System and method for lithium sulfide quality anomaly diagnosis and process adjustment

Abstract

The application discloses a lithium sulfide quality anomaly diagnosis and process adjustment system and method, and belongs to the technical field of intelligent management and control of solid-state battery key material preparation. The system comprises a multi-source data acquisition module, a visual quality detection module, a process and quality state reasoning module, a quality anomaly diagnosis module, an equipment fault prediction module, a multi-target optimization decision module, a closed-loop control execution module and a safety linkage module. The system obtains lithium sulfide production process data and material visual images, performs quality trend prediction, quality transfer tracking, anomaly root cause reasoning and equipment fault prediction, generates process adjustment strategies and set value adjustment instructions, and issues the instructions to the bottom control system for execution; when the hydrogen sulfide leakage concentration reaches a preset threshold, emergency shutdown and nitrogen purging are triggered. The application can be used for quality management, process optimization and safety linkage control of the lithium sulfide production process.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for the preparation of key materials for solid-state batteries, and in particular to a system and method for diagnosing and adjusting the quality anomalies of lithium sulfide. Background Technology

[0002] All-solid-state batteries have become an important direction for next-generation battery technology due to their high energy density and safety advantages. Sulfide solid electrolytes possess high ionic conductivity and good mechanical ductility. Lithium sulfide, as a key raw material for sulfide electrolytes, has its synthesis process directly affecting the cost and electrochemical performance of the material. Currently, the mainstream process for preparing high-purity lithium sulfide involves a two-step reaction of hydrogen sulfide and lithium hydroxide in an N-methylpyrrolidone solvent, involving neutralization and thermal decomposition. For example, Chinese invention patent application CN122010059A, published on May 12, 2026, entitled "Preparation method of lithium sulfide synthesized from industrial-grade lithium hydroxide and sodium sulfide, preparation method of solid electrolyte, and solid-state battery," discloses the use of industrial-grade lithium hydroxide and sodium sulfide as raw materials. Impurities such as iron, calcium, and magnesium are removed through multi-stage chemical precipitation by adding hydrogen peroxide, sodium sulfide, sodium carbonate, and barium nitrate. Subsequently, the hydrogen sulfide gas generated in the reaction is passed into a purification liquid for concentration and crystallization purification. Chinese invention patent application with publication number CN122010058A, publication date May 12, 2026, and titled "Preparation method of high-purity lithium sulfide by liquid-phase purification and impurity removal, preparation method of sulfide solid electrolyte material and power battery", discloses a method for physical-level deep liquid-phase impurity removal of the initial solution through a stepped resin column assembly and nanofiltration membrane, and a reaction synthesis of high-purity hydrogen sulfide gas after low-temperature dehydration and catalytic deoxygenation to improve the purity of the product.

[0003] However, the aforementioned existing technologies primarily focus on controlling impurities through pretreatment methods such as chemical precipitation or physical membrane separation, failing to address the process control challenges in the actual dynamic production of lithium sulfide. This preparation process is characterized by strong coupling of multiple parameters, with reaction temperature, pressure, hydrogen sulfide flow rate, and stirring rate all intricately affecting each other, and the entire process faces extremely high safety risks from highly toxic and flammable gases. Existing technologies still heavily rely on manual experience in production control, lacking data-driven dynamic adaptive optimization methods. This results in quality inspection methods typically relying on offline analysis such as microscopic sampling, which significantly lags behind the pace of production control. Furthermore, the maintenance of critical production equipment often employs a passive, periodic inspection model, making unexpected downtime highly likely. Moreover, data silos exist between various industrial information systems, such as distributed control systems and manufacturing execution systems, hindering rapid diagnosis and safety interlocking through deep inter-system collaboration in the event of abnormal fluctuations or hazardous gas leaks.

[0004] Therefore, how to solve the problems of lagging manual control, offline quality inspection, passive equipment maintenance, and high safety risks caused by strong coupling of multiple parameters in the lithium sulfide production process, and how to achieve intelligent dynamic quality control and process collaborative optimization throughout the entire process from raw materials to products, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The main objective of this invention is to provide a lithium sulfide quality anomaly diagnosis and process adjustment system and method, which aims to solve the problems of lagging manual control, offline quality inspection, passive equipment maintenance and high safety risks caused by strong coupling of multiple parameters in the lithium sulfide production process, and realize intelligent dynamic quality control and process collaborative optimization throughout the entire process from raw material end to product end.

[0006] To achieve the above objectives, this invention proposes a lithium sulfide quality anomaly diagnosis and process adjustment system, applicable to lithium sulfide production processes using hydrogen sulfide, lithium hydroxide, and N-methylpyrrolidone as raw materials or solvents, comprising: The multi-source data acquisition module is used to communicate with the distributed control system, safety instrumented system, manufacturing execution system, gas detection alarm system and laboratory information management system respectively, and to acquire real-time production data in the lithium sulfide production process. The real-time production data includes process parameters, safety status parameters, production execution data, raw material quality data, equipment operating status parameters and offline quality detection data. The visual quality detection module is used to acquire visual images of the lithium sulfide material produced by the reaction in an inert gas protective environment, and input the visual images into a pre-trained image segmentation and classification model to output real-time visual quality detection results, including crystal form distribution, particle size distribution and surface defect information. The process and quality status reasoning module is used to input the real-time production data and the real-time visual quality detection results into the pre-established process and quality mapping model to predict the product quality trend, and to track the quality transfer of lithium sulfide from raw materials to products based on the raw material-reaction-product knowledge graph containing graph neural networks, and output the quality transfer tracking results. The quality anomaly diagnosis module is used to perform root cause reasoning based on the quality transfer tracking results when the product quality trend prediction result or the real-time visual quality detection result deviates from the preset quality standard, and to determine the key process nodes and causes of the quality anomaly. The equipment failure prediction module is used to input the equipment operating status parameters into a pre-trained equipment failure prediction model and output the probability of failure of the lithium sulfide reactor within a future preset time period. The multi-objective optimization decision module is used to generate a process adjustment strategy for the current production state based on the key process nodes, the causes of the anomalies, the probability of the failure, and the preset safe operation boundary, and to generate setpoint adjustment instructions based on the process adjustment strategy. The setpoint adjustment instructions include adjustment instructions for one or more setpoints among temperature, pressure, hydrogen sulfide flow rate, and stirring rate. The closed-loop control execution module is used to send the set value adjustment command to the underlying control system through the industrial control protocol, and to receive the execution status data fed back by the underlying control system in order to dynamically adjust the lithium sulfide reactor. The safety linkage module is used to communicate and link with the gas detection alarm system and the safety instrument system. When the gas detection alarm system detects that the hydrogen sulfide leakage concentration reaches or exceeds a preset safety threshold, it triggers the safety instrument system to perform an emergency cut-off interlock action and a nitrogen purging action.

[0007] Preferably, the process parameters include reaction temperature, reaction pressure, stirring rate, hydrogen sulfide flow rate, liquid level, and N-methylpyrrolidone water content; the safety status parameters include hydrogen sulfide concentration, combustible gas concentration, and equipment sealing pressure; the production execution data includes batch information, raw material batch number, operator information, and equipment runtime; the raw material quality data includes the chemical purity of lithium hydroxide, magnetic foreign matter content, particle size distribution, and trace element spectrum; the offline quality detection data includes X-ray diffraction crystal form analysis data, laser particle size analyzer particle size distribution data, inductively coupled plasma atomic emission spectrometry (ICP-AES) impurity content data, specific surface area data, and scanning electron microscope morphology data.

[0008] Preferably, the visual quality detection module includes a closed inert gas protection chamber located at the outlet or post-processing outlet of the lithium sulfide reaction device, an industrial camera located within the closed inert gas protection chamber, and a ring light source; the industrial camera has 20 megapixels, and the ring light source has a color temperature of [missing information]. The image segmentation and classification model is used to sequentially perform Gaussian filtering for noise reduction, adaptive contrast enhancement, lithium sulfide particle instance segmentation, feature extraction, and classification output on the visual image.

[0009] Preferably, the image segmentation and classification model includes a particle segmentation model based on the U-Net network architecture and a classification model with ResNet-50 as the backbone network; the particle segmentation model is used to segment lithium sulfide particles in the visual image; the classification model is used to extract features and classify crystal forms of the segmented particle images to identify lithium sulfide. Mutually, Harmony The phase is processed and outputs surface defect information including agglomerates, foreign matter, and color difference.

[0010] Preferably, the process-quality mapping model includes at least two of the following: a random forest model, a gradient boosting tree model, and a deep neural network model; the input variables of the deep neural network model include temperature, pressure, flow rate, time, ratio, solid content, and stirring rate, and the output variables include product purity, particle size, impurity content, and moisture content; the random forest model includes 500 decision trees, a maximum depth of 15, and a minimum number of leaf node samples of 5, and is used to identify key process parameters through feature importance analysis; the gradient boosting tree model has a learning rate of 0.05, a tree depth of 6, and a subsampling rate of 0.8, and is used for continuous process parameter optimization.

[0011] Preferably, the raw material-reaction-product knowledge graph includes raw material nodes, equipment nodes, process nodes, and quality indicator nodes; the raw material nodes include lithium hydroxide, hydrogen sulfide, and N-methylpyrrolidone; the equipment nodes include reaction vessels and dryers; the process nodes include reaction stages and drying stages; and the quality indicator nodes include purity, particle size, moisture content, and impurities; the relationship types of the raw material-reaction-product knowledge graph include chemical transformation relationships, equipment association relationships, and quality transfer relationships, wherein the chemical transformation relationships include... as well as The graph neural network adopts the GraphSAGE architecture, with mean aggregation as the aggregation function and an embedding dimension of 128. It calculates the correlation probability between raw material batches, reaction conditions, and final product quality through node classification and link prediction.

[0012] Preferably, the equipment failure prediction model includes an isolated forest algorithm model and a long short-term memory network model; the isolated forest algorithm model has 100 trees, a subsampling rate of 0.8, and an anomaly score threshold of 0.6, and is used to detect anomalies in the equipment operating status parameters; the long short-term memory network model is used to construct time series data of length 72 according to a 5-minute sampling interval, and outputs the failure probability of the lithium sulfide reactor in the next 72 hours based on the time series data.

[0013] Preferably, the quality anomaly diagnosis module includes a Bayesian network model, which is used to calculate the posterior probability of the deviation cause leading to the quality anomaly based on the conditional probability relationship between raw material impurity content, raw material moisture content, temperature stability, stirring efficiency, product purity, and particle size distribution. The multi-objective optimization decision module uses a Q-Learning-based reinforcement learning algorithm combined with a non-dominated sorting genetic algorithm II to generate the process adjustment strategy. The optimization objectives include product purity, particle size distribution, energy consumption, and production efficiency, and the adjustment range of the temperature setpoint generated in a single iteration is constrained to within a certain range. Within the range, the pressure setpoint adjustment range is constrained. Within the specified range, the adjustment range of the hydrogen sulfide flow rate setpoint is constrained. Within the range.

[0014] Preferably, the closed-loop control execution module sends the setpoint adjustment command to the distributed control system via the OPC UA protocol; the safety linkage module is activated when the gas detection alarm system detects that the hydrogen sulfide leakage concentration reaches or exceeds [a certain value]. When the emergency shut-off valve is closed and the nitrogen purging action is triggered by the safety instrumented system, the closed-loop control execution module is also used to link with the manufacturing execution system and the laboratory information management system to automatically generate electronic batch records and associate process curves, quality inspection results and abnormal diagnosis results with the corresponding batches.

[0015] This application also discloses an AI algorithm-based method for diagnosing quality anomalies and adjusting processes in lithium sulfide production, applicable to lithium sulfide production processes using hydrogen sulfide, lithium hydroxide, and N-methylpyrrolidone as raw materials or solvent systems, including: Acquire real-time production data during the lithium sulfide production process. The real-time production data includes process parameters, safety status parameters, production execution data, raw material quality data, equipment operating status parameters, and offline quality testing data. Visual images of the lithium sulfide material produced by the reaction are acquired in an inert gas protective environment, and the visual images are input into a pre-trained image segmentation and classification model to output real-time visual quality detection results, which include crystal form distribution, particle size distribution and surface defect information. The real-time production data and the real-time visual quality inspection results are input into a pre-established process and quality mapping model to predict product quality trends. Based on a raw material-reaction-product knowledge graph containing a graph neural network, the quality transfer of lithium sulfide from raw material to product is tracked, and the quality transfer tracking results are output. When the product quality trend prediction result or the real-time visual quality detection result deviates from the preset quality standard, root cause reasoning is performed based on the quality transfer tracking result to determine the key process nodes and causes of the quality abnormality. The equipment operating status parameters are input into a pre-trained equipment fault prediction model, which outputs the probability of failure of the lithium sulfide reactor within a future preset time period. Based on the key process nodes, the causes of the anomalies, the probability of failure, and the preset safe operation boundaries, a process adjustment strategy for the current production state is generated, and a set value adjustment instruction is generated based on the process adjustment strategy. The setpoint adjustment command is sent to the underlying control system through the industrial control protocol, and the execution status data fed back by the underlying control system is received to dynamically adjust the lithium sulfide reactor. When the gas detection alarm system detects that the concentration of hydrogen sulfide leakage reaches or exceeds the preset safety threshold, it triggers the safety instrumented system to perform emergency shut-off interlocking and nitrogen purging actions.

[0016] The above technical solution has the following advantages: This invention provides a lithium sulfide quality anomaly diagnosis and process adjustment system and method. It centrally integrates data from multiple platforms, including distributed control systems and safety instrumented systems, through a multi-source data acquisition module, breaking down information silos between industrial systems and providing multi-dimensional, heterogeneous, and complete data support for artificial intelligence decision-making. A visual quality detection module, deployed within an inert gas protection chamber, extracts microscopic quality features such as particle size and crystal defects in real time, improving the adjustment lag problem caused by traditional offline laboratory analysis. A process and quality status reasoning module, in conjunction with a quality anomaly diagnosis module, inputs real-time production data and visual features into an artificial intelligence atlas for dynamic quality trend tracking and root cause inference, improving positioning efficiency and accuracy to pinpoint key process nodes and potential causes of quality index deviations. An equipment failure prediction model, through continuous time series analysis, enables predictive maintenance of core reaction equipment, reducing the probability of unplanned downtime due to sudden mechanical failures. A multi-objective optimization decision-making module dynamically generates setpoint adjustment commands within strict safety boundaries based on the current comprehensive production status and sends them to the underlying control system through a closed-loop control execution module, achieving high-precision automatic optimization of complex nonlinear chemical reactions. The safety linkage module establishes a high-priority cross-system hardware linkage response for excessive concentrations of harmful gases, eliminating the reliance on manual emergency assessment in high-risk production scenarios and reducing the risk of safety accidents caused by leaks of toxic and flammable gases. Attached Figure Description

[0017] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein: Figure 1This is a schematic diagram of the overall architecture of the system provided in the embodiments of this application.

[0018] Figure 2 This is a flowchart illustrating the method for diagnosing and adjusting abnormal lithium sulfide production quality according to an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0020] Example 1 like Figure 1 As shown, this embodiment provides a lithium sulfide quality anomaly diagnosis and process adjustment system. This system is applied to the lithium sulfide production process using hydrogen sulfide, lithium hydroxide, and N-methylpyrrolidone as raw materials or solvents. The lithium sulfide production process includes two main stages: neutralization reaction and thermal decomposition reaction. The specific chemical equation for the neutralization reaction is as follows: The specific chemical equation for the thermal decomposition reaction is: Traditional lithium sulfide preparation processes typically rely on manual experience and suffer from outdated quality control methods, making them ill-suited for complex reaction environments with highly coupled multi-parameter parameters. They also pose significant safety risks due to the presence of highly toxic and flammable gases. This embodiment presents a system with an integrated four-layer architecture: perception, cognition, decision-making, and execution. This architecture aims to achieve intelligent quality control and process optimization throughout the entire process. Specifically, the system includes a multi-source data acquisition module, a visual quality inspection module, a process and quality status reasoning module, a quality anomaly diagnosis module, an equipment fault prediction module, a multi-objective optimization decision-making module, a closed-loop control execution module, and a safety linkage module.

[0021] The multi-source data acquisition module, serving as the core interface of the perception layer, communicates with the distributed control system, safety instrumented system, manufacturing execution system, gas detection and alarm system, and laboratory information management system. Through these industrial information systems, the multi-source data acquisition module acquires real-time production data during the lithium sulfide production process. This real-time production data specifically includes process parameters, safety status parameters, production execution data, raw material quality data, equipment operating status parameters, and offline quality testing data. The centralized acquisition of multi-source heterogeneous data breaks down information silos, providing a complete data foundation for subsequent artificial intelligence cognition and decision-making.

[0022] A visual quality inspection module is deployed at the lithium sulfide production stage. Under a closed, inert gas environment, the module acquires visual images of the produced lithium sulfide material. These images are then processed by a pre-trained image segmentation and classification model. The model outputs real-time visual quality inspection results, including crystal form distribution, particle size distribution, and surface defect information. This online visual inspection mechanism enables the system to acquire microscopic quality characteristics of the product in real time during production, effectively avoiding the control lag issues associated with traditional offline laboratory analysis.

[0023] The process and quality status reasoning module belongs to the cognitive layer. It receives real-time production data from the multi-source data acquisition module and real-time visual quality inspection results from the visual quality inspection module. This data is then input into a pre-established process and quality mapping model for calculation, resulting in a product quality trend prediction. Simultaneously, based on a raw material-reaction-product knowledge graph incorporating a graph neural network, the module tracks the entire quality transfer process of lithium sulfide from raw material to final product and outputs the corresponding quality transfer tracking results.

[0024] The quality anomaly diagnosis module communicates with the process and quality status reasoning module. When the product quality trend prediction or real-time visual quality inspection results deviate from the preset quality standards, the quality anomaly diagnosis module is triggered. Based on the aforementioned quality transfer tracing results, the quality anomaly diagnosis module performs root cause reasoning. Through root cause reasoning, the module accurately identifies the key process nodes and specific causes of the quality anomalies. This mechanism can quickly locate hidden problems such as unstable temperature control or excessive impurities in raw materials.

[0025] The equipment failure prediction module receives equipment operating status parameters from the multi-source data acquisition module. These parameters include, for example, vibration and temperature data from the high-pressure reactor's agitator motor. The module then inputs these parameters into a pre-trained equipment failure prediction model for analysis. The model outputs the probability of failure occurring in the lithium sulfide reactor within a preset future timeframe. Through this predictive maintenance mechanism, the system can identify potential mechanical or electrical failures in critical equipment in advance, significantly reducing unplanned downtime.

[0026] The multi-objective optimization decision-making module constitutes the system's decision-making layer. This module integrates the key process nodes and causes of anomalies determined by the quality anomaly diagnosis module, as well as the failure probability output by the equipment failure prediction module. Based on this information and preset safe operating boundaries, the multi-objective optimization decision-making module generates the optimal process adjustment strategy for the current production state. Subsequently, based on this process adjustment strategy, the multi-objective optimization decision-making module generates specific setpoint adjustment instructions. These setpoint adjustment instructions specifically include setpoint adjustment instructions for one or more control loops among reaction temperature, reaction pressure, hydrogen sulfide flow rate, and stirring rate.

[0027] The closed-loop control execution module, acting as an execution layer component, communicates with the underlying control system. It sends setpoint adjustment commands to the underlying control system via standard industrial control protocols for automatic adjustment. Simultaneously, the module receives execution status data from the underlying control system, forming a closed-loop control circuit and dynamically adjusting the operating status of the lithium sulfide reactor. This closed-loop mechanism achieves seamless integration from intelligent decision-making to physical equipment actions.

[0028] The safety linkage module establishes high-priority communication linkage mechanisms with both the gas detection alarm system and the safety instrumented system. When the gas detection alarm system detects a hydrogen sulfide leak at the production site that the concentration reaches or exceeds a preset safety threshold, the safety linkage module immediately triggers the safety instrumented system to perform an emergency shut-off interlock action and simultaneously initiates a nitrogen purging operation. This safety linkage mechanism eliminates the system's reliance on manual emergency response, effectively preventing the escalation of toxic or flammable gas leaks and improving the operational safety of high-risk production processes.

[0029] Based on the above-described Example 1, this example provides a detailed description of the various data parameters collected by the system. Specifically, process parameters include reaction temperature, reaction pressure, stirring rate, hydrogen sulfide flow rate, liquid level, and N-methylpyrrolidone water content. Safety status parameters include hydrogen sulfide concentration, combustible gas concentration, and equipment sealing pressure. Production execution data includes batch information, raw material batch number, operator information, and equipment runtime. Raw material quality data includes the chemical purity of lithium hydroxide, magnetic foreign matter content, particle size distribution, and trace element spectrum. Offline quality testing data includes X-ray diffraction crystal form analysis data, laser particle size analyzer particle size distribution data, inductively coupled plasma atomic emission spectrometry (ICP-AES) impurity content data, specific surface area data, and scanning electron microscopy morphology data. The acquisition of these multi-dimensional data ensures comprehensive coverage and dynamic monitoring of the production process by the sensing layer.

[0030] like Figure 2As shown, this embodiment further defines the hardware deployment and algorithm flow for the visual quality detection module. The visual quality detection module includes a sealed inert gas protection chamber located at the outlet of the lithium sulfide reaction device or the post-processing outlet, an industrial camera located within the sealed inert gas protection chamber, and a ring light source. The industrial camera has a resolution of 20 megapixels. The ring light source has a color temperature of [missing information]. The aforementioned hardware configuration ensures high-resolution image acquisition and consistent lighting, while the inert gas protective chamber effectively prevents oxidation or moisture absorption failure of lithium sulfide products during the detection process. The image segmentation and classification model running within the visual quality detection module sequentially performs Gaussian filtering for noise reduction, adaptive contrast enhancement, lithium sulfide particle instance segmentation, feature extraction, and classification output on the acquired visual images.

[0031] The image segmentation and classification model employs a cascaded deep learning network structure. This model includes a particle segmentation model based on the U-Net network architecture and a classification model with ResNet-50 as the backbone network. The particle segmentation model performs high-precision instance segmentation of lithium sulfide particles in visual images. The classification model extracts features and classifies crystal forms from the segmented particle images. The classification model can identify the α, β, and γ phases of lithium sulfide. It also outputs surface defect information including particle agglomeration, foreign matter, and particle color differences. Compared to traditional manual microscopic inspection, this machine vision-based detection scheme significantly improves the detection rate of crystal defects and greatly reduces the inspection time per batch.

[0032] To delve deeper into the complex nonlinear relationship between process parameters and product quality, this embodiment elaborates on the internal design details of the process-quality mapping model. The process-quality mapping model includes at least two of the following: random forest model, gradient boosting tree model, and deep neural network model. The input variables of the deep neural network model include temperature, pressure, flow rate, time, proportion, solid content, and stirring rate. The output variables of the deep neural network model include product purity, particle size, impurity content, and moisture content. Through the deep feature extraction and nonlinear fitting capabilities of the neural network, the system can accurately predict the overall product quality trend under complex operating conditions.

[0033] As a supplement to deep neural networks and to interpret feature importance, the system uses a random forest model containing 500 decision trees. The maximum depth of this random forest model is set to 15, and the minimum number of leaf node samples is set to 5. Through feature importance analysis algorithms, the random forest model accurately identifies the most critical process parameters affecting the current quality indicators. The gradient boosting tree model has a learning rate of 0.05, a tree depth of 6, and a subsampling rate of 0.8. The gradient boosting tree model is specifically designed for high-precision optimization of continuous process parameters and time series trend prediction. Through the integrated application of multiple algorithms, the prediction accuracy and model generalization ability of the cognitive layer are significantly enhanced.

[0034] In terms of quality transfer and traceability, the raw material-reaction-product knowledge graph includes raw material nodes, equipment nodes, process nodes, and quality indicator nodes. The raw material node specifically includes lithium hydroxide, hydrogen sulfide, and N-methylpyrrolidone. The equipment node includes the reactor and dryer. The process node includes the reaction stage and the drying stage. The quality indicator nodes include purity, particle size, moisture content, and impurities. The relationship types in this knowledge graph include chemical transformation relationships, equipment association relationships, and quality transfer relationships. Among them, the chemical transformation relationships cover the two core reactions in lithium sulfide preparation, and the specific chemical equation for the neutralization reaction is as follows: The specific chemical equation for the thermal decomposition reaction is: The system's graph neural network specifically adopts the GraphSAGE architecture. The aggregation function of this graph neural network is mean aggregation, and the embedding dimension is set to 128. Through node classification and link prediction algorithms, the graph neural network quantitatively calculates the correlation probability between raw material batches, reaction conditions, and final product quality, thereby realizing knowledge reasoning and traceability management throughout the entire manufacturing lifecycle.

[0035] To address the issues of reactive maintenance and delayed fault detection in chemical equipment maintenance, this embodiment details the internal computational architecture of an equipment fault prediction model. The model comprises an isolated forest algorithm and a long short-term memory (LSTM) network model. The isolated forest algorithm has 100 trees, a subsampling rate of 0.8, and an anomaly score threshold of 0.6. It is used for real-time unsupervised anomaly detection of equipment operating parameters. The LTM network model extracts equipment operating data at 5-minute sampling intervals and constructs a 72-dimensional time series dataset. Based on this multi-dimensional time series data, the LTM network model outputs the probability of failure of the lithium sulfide reactor within the next 72 hours. This predictive maintenance strategy effectively reduces the probability of sudden failures and unplanned downtime risks of critical core equipment.

[0036] This embodiment further refines the execution logic of quality anomaly diagnosis and multi-objective optimization. The core algorithm of the quality anomaly diagnosis module includes a Bayesian network model. Based on the conditional probability relationship between raw material impurity content, raw material moisture content, temperature stability, stirring efficiency, product purity, and particle size distribution, the Bayesian network model calculates the posterior probability of the causes of various quality anomalies. The multi-objective optimization decision module uses a Q-Learning-based reinforcement learning algorithm combined with a non-dominated sorting genetic algorithm II to generate process adjustment strategies. The comprehensive optimization objectives cover product purity, particle size distribution, energy consumption, and production efficiency. To improve the safety and control stability of the production process, the system implements strict physical boundary control when issuing automatic adjustment commands. The adjustment range of a single generated temperature setpoint is constrained to... Within the range. The pressure setpoint adjustment range is constrained within... Within the specified range. The adjustment range of the hydrogen sulfide flow rate setpoint is constrained to within... Within the range.

[0037] At the level of coordinated execution of the underlying control equipment, the closed-loop control execution module uses the standardized OPC UA protocol to send setpoint adjustment commands to the distributed control system. This protocol configuration ensures the real-time performance and reliability of data command interaction between heterogeneous systems. When the on-site gas detection alarm system detects that the hydrogen sulfide leakage concentration reaches or exceeds [a certain level], [the system will respond accordingly]. When the safety red line is crossed, the safety linkage module bypasses the regular business logic and quickly triggers the safety instrumented system to physically close the emergency shut-off valve, simultaneously initiating a nitrogen purging operation on the pipeline. Furthermore, the closed-loop control execution module is deeply integrated with the manufacturing execution system and the laboratory information management system to automate the generation of electronic batch records. The system automatically and strongly correlates process control curves, quality visual inspection results, and AI anomaly diagnostic results with the corresponding production batch numbers, thus providing a complete audit trail data chain that meets modern industrial compliance requirements.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A lithium sulfide quality anomaly diagnosis and process adjustment system, applied to a lithium sulfide production process using hydrogen sulfide, lithium hydroxide, and N-methylpyrrolidone as raw materials or solvents, characterized in that, include: The multi-source data acquisition module communicates with the distributed control system, safety instrumented system, manufacturing execution system, gas detection and alarm system, and laboratory information management system to acquire real-time production data, including process parameters, safety status parameters, production execution data, raw material quality data, equipment operating status parameters, and offline quality testing data. The visual quality inspection module is used to acquire visual images of the lithium sulfide material produced by the reaction in an inert gas protective environment, and output real-time visual quality inspection results including crystal distribution, particle size distribution and surface defect information through a pre-trained image segmentation and classification model. The process and quality status reasoning module is used to predict product quality trends based on the real-time production data and the real-time visual quality inspection results, and to track the quality transfer of lithium sulfide from raw materials to products based on a raw material-reaction-product knowledge graph containing a graph neural network, and output the quality transfer tracking results. The quality anomaly diagnosis module is used to perform root cause reasoning based on the quality transfer tracking results when the product quality trend prediction result or the real-time visual quality detection result deviates from the preset quality standard, and to determine the key process nodes and causes of the quality anomaly. The equipment failure prediction module is used to output the probability of failure of the lithium sulfide reactor within a future preset time period based on the equipment operating status parameters. The multi-objective optimization decision module is used to generate process adjustment strategies based on the key process nodes, abnormal causes, failure probability and preset safe operation boundaries, and generate setpoint adjustment instructions for one or more setpoints among temperature, pressure, hydrogen sulfide flow rate and stirring rate. The closed-loop control execution module is used to send the set value adjustment command to the underlying control system through the industrial control protocol, and to receive execution status data to dynamically adjust the lithium sulfide reaction device. The safety linkage module communicates and links with the gas detection alarm system and the safety instrument system to trigger emergency shut-off interlocking and nitrogen purging actions when the hydrogen sulfide leakage concentration reaches or exceeds a preset safety threshold.

2. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The process parameters include reaction temperature, reaction pressure, stirring rate, hydrogen sulfide flow rate, liquid level, and N-methylpyrrolidone water content; the safety parameters include hydrogen sulfide concentration, combustible gas concentration, and equipment sealing pressure; the production execution data includes batch information, raw material batch number, operator information, and equipment runtime; the raw material quality data includes the chemical purity of lithium hydroxide, magnetic foreign matter content, particle size distribution, and trace element spectrum; the offline quality testing data includes X-ray diffraction crystal form analysis data, laser particle size analyzer particle size distribution data, inductively coupled plasma atomic emission spectrometry (ICP-AES) impurity content data, specific surface area data, and scanning electron microscope morphology data.

3. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The visual quality inspection module includes a sealed inert gas protection chamber located at the outlet or post-processing outlet of the lithium sulfide reactor, an industrial camera located within the sealed inert gas protection chamber, and a ring light source; the industrial camera has 20 megapixels, and the ring light source has a color temperature of [missing information]. The image segmentation and classification model is used to sequentially perform Gaussian filtering for noise reduction, adaptive contrast enhancement, lithium sulfide particle instance segmentation, feature extraction, and classification output on the visual image.

4. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 3, characterized in that, The image segmentation and classification model includes a particle segmentation model based on the U-Net network architecture and a classification model with ResNet-50 as the backbone network. The particle segmentation model is used to segment lithium sulfide particles in the visual image. The classification model is used to extract features and classify crystal forms of the segmented particle images to identify lithium sulfide. Mutually, Harmony The phase is processed and outputs surface defect information including agglomerates, foreign matter, and color difference.

5. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The process-quality mapping model includes at least two of the following: random forest model, gradient boosting tree model, and deep neural network model. The input variables of the deep neural network model include temperature, pressure, flow rate, time, ratio, solid content, and stirring rate, and the output variables include product purity, particle size, impurity content, and moisture content. The random forest model includes 500 decision trees, a maximum depth of 15, and a minimum number of leaf node samples of 5, and is used to identify key process parameters through feature importance analysis. The gradient boosting tree model has a learning rate of 0.05, a tree depth of 6, and a subsampling rate of 0.8, and is used for continuous process parameter optimization.

6. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The raw material-reaction-product knowledge graph includes raw material nodes, equipment nodes, process nodes, and quality indicator nodes. The raw material nodes include lithium hydroxide, hydrogen sulfide, and N-methylpyrrolidone; the equipment nodes include reaction vessels and dryers; the process nodes include reaction stages and drying stages; and the quality indicator nodes include purity, particle size, moisture content, and impurities. The relationship types in the raw material-reaction-product knowledge graph include chemical transformation relationships, equipment association relationships, and quality transfer relationships. The chemical transformation relationships include... as well as The graph neural network adopts the GraphSAGE architecture, with mean aggregation as the aggregation function and an embedding dimension of 128. It calculates the correlation probability between raw material batches, reaction conditions, and final product quality through node classification and link prediction.

7. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The equipment failure prediction model includes an isolated forest algorithm model and a long short-term memory network model. The isolated forest algorithm model has 100 trees, a subsampling rate of 0.8, and an anomaly score threshold of 0.6, and is used to detect anomalies in the equipment's operating status parameters. The long short-term memory network model is used to construct time series data of length 72 according to a 5-minute sampling interval, and outputs the probability of failure of the lithium sulfide reactor in the next 72 hours based on the time series data.

8. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The quality anomaly diagnosis module includes a Bayesian network model, which calculates the posterior probability of deviations leading to quality anomalies based on the conditional probability relationships between raw material impurity content, raw material moisture content, temperature stability, stirring efficiency, product purity, and particle size distribution. The multi-objective optimization decision module uses a Q-Learning-based reinforcement learning algorithm combined with a non-dominated sorting genetic algorithm II to generate the process adjustment strategy. The optimization objectives include product purity, particle size distribution, energy consumption, and production efficiency, and the adjustment range of the temperature setpoint generated in a single iteration is constrained to a certain limit. Within the range, the pressure setpoint adjustment range is constrained. Within the specified range, the adjustment range of the hydrogen sulfide flow rate setpoint is constrained. Within the range.

9. The lithium sulfide quality anomaly diagnosis and process adjustment system according to claim 1, characterized in that, The closed-loop control execution module sends the setpoint adjustment command to the distributed control system via the OPC UA protocol; the safety linkage module responds when the gas detection alarm system detects that the hydrogen sulfide leakage concentration reaches or exceeds [a certain level]. When this occurs, the safety instrumented system is triggered to perform an emergency shut-off valve closure and a nitrogen purging action; The closed-loop control execution module is also used to link with the manufacturing execution system and the laboratory information management system to automatically generate electronic batch records and associate process curves, quality inspection results and anomaly diagnosis results with the corresponding batches for storage.

10. A method for diagnosing quality anomalies and adjusting processes in lithium sulfide production, applied to a lithium sulfide production process using hydrogen sulfide, lithium hydroxide, and N-methylpyrrolidone as raw materials or solvent systems, characterized in that... include: Acquire real-time production data during the lithium sulfide production process. The real-time production data includes process parameters, safety status parameters, production execution data, raw material quality data, equipment operating status parameters, and offline quality testing data. Visual images of the lithium sulfide material produced by the reaction are acquired in an inert gas protective environment, and the visual images are input into a pre-trained image segmentation and classification model to output real-time visual quality detection results, which include crystal form distribution, particle size distribution and surface defect information. The real-time production data and the real-time visual quality inspection results are input into a pre-established process and quality mapping model to predict product quality trends. Based on a raw material-reaction-product knowledge graph containing a graph neural network, the quality transfer of lithium sulfide from raw material to product is tracked, and the quality transfer tracking results are output. When the product quality trend prediction result or the real-time visual quality detection result deviates from the preset quality standard, root cause reasoning is performed based on the quality transfer tracking result to determine the key process nodes and causes of the quality abnormality. The equipment operating status parameters are input into a pre-trained equipment fault prediction model, which outputs the probability of failure of the lithium sulfide reactor within a future preset time period. Based on the key process nodes, the causes of the anomalies, the probability of failure, and the preset safe operation boundaries, a process adjustment strategy for the current production state is generated, and a set value adjustment instruction is generated based on the process adjustment strategy. The setpoint adjustment command is sent to the underlying control system through the industrial control protocol, and the execution status data fed back by the underlying control system is received to dynamically adjust the lithium sulfide reactor. When the gas detection alarm system detects that the concentration of hydrogen sulfide leakage reaches or exceeds the preset safety threshold, it triggers the safety instrumented system to perform emergency shut-off interlocking and nitrogen purging actions.

Images