to adapt to the dynamically changing fluid flow topology. For example, if the edge weights between a production control unit and other production control units change, the parameters corresponding to this connection relationship in the graph convolution kernel will also be updated accordingly to better adapt to the changes in the fluid flow topology that may occur during the production of aqueous gels.

[0119] For example, the feature importance evaluation of the conversion result is performed to generate a feature contribution score for each production control unit, including: The cosine similarity between the feature vector of each production control unit in the conversion result and the global feature mean is calculated to generate a first importance index.

[0120] The coefficient of variation of the characteristic vector of each production control unit in historical abnormal events is counted to generate the second importance index.

[0121] The first importance indicator and the second importance indicator are fused to generate the feature contribution score, wherein the fusion weight is dynamically adjusted according to the safety level of the current production stage.

[0122] For example, the step of generating a graph expression vector of the first benchmark system running sample through cascade processing of Y graph convolution transformation layers includes: A cross-layer feature fusion module is inserted between every two adjacent graph convolution transformation layers to perform a residual connection between the aggregated feature matrix of the previous layer and the initial convolution feature map of the next layer.

[0123] The conversion results of each layer in the cascade processing are summarized in an attention-weighted manner, where the attention weight is determined by the real-time change trend of the gel consistency detection parameter.

[0124] The conversion result of the final layer is globally averaged pooled to generate the graph expression vector, wherein the dimension of the graph expression vector matches the requirements of the diagnosis prediction task of the abnormal state chain.

[0125] A cross-layer feature fusion module is inserted between every two adjacent graph convolution transformation layers, and the aggregated feature matrix of the previous layer is residually connected with the initial convolutional feature map of the next layer. In the water-based gel production control system, this cross-layer feature fusion and residual connection helps to retain the useful information of the previous layer while fusing the new features of the next layer. For example, the aggregated feature matrix of the previous layer contains some feature information about the production control unit, and the initial convolutional feature map of the next layer contains new process topology features, which are combined through residual connections.

[0126] The conversion results of each layer in the cascade process are summarized with attention weights, where the attention weights are determined by the real-time change trend of the gel consistency detection parameters. If the real-time change trend of the gel consistency detection parameters indicates that the gel consistency is changing rapidly, the conversion results of the process parameter channels related to the gel consistency will be given higher attention weights when summarized to better reflect the current production status.

[0127] The conversion results of the final layer are globally averaged pooled to generate a graph expression vector, where the dimension of the graph expression vector matches the requirements of the abnormal state chain diagnosis prediction task. For example, in the abnormal state chain diagnosis prediction task of aqueous gel production, the graph expression vector may need to have a specific dimension to represent various state information in the production process. The conversion results of the final layer are converted into a graph expression vector that meets the requirements through global average pooling.

[0128] In a possible implementation, the first sample system operation data includes a first baseline system operation sample, a first associated system operation sample and a first non-associated system operation sample, the i-1th prior anomaly diagnosis neural network is a dual graph convolution module, and the i-th prior anomaly diagnosis neural network is an interactive graph convolution module.

[0129] In this embodiment, the first benchmark system operation sample can be regarded as the standard operation in the aqueous gel production process, for example, the raw materials are put into the reactor according to the precise ratio, the temperature in the reactor is accurately controlled according to the predetermined heating curve, the stirring speed is stably maintained at an appropriate value, the gel reaches the specified consistency at the appropriate time, and the subsequent packaging process is also normal. All parameters in the whole process are within the ideal normal range. The first associated system operation sample is a situation that is directly related to the operation of the benchmark system. For example, when the temperature control system of the reactor fluctuates slightly, the changes in related parameters such as the stirring speed, gel consistency, and sealing pressure in the packaging process are associated with it. The first non-associated system operation sample is a data situation that is not directly related to the operation of the benchmark system. For example, when equipment maintenance is performed in another area of ​​the production workshop, although the operation data of the maintenance equipment has no direct connection with the core process of gel production, it is collected as part of the operation data of the entire system.

[0130] Step S131 includes: Step S1311, aggregating the first benchmark system operation sample and the first associated system operation sample to generate a first interactive system operation sample, and aggregating the first benchmark system operation sample and the first non-associated system operation sample to generate a second interactive system operation sample.

[0131] In this embodiment, for the i-th prior abnormal diagnosis neural network, the first operation performed is to aggregate related system operation samples. The first benchmark system operation sample and the first associated system operation sample are aggregated to generate a first interactive system operation sample. In the production of aqueous gels, this aggregation operation has practical significance. For example, the normal change curve of the reactor temperature in the benchmark system operation sample is combined with the stirring speed adjustment data caused by temperature fluctuations in the associated system operation sample to form a first interactive system operation sample that more comprehensively reflects the relationship between temperature and stirring speed. Similarly, the first benchmark system operation sample and the first non-associated system operation sample are aggregated to generate a second interactive system operation sample. For example, the normal change data of gel consistency in the benchmark system is combined with the environmental parameters during equipment maintenance in the non-associated system (such as workshop humidity, etc., although not directly related, it may affect gel production to a certain extent) to form a second interactive system operation sample.

[0132] Step S1312: normalize the first interactive system operation sample and the second interactive system operation sample respectively to generate network learning data of the first interactive system operation sample and network learning data of the second interactive system operation sample.

[0133] For the first interactive system operation sample, it contains a variety of data from different aspects. For example, in the first interactive system operation sample, there are both temperature and pressure data of the reactor, stirring speed and gel consistency data, and other related parameters involved due to the association relationship. When normalizing these data, special processing is required for temperature data because its numerical range and change law are different from other parameters. For example, if the normal temperature range is between 20-80 degrees Celsius, it may be mapped to a specific interval, such as between 0-1, according to the specific neural network processing requirements. At the same time, the pressure data is filtered to remove noise, and the stirring speed and gel consistency data are standardized so that they are within the same order of magnitude, thereby generating network learning data for the first interactive system operation sample. Similar processing is also performed on the second interactive system operation sample, such as uniformly normalizing data such as workshop humidity and data related to gel production, such as converting humidity data into dimensionless values ​​according to a certain ratio, and together with data such as gel consistency, it constitutes the network learning data for the second interactive system operation sample.

[0134] Step S1313, using the i-th prior abnormality diagnosis neural network to perform graph convolution conversion on the network learning data of the first interactive system operation sample and the network learning data of the second interactive system operation sample, respectively, to generate a graph expression vector of the first interactive system operation sample and a graph expression vector of the second interactive system operation sample.

[0135] When the network learning data of the first interactive system running sample is converted into graph convolution, the i-th prior anomaly diagnosis neural network (interactive graph convolution module) will operate according to the graph structure characteristics of the data. In the water-based gel production control system, the entire production process can be regarded as a graph structure, and each production link (such as reactor, agitator, packaging equipment, etc.) is a node in the graph, and the process association between the links (such as the effect of temperature on gel consistency, the relationship between stirring speed and reactor pressure, etc.) is the edge in the graph. The graph convolution conversion extracts new features based on the connection relationship between nodes and the characteristics of the nodes themselves (such as the temperature of the reactor, stirring speed, etc.). After a series of complex calculations, the graph expression vector of the first interactive system running sample is generated. This vector contains the feature information about the first interactive system running sample after the graph convolution conversion, for example, it may contain the relationship characteristics between the temperature and stirring speed in different areas of the reactor, the coupling characteristics of the raw materials with the temperature and stirring speed at different reaction stages, etc. Similarly, after the network learning data of the second interactive system running sample is converted into graph convolution, the graph expression vector of the second interactive system running sample is obtained.

[0136] Step S1314: Use the graph expression vector of the first interactive system running sample and the graph expression vector of the second interactive system running sample as the graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

[0137] In this embodiment, the two graph expression vectors comprehensively reflect the feature extraction results of the i-th prior abnormality diagnosis neural network on the first sample system operation data after specific aggregation, normalization processing and graph convolution transformation. These results will be subsequently compared with the relevant results of the basic abnormality diagnosis neural network after the i-1th round of guided learning, so as to achieve guided learning of the basic abnormality diagnosis neural network, thereby improving the diagnostic and predictive capabilities of the entire abnormality diagnosis system for the abnormal state chain of the aqueous gel production control system.

[0138] In a possible implementation, step S133 may include: Step S1331, determining a first error term according to the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning.

[0139] Step S1332, determining a second error term according to the graph convolution conversion result of the i-th prior abnormality diagnosis neural network and the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1-th round of guided learning.

[0140] Step S1333: fuse the first error term and the second error term to generate a global training error.

[0141] Among them, the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning includes the graph expression vector of the first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, the graph expression vector of the first associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, and the graph expression vector of the first non-associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning.

[0142] Step S1331 includes: Step S1331-1, calculate the feature distance between the graph expression vector of the first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph expression vector of the first associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, and generate a first feature distance.

[0143] Step S1331-2, calculate the feature distance between the graph expression vector of the first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph expression vector of the first non-associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, and generate a second feature distance.

[0144] Step S1331 - 3 , determining a first error term according to the first characteristic distance and the second characteristic distance.

[0145] Wherein, step S1332 includes: Step S1332-1, determining a first abnormality diagnosis result of the i-th prior abnormality diagnosis neural network according to the graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

[0146] Step S1332-2, determining a second abnormality diagnosis result of the basic abnormality diagnosis neural network after the i-1th round of guided learning according to the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning.

[0147] Step S1332-3: determine a second error term based on the first abnormality diagnosis result and the second abnormality diagnosis result.

[0148] In this embodiment, the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning includes the graph expression vector of the first benchmark system operation sample, the graph expression vector of the first associated system operation sample, and the graph expression vector of the first non-associated system operation sample.

[0149] For the graph expression vector of the first benchmark system operation sample and the graph expression vector of the first associated system operation sample, a characteristic distance calculation is performed to generate a first characteristic distance. In the production process of aqueous gel, the first benchmark system operation sample contains a graph structure characteristic representation of various parameter data under the normal production process, such as the graph expression vector formed by the mutual relationship between parameters such as the temperature in the reactor, the stirring speed, and the gel consistency during normal production. The first associated system operation sample contains a graph expression vector of the parameter relationship under the influence of certain associated factors, such as the graph expression vector formed by the new relationship between these related parameters after the stirring speed is adjusted due to the temperature fluctuation of the reactor. Calculate the characteristic distance between them, for example, by using the Euclidean distance calculation method, take the square root of the sum of the squares of the differences between the corresponding elements in the two graph expression vectors, and obtain a value as the first characteristic distance. This value reflects the distance between the two graph expression vectors in the feature space, that is, the degree of difference between the normal operation of the benchmark system and the operation of the associated system after the graph convolution conversion by the basic abnormal diagnosis neural network.

[0150] Similarly, the feature distance calculation is performed on the graph expression vector of the first benchmark system operation sample and the graph expression vector of the first non-associated system operation sample of the basic abnormal diagnosis neural network after the i-1th round of guided learning to generate a second feature distance. The graph expression vector of the first non-associated system operation sample contains the graph structure characteristics of data that are not directly related to the benchmark system, such as the graph expression vector composed of a certain relationship between the environmental parameters (such as humidity, equipment noise, etc.) during equipment maintenance in another area of ​​the production workshop and the benchmark parameters for gel production. The feature distance of these two graph expression vectors is calculated using a similar calculation method. The second feature distance obtained reflects the degree of difference between the basic abnormal diagnosis neural network for the benchmark system and the non-associated system after graph convolution conversion.

[0151] The first error term is determined based on the first characteristic distance and the second characteristic distance. This determination process may be achieved through a specific functional relationship, such as weighted summing the first characteristic distance and the second characteristic distance, and the weight may be determined based on the relative importance of the associated system and the non-associated system to the overall production control during the production of aqueous gel. If the operating status of the associated system has a greater impact on the quality of gel production, the weight of the first characteristic distance may be set higher; if the non-associated system does not directly affect gel production, but has a certain impact on the overall production environment, then the second characteristic distance will also have a corresponding weight. The first error term is obtained through this weighted summation, which reflects the degree of difference in the graph convolution conversion results between different types of system operation samples of the basic abnormal diagnosis neural network.

[0152] Next, a second error term is determined based on the graph convolution transformation result of the i-th prior abnormality diagnosis neural network and the graph convolution transformation result of the basic abnormality diagnosis neural network after the i-1-th round of guided learning.

[0153] The first abnormality diagnosis result of the i-th prior abnormality diagnosis neural network is determined based on the graph convolution conversion result of the i-th prior abnormality diagnosis neural network. After the i-th prior abnormality diagnosis neural network performs graph convolution conversion on the operating data of the aqueous gel production control system, its result can be used for abnormality diagnosis. For example, in the graph convolution conversion result, certain features may indicate that the relationship between the temperature and the stirring speed in the reactor exceeds the normal range, or that the change trend of the gel consistency does not meet the expectations of normal production. Based on these features, the first abnormality diagnosis result is determined according to pre-set diagnostic rules. These diagnostic rules may be established based on a large amount of historical data on aqueous gel production and known abnormal situations. For example, if the ratio of temperature to stirring speed exceeds a certain threshold and lasts for a certain period of time, it is determined to be an abnormal situation; if the change rate of gel consistency in a certain production stage exceeds a certain multiple of the normal change rate, it is also determined to be an abnormal situation, etc.

[0154] The second abnormality diagnosis result of the basic abnormality diagnosis neural network after the i-1th round of guided learning is determined based on the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning. Similarly, after the graph convolution conversion, the basic abnormality diagnosis neural network determines the second abnormality diagnosis result according to its result and similar diagnostic rules. For example, in the graph expression vector of the basic abnormality diagnosis neural network, the relationship between the pressure of the reactor and the gel consistency is analyzed to see whether it conforms to the normal production mode. If not, it is determined to be an abnormal situation; or whether there is an abnormality is determined based on whether the relationship between the raw material input flow rate and the reactor temperature is within the normal range.

[0155] The second error term is determined based on the first abnormal diagnosis result and the second abnormal diagnosis result. This determination process may involve comparing the degree of difference between the two diagnostic results. For example, if the first abnormal diagnosis result determines that there is a certain abnormality in a certain production link, while the second abnormal diagnosis result determines that the link is normal, this indicates that there is a large difference between the two. This difference can be quantified by calculating the degree of inconsistency between the two diagnostic results on all possible diagnostic items. For example, assuming that there are 10 diagnostic items, and the two diagnostic results are inconsistent on 5 of them, then a weighted calculation can be performed based on the importance of these 5 inconsistent items in the entire diagnosis to obtain a numerical value as the second error term. This second error term reflects the degree of difference in abnormal diagnostic results between the i-th prior abnormal diagnosis neural network and the basic abnormal diagnosis neural network after the i-1th round of guided learning.

[0156] Finally, the first error term and the second error term are fused to generate a global training error. This fusion process may be achieved through a specific weighted summation or other mathematical operation. For example, since the first error term reflects the difference between the graph convolution conversion results of different system operation samples, and the second error term reflects the difference in the abnormal diagnosis results, the weight is determined according to the relative importance of these two differences for the overall training in the aqueous gel production control system. If it is believed that the consistency of the graph convolution conversion results is more basic and important for the training of the neural network, the weight of the first error term may be higher; if the accuracy of the abnormal diagnosis results is more critical for production control, the weight of the second error term may be adjusted accordingly. The global training error is obtained through this weighted summation or other fusion method. This global training error will be used to optimize the neuron weight information of the basic abnormal diagnosis neural network after the i-1th round of guided learning to improve its diagnostic prediction ability for the abnormal state chain of the aqueous gel production control system.

[0157] In a possible implementation manner, the first sample system operation data includes a first reference system operation sample, a first associated system operation sample, and a first non-associated system operation sample.

[0158] Step S1332-1 includes: Step S1332-11, based on the graph convolution conversion result of the i-th prior abnormality diagnosis neural network, determine the third characteristic distance between the first benchmark system operation sample and the first associated system operation sample, and determine the fourth characteristic distance between the first benchmark system operation sample and the first non-associated system operation sample.

[0159] Step S1332-12, performing standardization conversion on the third characteristic distance and the fourth characteristic distance to generate a fifth characteristic distance.

[0160] Step S1332-13: output the ratio of the third characteristic distance to the fifth characteristic distance as a first abnormality diagnosis confidence, and output the ratio of the fourth characteristic distance to the fifth characteristic distance as a second abnormality diagnosis confidence.

[0161] Step S1332-14: output the first abnormality diagnosis confidence and the second abnormality diagnosis confidence as the first abnormality diagnosis result of the i-th prior abnormality diagnosis neural network.

[0162] In this embodiment, the first benchmark system operation sample reflects the standard operation mode in the production process of aqueous gel, including various parameters and their interrelationships in the complete normal production process, such as the raw materials are added to the reactor in a specific proportion, the temperature in the reactor is accurately controlled according to a predetermined curve, the stirring speed is stabilized at an appropriate value, the gel reaches a specified consistency, and the packaging process is normal. The first associated system operation sample reflects the situation that is directly related to the operation of the benchmark system, such as when the temperature control system of the reactor fluctuates slightly, the changes in related parameters such as the stirring speed, gel consistency, and sealing pressure in the packaging process. The first non-associated system operation sample includes data that is not directly related to the operation of the benchmark system, such as the equipment operation data when the equipment is maintained in another area of ​​the production workshop. Although it has no direct connection with the core process of gel production, it is collected as part of the entire system operation data.

[0163] According to the graph convolution conversion result of the i-th prior abnormal diagnosis neural network, the third characteristic distance between the first reference system operation sample and the first associated system operation sample is first determined. In the aqueous gel production process, the graph convolution conversion result of the i-th prior abnormal diagnosis neural network contains the characteristic representation of each system operation sample. The graph convolution conversion result of the first reference system operation sample reflects the characteristic representation of the relationship between each parameter under the normal production process, such as the characteristic representation of the relationship between parameters such as the temperature in the reactor, the stirring speed, and the gel consistency during normal production. The graph convolution conversion result of the first associated system operation sample reflects the characteristic representation of the relationship between each parameter under the influence of the associated factors, such as the characteristic representation of the relationship between the related parameters after the stirring speed is adjusted due to temperature fluctuations. To calculate the third characteristic distance between the two graph convolution conversion results, a specific distance calculation method, such as the Euclidean distance calculation method, can be used. Each characteristic element in the graph convolution conversion result of the first reference system operation sample is subtracted from the corresponding characteristic element in the graph convolution conversion result of the first associated system operation sample, and then the square of the difference is summed, and finally the square root is taken to obtain the third characteristic distance. This distance value reflects the degree of difference between the baseline system and the associated system under the feature representation of the i-th prior anomaly diagnosis neural network.

[0164] Similarly, the fourth characteristic distance between the first benchmark system operation sample and the first non-associated system operation sample is determined. The graph convolution transformation result of the first non-associated system operation sample contains the feature representation of data that is not directly related to the benchmark system, such as the feature representation of the relationship between the environmental parameters (such as humidity, equipment noise, etc.) during equipment maintenance in another area of ​​the production workshop and the benchmark parameters for gel production. The fourth characteristic distance between the first benchmark system operation sample and the first non-associated system operation sample is calculated using a similar Euclidean distance calculation method. This distance value reflects the degree of difference between the benchmark system and the non-associated system under the feature representation of the i-th prior abnormality diagnosis neural network.

[0165] The third characteristic distance and the fourth characteristic distance are standardized and transformed to generate the fifth characteristic distance. In the aqueous gel production control system, since the numerical range and calculation method of the third characteristic distance and the fourth characteristic distance may be different, in order to be able to compare and perform subsequent calculations under a unified scale, a standardized transformation is required. The standardized transformation method may be based on the statistical characteristics of the entire system operation data. For example, the mean and standard deviation of all possible characteristic distances are calculated, and then the third characteristic distance and the fourth characteristic distance are respectively subtracted from the mean and divided by the standard deviation to obtain a standardized value. The two standardized values ​​are combined or further processed to generate the fifth characteristic distance. This fifth characteristic distance is a representation of the difference between the first benchmark system operation sample and other samples (associated and non-associated) under a unified standard.

[0166] The ratio of the third characteristic distance to the fifth characteristic distance is output as the first abnormal diagnosis confidence, and the ratio of the fourth characteristic distance to the fifth characteristic distance is output as the second abnormal diagnosis confidence. In the abnormal diagnosis scenario of aqueous gel production, this ratio is of great significance. The first abnormal diagnosis confidence represents the proportional relationship between the difference between the first reference system operation sample and the first associated system operation sample in the i-th prior abnormal diagnosis neural network relative to the difference under the unified standard (fifth characteristic distance). If this ratio is high, it means that the difference between the first reference system operation sample and the first associated system operation sample is large, which may imply that the operation of the associated system has a large deviation from the normal state relative to the reference system, thereby reflecting the possibility of abnormality to a certain extent. Similarly, the second abnormal diagnosis confidence represents the proportional relationship between the difference between the first reference system operation sample and the first non-associated system operation sample relative to the difference under the unified standard. If this ratio is high, it also implies that there is a large difference between the operation of the non-associated system and the reference system. Although the non-associated system has no direct connection with the core process of gel production, this difference may have an impact on the overall production environment or other indirect aspects, and may also be a potential abnormal performance.

[0167] Finally, the first abnormal diagnosis confidence and the second abnormal diagnosis confidence are output as the first abnormal diagnosis result of the i-th prior abnormal diagnosis neural network. This first abnormal diagnosis result integrates the difference relationship between the first reference system operation sample and the first associated system operation sample and the first non-associated system operation sample, and is expressed in the form of confidence. In the abnormal diagnosis of the aqueous gel production control system, these confidences can be used as an important basis for subsequent judgment of whether there is an abnormality in the system, the severity of the abnormality, and taking corresponding measures. For example, if the first abnormal diagnosis confidence exceeds a certain preset threshold, it may be necessary to conduct a more detailed inspection or adjustment of the associated system; if the second abnormal diagnosis confidence exceeds the corresponding threshold, although it may not be necessary to directly intervene in the core process of gel production, it may be necessary to pay attention to whether the production environment or other indirect factors have a potential impact on production.

[0168] In a possible implementation, before step S110, the method further includes: Acquire second sample system operation data and an i-th initial priori abnormality diagnosis neural network, wherein the second sample system operation data includes a second reference system operation sample, a second associated system operation sample, and a second non-associated system operation sample.

[0169] The second benchmark system operation sample, the second associated system operation sample and the second non-associated system operation sample are respectively normalized to generate network learning data of the second benchmark system operation sample, network learning data of the second associated system operation sample and network learning data of the second non-associated system operation sample.

[0170] The i-th initial prior abnormal diagnosis neural network is used to perform graph convolution transformation on the network learning data of the second benchmark system operation sample, the network learning data of the second associated system operation sample, and the network learning data of the second non-associated system operation sample, to generate a graph expression vector of the second benchmark system operation sample, a graph expression vector of the second associated system operation sample, and a graph expression vector of the second non-associated system operation sample.

[0171] The i-th initial prior anomaly diagnosis neural network is trained based on the graph expression vector of the second benchmark system operation sample, the graph expression vector of the second associated system operation sample and the graph expression vector of the second non-associated system operation sample to generate the i-th prior anomaly diagnosis neural network.

[0172] In this embodiment, the second benchmark system operation example is another example of standard operation in the production process of aqueous gel. For example, in this example, starting from the source of raw material procurement, the quality and composition ratio of the raw materials meet specific high quality standards, and then in the production process, the heating system of the reactor accurately follows a specific power curve to increase the temperature to ensure that the temperature is in the ideal range at different reaction stages, and the stirring speed of the agitator maintains a stable and accurate value according to the progress of the reaction and the characteristics of the raw materials. During the formation of the gel, its consistency gradually increases at the expected rate until it reaches the perfect finished product consistency. The final packaging link also strictly follows all quality and specification requirements without any deviation. This complete set of production process parameters constitutes the second benchmark system operation example.

[0173] The second associated system operation sample reflects the operation conditions that are directly related to the second benchmark system operation. For example, when the reactor temperature control system in the second benchmark system fluctuates in a different mode from the previous one, this fluctuation may be due to new external environmental factors (such as a small fluctuation in the power supply, but still within an acceptable range) or a special state of the internal equipment (such as slight scaling of the reactor heat exchanger). This temperature fluctuation in turn affects other parameters associated with it, such as the stirring speed needs to be adjusted accordingly to maintain the normal progress of the reaction, the consistency of the gel may also fluctuate to a certain extent due to the combined changes in temperature and stirring speed, and the sealing pressure in the packaging process may also need to be fine-tuned to adapt to this slight change in the gel. The overall changes in these related parameters constitute the second associated system operation sample.

[0174] The second non-correlated system operation sample is a data situation that is not directly related to the operation of the second benchmark system. For example, in a remote corner of the production workshop, there is a set of equipment for environmental monitoring in operation. It monitors some environmental parameters in the workshop that are not directly related to the core process of gel production, such as the light intensity in a specific area, the dust content in the air, etc. Although these data are not related to the direct operation of gel production (such as the reaction in the reactor, the stirring and molding of the gel, etc.), they are recorded as part of the entire production environment, forming the second non-correlated system operation sample.

[0175] Then, the second benchmark system operation samples, the second associated system operation samples and the second non-associated system operation samples are respectively normalized to generate respective network learning data.

[0176] For the second benchmark system operation sample, it contains various types of parameter data, such as temperature, pressure, speed, consistency, etc. The magnitude, unit and distribution range of these data may be different. Taking temperature data as an example, its numerical range may be from room temperature to high temperature values ​​during the reaction process, such as 20-80 degrees Celsius. In order to make these data more suitable for neural network learning, they need to be normalized. A linear transformation method can be used to map the temperature data to a specific interval, such as between 0 and 1. The specific calculation method can be to calculate the corresponding value of each temperature value in the new interval through a simple linear function based on the actual temperature range. For pressure data, if its original unit is Pascal, the numerical range is large and the fluctuation is different from other parameters, it may be converted into a dimensionless relative pressure value so that it is in the same magnitude range as other normalized parameters. Similar processing is also performed for data such as stirring speed and gel consistency. By adjusting the numerical range, converting units, etc., all parameter data are processed into a form suitable for neural network processing, thereby generating network learning data for the second benchmark system operation sample.

[0177] For the second association system operation sample, since the parameter data therein is related to the data in the second benchmark system operation sample and has similar properties, it also needs to be normalized. For example, when the stirring speed data in the association system changes due to temperature fluctuations, its original range of change may be more complicated. In order to enable the neural network to better understand and process these data, the stirring speed data is normalized in the same way as the stirring speed data in the second benchmark system operation sample. For example, its actual speed range is converted into a relative numerical range so that it is consistent with other parameters in magnitude and representation, thereby generating network learning data for the second association system operation sample.

[0178] For the second non-correlated system operation sample, although its data is not directly related to the core process of gel production, it also needs to be normalized for unified processing and analysis under the overall system framework. Taking light intensity data as an example, its original measurement unit and value range may be completely different from the gel production parameters. Through a specific conversion method, the light intensity data is converted into a dimensionless relative value, and its value range matches the other normalized parameters. Similarly, the data such as air dust content is also converted accordingly to generate network learning data for the second non-correlated system operation sample.

[0179] Next, the i-th initial prior anomaly diagnosis neural network is used to perform graph convolution transformation on the network learning data of the second benchmark system operation sample, the network learning data of the second associated system operation sample, and the network learning data of the second non-associated system operation sample to generate their respective graph expression vectors.

[0180] When the network learning data of the second benchmark system running sample is converted to graph convolution, the i-th initial prior anomaly diagnosis neural network will process the data according to its internal graph convolution structure. In the water-based gel production control system, the entire production process can be regarded as a graph structure, in which each production link (such as raw material supply link, reactor, agitator, gel forming equipment, packaging equipment, etc.) is a node in the graph, and the relationship between the links (such as the effect of raw material supply on the reaction in the reactor, the effect of reactor temperature on gel consistency, the effect of stirring speed on gel forming, etc.) is the edge in the graph. The graph convolution conversion extracts new features based on the connection relationship (edge) between nodes and the characteristics of the nodes themselves (various parameter values ​​in the network learning data). For example, for the reactor node, its parameter characteristics such as temperature and pressure and the relationship with other nodes (such as agitator) will be processed by the graph convolution operation. After a series of complex graph convolution calculations, a graph expression vector of the second benchmark system operation sample is generated. This vector contains the characteristic information about the second benchmark system operation sample after graph convolution conversion, such as the relationship characteristics between the temperature and stirring speed in different areas of the reactor, the coupling characteristics of the raw materials with the temperature and stirring speed at different reaction stages, etc.

[0181] When the network learning data of the second associated system operation sample is converted into a graph convolution, since its data is associated with the data of the second benchmark system operation sample and is similar in form to the benchmark system data after normalization, the i-th initial prior anomaly diagnosis neural network also operates according to the graph convolution structure. For example, when the temperature fluctuation in the associated system causes the stirring speed to be adjusted, the graph convolution conversion will capture the impact of this change in the entire graph structure. By analyzing the relationship between the nodes related to the temperature fluctuation and the nodes related to the stirring speed adjustment and their mutual influence with other nodes (such as the gel consistency related nodes), the graph expression vector of the second associated system operation sample is generated through graph convolution calculation. This vector reflects the relationship characteristics between the parameters of each production link under the operation of the associated system after the graph convolution conversion.

[0182] When performing graph convolution transformation on the network learning data of the second non-correlated system operation sample, although its data has no direct correlation with the core process of gel production, the i-th initial prior anomaly diagnosis neural network will still perform graph convolution operation under the overall system operation data framework. For example, the environmental parameters in the non-correlated system (such as light intensity, air dust content, etc.) are regarded as special nodes in the graph structure. Although the relationship between them and the core production link nodes is relatively weak, there is still a certain potential connection (such as light intensity may affect the temperature distribution in the workshop, and then indirectly affect gel production). Through graph convolution transformation, the graph expression vector of the second non-correlated system operation sample is generated. This vector reflects the characteristic representation of the non-correlated system operation data in the graph structure.

[0183] Then, based on the graph expression vector of the second benchmark system operation sample, the graph expression vector of the second associated system operation sample and the graph expression vector of the second non-associated system operation sample, the i-th initial prior anomaly diagnosis neural network is trained to generate the i-th prior anomaly diagnosis neural network.

[0184] During the training process, these graph expression vectors are used to adjust the internal parameters of the i-th initial prior anomaly diagnosis neural network to improve its anomaly diagnosis capability for the aqueous gel production control system. For example, if the graph expression vector of the second benchmark system operation sample indicates that there is a specific relationship pattern between certain parameters under normal production conditions, and a deviation from this pattern appears in the graph expression vector of the second associated system operation sample, then the neural network will adjust its weight parameters according to this deviation, so that it can better identify the difference from the normal pattern, thereby improving the ability to diagnose abnormal conditions in the associated system. For the graph expression vector of the second non-associated system operation sample, although it has no direct correlation with the core production process, the neural network can be trained to learn the potential relationship between this non-associated data and the operating status of the entire system, so as to play a role in more comprehensive system monitoring and anomaly diagnosis.

[0185] In a possible implementation, before step S110, the method further includes: Acquire second sample system operation data and an i-th initial priori abnormality diagnosis neural network, wherein the second sample system operation data includes a second reference system operation sample, a second associated system operation sample, and a second non-associated system operation sample.

[0186] The second benchmark system operation sample and the second associated system operation sample are aggregated to generate a third interactive system operation sample, and the second benchmark system operation sample and the second non-associated system operation sample are aggregated to generate a fourth interactive system operation sample.

[0187] The third interactive system operation sample and the fourth interactive system operation sample are respectively normalized to generate network learning data of the third interactive system operation sample and network learning data of the fourth interactive system operation sample.

[0188] The i-th initial prior abnormal diagnosis neural network is used to perform graph convolution conversion on the network learning data of the third interactive system operation sample and the network learning data of the fourth interactive system operation sample, respectively, to generate a graph expression vector of the third interactive system operation sample and a graph expression vector of the fourth interactive system operation sample.

[0189] The i-th initial a priori abnormality diagnosis neural network is trained based on the graph expression vector of the third interactive system running sample and the graph expression vector of the fourth interactive system running sample to generate the i-th a priori abnormality diagnosis neural network.

[0190] For another method, after obtaining the second sample system operation data and the i-th initial prior abnormal diagnosis neural network, the second benchmark system operation sample and the second associated system operation sample are aggregated to generate a third interactive system operation sample, and the second benchmark system operation sample and the second non-associated system operation sample are aggregated to generate a fourth interactive system operation sample.

[0191] In the production of aqueous gels, the third interactive system operation sample is generated by combining the normal production process data in the second benchmark system operation sample with the associated data in the second associated system operation sample. For example, the normal temperature control data of the reactor in the second benchmark system is aggregated with the stirring speed adjustment data caused by temperature fluctuations in the second associated system. This aggregation is not just a simple data superposition, but a combination based on the actual logical relationship in the production process to form a third interactive system operation sample that more comprehensively reflects the interaction between the benchmark system and the associated system. Similarly, the fourth interactive system operation sample is generated by aggregating the normal production data in the second benchmark system with the non-directly associated data in the second non-associated system. For example, the normal data of gel consistency in the second benchmark system is combined with the workshop light intensity data in the second non-associated system. Considering that the light intensity may have an indirect effect on the workshop temperature, which may in turn affect the gel consistency, this combination forms the fourth interactive system operation sample.

[0192] Then, the third interactive system running sample and the fourth interactive system running sample are respectively normalized to generate respective network learning data.

[0193] For the third interactive system operation sample, since it contains different types of data from the benchmark system and the associated system, the normalization processing needs to comprehensively consider the characteristics of the two types of data. For example, in the third interactive system operation sample, there are both core production parameters such as the temperature and stirring speed of the reactor, as well as special adjustment data generated by the associated relationship. For temperature data, it may be necessary to perform unified normalization processing based on its normal range in the benchmark system and its fluctuation range in the associated system. For stirring speed data, it is necessary to take into account the normal stirring speed in the benchmark system and the adjustments made in the associated system due to temperature fluctuations, and convert it into a form suitable for neural network learning through a specific calculation method, thereby generating network learning data for the third interactive system operation sample.

[0194] For the fourth interactive system operation sample, which contains data from the benchmark system and the non-associated system, normalization also requires special methods. For example, the gel consistency data and workshop light intensity data in the fourth interactive system operation sample are normalized. For the gel consistency data, it is normalized in a similar manner to the gel consistency data in the separate benchmark system operation sample so that it meets the requirements in terms of numerical range and representation. For the light intensity data, it is converted into a dimensionless relative value that matches the gel consistency data, thereby generating the network learning data for the fourth interactive system operation sample.

[0195] Then, the network learning data of the third interactive system running sample and the network learning data of the fourth interactive system running sample are respectively converted into graph convolutions using the i-th initial prior anomaly diagnosis neural network to generate respective graph expression vectors.

[0196] When performing graph convolution transformation on the network learning data of the third interactive system operation sample, the i-th initial prior anomaly diagnosis neural network will process the data therein according to the graph convolution structure. Since the third interactive system operation sample contains data on the interaction between the reference system and the associated system, the graph convolution transformation will capture the characteristic representation of this interaction relationship in the graph structure. For example, by analyzing the relationship between the reactor temperature in the reference state and the associated fluctuation state and the stirring speed adjustment in the graph structure, the graph expression vector of the third interactive system operation sample is generated through graph convolution calculation. This vector reflects the relationship characteristics between the parameters of each production link in the interaction between the reference system and the associated system after graph convolution transformation.

[0197] When performing graph convolution transformation on the network learning data of the fourth interactive system operation sample, the i-th initial prior anomaly diagnosis neural network also operates according to the graph convolution structure. For example, in the fourth interactive system operation sample, there is a potential indirect relationship between the gel consistency and the light intensity in the workshop. The graph convolution transformation will calculate based on the representation of this relationship and other possible associations (such as light intensity indirectly affects gel consistency by affecting temperature) in the graph structure to generate a graph expression vector of the fourth interactive system operation sample. This vector reflects the characteristic representation of the interaction data between the benchmark system and the non-association system in the graph structure.

[0198] Finally, according to the graph expression vector of the third interactive system running sample and the graph expression vector of the fourth interactive system running sample, the i-th initial prior anomaly diagnosis neural network is trained to generate the i-th prior anomaly diagnosis neural network.

[0199] During this training process, the graph expression vector of the third interactive system operation sample and the graph expression vector of the fourth interactive system operation sample provide the neural network with feature information about the interaction between the baseline system and the associated system, and the baseline system and the non-associated system. The neural network adjusts its own parameters based on this feature information to improve its ability to diagnose abnormal situations in different interaction situations. For example, if the graph expression vector of the third interactive system operation sample shows that an abnormal pattern appears in a parameter relationship when the baseline system interacts with the associated system, the neural network will adjust the weight parameters to better identify this abnormality. For the graph expression vector of the fourth interactive system operation sample, the neural network will learn the potential abnormal features in the interaction between the baseline system and the non-associated system, so as to be more accurate and comprehensive in the overall abnormal diagnosis of the aqueous gel production control system.

[0200] Figure 2 The hardware structure of the intelligent production service system 100 provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the intelligent production service system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0201] In one possible design, the intelligent production service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the intelligent production service system 100 can be a distributed system). In some embodiments, the intelligent production service system 100 can be local or remote. For example, the intelligent production service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the intelligent production service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the intelligent production service system 100 can be implemented on an intelligent production service system. By way of example only, the intelligent production service system may include a private cloud, a semantically related cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0202] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the intelligent production service system 100 uses to execute or use to complete the exemplary methods described in this application.

[0203] In the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the abnormality analysis method applied to the aqueous gel production control system of the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0204] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned intelligent production service system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0205] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When a processor executes the computer executable instructions, the abnormality analysis method applied to the aqueous gel production control system as above is implemented.

[0206] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. An abnormality analysis method applied to a water-based gel production control system, characterized in that: The method comprises: Obtaining an initialized basic abnormality diagnosis neural network and X a priori abnormality diagnosis neural networks, wherein the a priori abnormality diagnosis neural networks have different abnormality diagnosis capabilities, wherein the basic abnormality diagnosis neural network and the X a priori abnormality diagnosis neural networks are used to diagnose and predict abnormal state chains of system operation data of the aqueous gel production control system; Using a first a priori abnormality diagnosis neural network to guide learning of the initialized basic abnormality diagnosis neural network, to generate a basic abnormality diagnosis neural network after a first round of guided learning, wherein the first a priori abnormality diagnosis neural network is a priori abnormality diagnosis neural network with the weakest abnormality diagnosis ability among the X a priori abnormality diagnosis neural networks; Using the i-th prior abnormality diagnosis neural network to guide the learning of the basic abnormality diagnosis neural network after the i-1th round of guided learning, to generate the basic abnormality diagnosis neural network after the i-1th round of guided learning, wherein i is an increasing integer greater than 1 and not greater than X, the basic abnormality diagnosis neural network after the i-1th round of guided learning is generated by using the i-1th prior abnormality diagnosis neural network to guide the learning, the i-th prior abnormality diagnosis neural network and the i-1th prior abnormality diagnosis neural network are any one of the X prior abnormality diagnosis neural networks, and the abnormality diagnosis ability of the i-th prior abnormality diagnosis neural network is stronger than the abnormality diagnosis ability of the i-1th prior abnormality diagnosis neural network; The basic abnormality diagnosis neural network after the X-th round of guided learning is output as a target abnormality diagnosis neural network, so as to perform diagnosis and prediction of abnormal state chains for any input system operation data through the target abnormality diagnosis neural network.

2. The abnormality analysis method applied to the aqueous gel production control system according to claim 1, characterized in that: The method of using the i-th prior abnormality diagnosis neural network to guide the learning of the basic abnormality diagnosis neural network after the i-1-th round of guided learning to generate the basic abnormality diagnosis neural network after the i-th round of guided learning includes: Using the i-th prior abnormality diagnosis neural network to perform graph convolution conversion on the first sample system operation data, to generate a graph convolution conversion result of the i-th prior abnormality diagnosis neural network; Performing graph convolution conversion on the first sample system operation data using the basic abnormality diagnosis neural network after the i-1th round of guided learning to generate a graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning; Determining a global training error according to a graph convolution conversion result of the basic abnormality diagnosis neural network after the (i-1)th round of guided learning and a graph convolution conversion result of the (i)th prior abnormality diagnosis neural network; According to the global training error, the neuron weight information of the basic abnormality diagnosis neural network after the i-1th round of guided learning is optimized to generate the basic abnormality diagnosis neural network after the i-th round of guided learning.

3. The abnormality analysis method applied to the aqueous gel production control system according to claim 2, characterized in that: The first sample system operation data includes a first reference system operation sample, a first associated system operation sample and a first non-associated system operation sample, and the i-1th prior anomaly diagnosis neural network and the i-th prior anomaly diagnosis neural network are both dual graph convolution modules; The step of performing graph convolution conversion on the first sample system operation data using the i-th priori abnormality diagnosis neural network to generate a graph convolution conversion result of the i-th priori abnormality diagnosis neural network includes: Normalizing the first benchmark system operation sample, the first associated system operation sample, and the first non-associated system operation sample respectively to generate network learning data of the first benchmark system operation sample, network learning data of the first associated system operation sample, and network learning data of the first non-associated system operation sample; Using the i-th prior abnormality diagnosis neural network, respectively, the network learning data of the first reference system operation sample, the network learning data of the first associated system operation sample, and the network learning data of the first non-associated system operation sample are converted into graph convolutions to generate a graph expression vector of the first reference system operation sample, a graph expression vector of the first associated system operation sample, and a graph expression vector of the first non-associated system operation sample; The graph expression vector of the first benchmark system operation sample, the graph expression vector of the first associated system operation sample and the graph expression vector of the first non-associated system operation sample are used as graph convolution conversion results of the i-th prior abnormality diagnosis neural network.

4. The abnormality analysis method applied to the aqueous gel production control system according to claim 3, characterized in that: The graph convolution conversion of the network learning data of the first benchmark system running sample is implemented by using the Y sequentially connected graph convolution conversion layers included in the i-th prior anomaly diagnosis neural network; Using a first graph convolution transformation layer among the Y sequentially connected graph convolution transformation layers, perform graph convolution transformation on the network learning data of the first benchmark system running sample to generate a transformation result of the first graph convolution transformation layer; Using the j-th graph convolution transformation layer, perform graph convolution transformation on the transformation result of the j-1-th graph convolution transformation layer to generate the transformation result of the j-th graph convolution transformation layer, and determine the graph expression vector of the first benchmark system running sample according to the transformation result of the Y-th graph convolution transformation layer, where j is an increasing integer greater than 1 and not greater than Y; The step of using the first graph convolution conversion layer among the Y sequentially connected graph convolution conversion layers to perform graph convolution conversion on the network learning data of the first benchmark system running sample to generate a conversion result of the first graph convolution conversion layer includes: Acquire a process parameter time series sequence corresponding to the first benchmark system operation sample, wherein the process parameter time series sequence includes continuously collected data of temperature control parameters, pressure feedback parameters, fluid flow rate monitoring parameters, and gel consistency detection parameters; Normalizing the process parameter timing sequence to generate a normalized process parameter timing sequence, wherein the normalization includes dynamic range compression of the temperature control parameter, noise filtering of the pressure feedback parameter, and time alignment of the fluid flow rate monitoring parameter; Based on the process correlation index between the process parameters in the normalized process parameter time series, construct an adjacency matrix corresponding to the fluid flow topology structure of the first benchmark system operation sample, wherein the nodes of the adjacency matrix represent production control units, and the edge weights represent the process parameter coupling strength between adjacent production control units; Based on the multidimensional parameter set of each production control unit in the normalized process parameter time series, a feature matrix corresponding to the adjacency matrix is ​​generated, wherein the row dimension of the feature matrix corresponds to the production control unit identifier, and the column dimension corresponds to the normalized eigenvalues ​​of the temperature control parameter, the pressure feedback parameter, the fluid flow rate monitoring parameter, and the gel consistency detection parameter; Combining the adjacency matrix with the characteristic matrix for multi-channel data to generate multi-channel graph data of the first benchmark system operation sample, wherein the multi-channel graph data includes a fluid flow topology channel, a parameter timing correlation channel, and a dynamic process state channel; Inputting the multi-channel graph data into a first graph convolution transformation layer among the Y sequentially connected graph convolution transformation layers, wherein the first graph convolution transformation layer includes a graph convolution kernel, a node feature aggregation module, and a nonlinear activation module; Extracting process topological features from the adjacency matrix and the feature matrix through the graph convolution kernel to generate an initial convolution feature graph, wherein the weight matrix of the graph convolution kernel dynamically adjusts the aggregation weight based on the fluid flow path between the production control units; Based on the node connection relationship in the initial convolution feature graph, the node feature aggregation module is used to perform cross-node aggregation on the process parameter features of adjacent production control units to generate an aggregated feature matrix, wherein the cross-node aggregation includes gradient weighted fusion of temperature control parameters and pressure feedback parameters, and dynamic correlation superposition of fluid flow rate monitoring parameters and gel consistency detection parameters; Inputting the aggregated feature matrix into the nonlinear activation module for feature nonlinear mapping to generate a conversion result of the first graph convolutional transformation layer, wherein the output dimension of the nonlinear activation module is consistent with the column dimension of the feature matrix, and the activation function parameters are dynamically calibrated based on the steady-state threshold of the aqueous gel production process; The conversion result is used as the input of the next graph convolutional transformation layer to generate a graph expression vector of the first benchmark system running sample through cascade processing of Y graph convolutional transformation layers.

5. The abnormality analysis method applied to the aqueous gel production control system according to claim 2, characterized in that: The first sample system operation data includes a first reference system operation sample, a first associated system operation sample and a first non-associated system operation sample, the i-1th prior anomaly diagnosis neural network is a dual graph convolution module, and the i-th prior anomaly diagnosis neural network is an interactive graph convolution module; The step of performing graph convolution conversion on the first sample system operation data using the i-th priori abnormality diagnosis neural network to generate a graph convolution conversion result of the i-th priori abnormality diagnosis neural network includes: Aggregating the first benchmark system operation sample and the first associated system operation sample to generate a first interactive system operation sample, and aggregating the first benchmark system operation sample and the first non-associated system operation sample to generate a second interactive system operation sample; Performing normalization processing on the first interactive system operation sample and the second interactive system operation sample respectively to generate network learning data of the first interactive system operation sample and network learning data of the second interactive system operation sample; Using the i-th prior abnormality diagnosis neural network, respectively, the network learning data of the first interactive system operation sample and the network learning data of the second interactive system operation sample are converted into graph convolutions to generate a graph expression vector of the first interactive system operation sample and a graph expression vector of the second interactive system operation sample; The graph expression vector of the first interactive system running sample and the graph expression vector of the second interactive system running sample are used as the graph convolution conversion result of the i-th prior abnormality diagnosis neural network.

6. The abnormality analysis method applied to the aqueous gel production control system according to claim 2, characterized in that: The determining of the global training error based on the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph convolution conversion result of the i-th prior abnormality diagnosis neural network comprises: Determining a first error term according to a graph convolution transformation result of the basic abnormality diagnosis neural network after the i-1th round of guided learning; Determining a second error term according to the graph convolution conversion result of the i-th prior abnormality diagnosis neural network and the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1-th round of guided learning; fusing the first error term and the second error term to generate a global training error; The graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning includes a graph expression vector of a first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, a graph expression vector of a first associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning, and a graph expression vector of a first non-associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning; The determining of the first error term according to the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning includes: Performing feature distance calculation on the graph expression vector of the first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph expression vector of the first associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning to generate a first feature distance; Performing feature distance calculation on the graph expression vector of the first benchmark system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning and the graph expression vector of the first non-associated system operation sample of the basic abnormality diagnosis neural network after the i-1th round of guided learning to generate a second feature distance; Determining a first error term according to the first characteristic distance and the second characteristic distance; Wherein, determining the second error term according to the graph convolution conversion result of the i-th prior abnormality diagnosis neural network and the graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1-th round of guided learning includes: Determining a first abnormality diagnosis result of the i-th prior abnormality diagnosis neural network according to a graph convolution conversion result of the i-th prior abnormality diagnosis neural network; Determining a second abnormality diagnosis result of the basic abnormality diagnosis neural network after the i-1th round of guided learning according to a graph convolution conversion result of the basic abnormality diagnosis neural network after the i-1th round of guided learning; A second error term is determined according to the first abnormality diagnosis result and the second abnormality diagnosis result.

7. The abnormality analysis method applied to the aqueous gel production control system according to claim 6, characterized in that: The first sample system operation data includes a first reference system operation sample, a first associated system operation sample and a first non-associated system operation sample; Determining the first abnormality diagnosis result of the i-th priori abnormality diagnosis neural network according to the graph convolution conversion result of the i-th priori abnormality diagnosis neural network includes: Determine, according to the graph convolution conversion result of the i-th prior abnormality diagnosis neural network, a third characteristic distance between the first reference system operation sample and the first associated system operation sample, and determine a fourth characteristic distance between the first reference system operation sample and the first non-associated system operation sample; Performing standardization conversion on the third characteristic distance and the fourth characteristic distance to generate a fifth characteristic distance; Outputting the ratio of the third characteristic distance to the fifth characteristic distance as a first abnormality diagnosis confidence, and outputting the ratio of the fourth characteristic distance to the fifth characteristic distance as a second abnormality diagnosis confidence; The first abnormality diagnosis confidence and the second abnormality diagnosis confidence are output as a first abnormality diagnosis result of the i-th prior abnormality diagnosis neural network.

8. The abnormality analysis method applied to the aqueous gel production control system according to claim 1, characterized in that: Before obtaining the initialized basic abnormality diagnosis neural network and X prior abnormality diagnosis neural networks, the method further includes: Acquire second sample system operation data and an i-th initial priori abnormality diagnosis neural network, wherein the second sample system operation data includes a second reference system operation sample, a second associated system operation sample, and a second non-associated system operation sample; Normalizing the second benchmark system operation sample, the second associated system operation sample, and the second non-associated system operation sample respectively to generate network learning data of the second benchmark system operation sample, network learning data of the second associated system operation sample, and network learning data of the second non-associated system operation sample; Using the i-th initial priori abnormal diagnosis neural network, respectively, the network learning data of the second reference system operation sample, the network learning data of the second associated system operation sample, and the network learning data of the second non-associated system operation sample are converted into graph convolutions to generate a graph expression vector of the second reference system operation sample, a graph expression vector of the second associated system operation sample, and a graph expression vector of the second non-associated system operation sample; The i-th initial prior anomaly diagnosis neural network is trained based on the graph expression vector of the second benchmark system operation sample, the graph expression vector of the second associated system operation sample and the graph expression vector of the second non-associated system operation sample to generate the i-th prior anomaly diagnosis neural network.

9. The abnormality analysis method applied to the aqueous gel production control system according to claim 1, characterized in that: Before obtaining the initialized basic abnormality diagnosis neural network and X prior abnormality diagnosis neural networks, the method further includes: Acquire second sample system operation data and an i-th initial priori abnormality diagnosis neural network, wherein the second sample system operation data includes a second reference system operation sample, a second associated system operation sample, and a second non-associated system operation sample; Aggregating the second benchmark system operation sample and the second associated system operation sample to generate a third interactive system operation sample, and aggregating the second benchmark system operation sample and the second non-associated system operation sample to generate a fourth interactive system operation sample; Performing normalization processing on the third interactive system operation sample and the fourth interactive system operation sample respectively, generating network learning data of the third interactive system operation sample and network learning data of the fourth interactive system operation sample; Using the i-th initial priori abnormal diagnosis neural network, respectively, the network learning data of the third interactive system operation sample and the network learning data of the fourth interactive system operation sample are subjected to graph convolution conversion to generate a graph expression vector of the third interactive system operation sample and a graph expression vector of the fourth interactive system operation sample; The i-th initial a priori abnormality diagnosis neural network is trained based on the graph expression vector of the third interactive system running sample and the graph expression vector of the fourth interactive system running sample to generate the i-th a priori abnormality diagnosis neural network.

10. An intelligent production service system, characterized in that: The intelligent production service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the abnormality analysis method applied to the aqueous gel production control system according to any one of claims 1 to 9.

Images