Automatic control system and control method for producing lithium carbonate

Abstract

The invention relates to the technical field of digital control, in particular to an automatic control system and method for producing lithium carbonate. The method comprises the following steps: acquiring sensing data of a lithium carbonate production line, and binding the sensing data to obtain a process section field state factor set; performing fine-tuning embedding time state transition according to the process section field state factor set to obtain a multi-dimensional control characterization flow; performing control instruction coding based on the multi-dimensional control representation flow to obtain a multi-channel control factor set; carrying out field load parameter re-calculation and micro-cycle feedback modulation by combining the obtained real-time lithium carbonate production line sensing data to obtain a control modulation response matrix; based on the control modulation response matrix, reverse adjustment is carried out on the process diagram structure, and a process anomaly feedback body is obtained; the real-time response capability of the lithium carbonate production line to raw liquid fluctuation, equipment state change and working condition abnormity is enhanced, and the product purity, the yield and the production line operation stability are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of digital control technology, and in particular to an automatic control system and a control method for producing lithium carbonate. Background Art

[0002] Currently, the industrial production of lithium carbonate primarily involves extracting lithium from salt lake brine, extracting lithium from spodumene by calcining it, and extracting lithium from recycled materials. Among these, extracting lithium from salt lake brine has become the mainstream industrial lithium extraction method both domestically and internationally due to its abundant raw material resources and low production costs. However, the lithium carbonate production process involves multiple stages of material separation, solution concentration, precipitation reaction, filtration, and drying. The process is long, the reaction conditions are complex, and it is significantly affected by factors such as temperature, pH, reaction time, and fluctuations in the stock solution composition. Traditional control methods make it difficult to achieve stable, continuous, and efficient industrial operation. Most existing production lines employ a distributed control structure based on a PLC system, implementing either on-off control or proportional-integral regulation with set thresholds for key process parameters such as concentration, temperature, and flow rate. However, this type of control method is generally subject to issues such as response lag, inability to dynamically adapt to fluctuations in stock solution quality, and fixed system parameters, resulting in significant fluctuations in product purity and yield. For example, during the lithium precipitation reaction stage, failure to precisely adjust the Na2CO3 dosage or temperature curve based on the real-time reaction rate can easily lead to excessive precipitation or incomplete crystallization, affecting subsequent filtration efficiency and product quality. During the concentration and evaporation stage, fluctuations in the stock solution concentration can easily cause premature crystallization or imbalance in the mother liquor, leading to equipment scaling and frequent maintenance. Furthermore, differences in the composition of raw materials from batch to batch often require manual intervention to set parameters, lacking intelligent identification and adaptive control capabilities, greatly limiting the level of process continuity and product standardization. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide an automatic control system and control method for producing lithium carbonate to solve at least one of the above technical problems.

[0004] To achieve the above object, an automatic control method for producing lithium carbonate comprises the following steps:

[0005] Step S1: Acquire sensor data of the lithium carbonate production line, and perform structured binding of physical quantities of process sections on the sensor data of the lithium carbonate production line to obtain a set of field status factors of the process sections;

[0006] Step S2: extracting multi-step prediction vectors based on the process section field state factor set, and fine-tuning the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow;

[0007] Step S3: Perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions for the digital twin control simulation results to obtain a multi-channel control factor set;

[0008] Step S4: transmitting the multi-channel control factor set to the production line management platform to execute the production line control task; combining the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain a control modulation response matrix;

[0009] Step S5: tracing the process deviation based on the control modulation response matrix, and reversely adjusting the process diagram structure of the process section field status factor set according to the process deviation tracing result to obtain a process abnormality feedback body;

[0010] Step S6: Using the process abnormality feedback body as the input of a new round of step S2, iterate steps S2 to S5 repeatedly until the process offset is less than a preset process offset threshold.

[0011] Optionally, step S1 specifically includes:

[0012] Step S11: collecting physical sensor data of each process section of the lithium carbonate production line and summarizing them according to a preset sampling period to obtain sensor data of the lithium carbonate production line;

[0013] Step S12: performing data cleaning and time series alignment on the sensor data of the lithium carbonate production line, removing missing values and outliers, and completing the unified time step to generate a preprocessed data set;

[0014] Step S13: performing physical quantity-process section mapping and binding according to the pre-processed data set to obtain a physical quantity-process section relationship table;

[0015] Step S14: classifying and integrating the pre-processed data sets based on the physical quantity-process section relationship table, and outputting a structured set by process section;

[0016] Step S15: convert the structured set into a state factor vector, and splice them in sequence according to the data collection time to form a process section field state factor set.

[0017] Optionally, step S13 is specifically as follows:

[0018] Step S131: extracting the sensor ID, sensor installation location number, and data collection time corresponding to each type of physical quantity in the preprocessed data set, thereby constructing a basic metadata index table;

[0019] Step S132: Connecting to the production line management platform, loading the lithium carbonate production line process flow configuration file, extracting the spatial location range, physical function label and corresponding process section number of each process section, and constructing a process section structure table;

[0020] Step S133: Compare the basic metadata index table with the process section structure table, establish a mapping relationship between the sensor installation position number and the spatial position range of the process section, and mark the corresponding process section number of each physical quantity to obtain a physical quantity-process section mapping relationship;

[0021] Step S134: The physical quantity-process section mapping relationship is structured and expressed according to a preset process section number-physical quantity type-sensor identifier-data acquisition frequency framework to form a structured mapping relationship table;

[0022] Step S135: Represent the structured mapping relationship table in a graph structure, where nodes represent process sections, edges represent binding paths of physical quantities, and node attributes are physical quantity sets, to generate a physical quantity-process section relationship table.

[0023] Optionally, the multi-step prediction vector extraction in step S2 is specifically as follows:

[0024] The window length is set to T = 60 seconds, the sliding step is set to 10 seconds, and the process section field status factor set is sliced in time series to generate a set of time series fragments;

[0025] Perform multi-channel trend deconstruction on the time series fragment set to obtain a set of standardized state trend trajectories;

[0026] Perform multi-layer sliding convolution based on the standardized state trend trajectory set to predict the trend prediction vector of the next 6 sliding steps and generate the initial multi-step prediction sequence set;

[0027] The initial multi-step prediction sequence set is cross-adjusted based on the data correlation between segments, low confidence vectors are eliminated, and a refined prediction vector set is obtained;

[0028] According to the formatting framework of process section number, physical quantity type and time series, the refined prediction vector set is formatted and packaged to obtain a multi-step prediction vector set.

[0029] Optionally, the fine-tuning embedding time state transition in step S2 is specifically as follows:

[0030] Calculate the time series similarity between the multi-step prediction vector set and the current state factor in the process section field state factor set, and perform reweighted fine-tuning fusion on the multi-step prediction vector set based on the time series similarity to generate a re-adjusted prediction vector group;

[0031] According to the process section structure table, the readjusted prediction vector group is synchronously embedded in the upstream and downstream sections to generate a section-level time migration sequence;

[0032] Extract the change rate, mutation frequency and steady-state interval characteristics in the segment-level time migration sequence and encode them into state migration vectors;

[0033] Compress and map the state transition vector and output a low-dimensional control representation code;

[0034] The low-dimensional control representation codes of each segment are spliced in time sequence and labeled with physical action labels to obtain a multi-dimensional control representation stream.

[0035] Optionally, step S3 specifically includes:

[0036] Step S31: Connecting to the digital twin physical simulation platform; loading the corresponding digital twin simulation model from the digital twin physical simulation platform according to the physical action label in the multi-dimensional control representation flow, initializing the simulation input parameters, and generating a simulation model configuration set;

[0037] Step S32: inputting the simulation model configuration set into the digital twin physical simulation platform, running multi-segment parallel process response simulation, and obtaining simulation response data;

[0038] Step S33: performing control target index matching on the simulation response data, and performing control simulation performance evaluation based on the control target index, selecting the control input trajectory with the highest performance evaluation, and obtaining the optimal control trajectory set;

[0039] Step S34: discretizing the control variables in the optimal control trajectory set, reorganizing the format and performing time synchronization to generate a standard control instruction stream;

[0040] Step S35: According to the control instruction execution relationship in the lithium carbonate production line process configuration file, the standard control instruction flow is split into control channels and concurrently mapped to generate a multi-channel control factor set.

[0041] Optionally, the on-site load parameter recalculation and micro-cycle feedback modulation in step S4 are specifically as follows:

[0042] Receive real-time lithium carbonate production line sensor data and actuator control logs of executed control instructions after production line control;

[0043] Based on the real-time lithium carbonate production line sensor data and the actuator control log, logical judgment is performed to extract the corresponding control variable execution feedback value to obtain the actuator operation status data set;

[0044] Align the actuator operating status data set with the multi-channel control factor set before production line control, calculate the response offset of each control channel, perform leveling, and generate a load parameter recalculation matrix.

[0045] Based on the load parameter recalculation matrix and using the inertia compensation time window and equipment dynamic response characteristics of each process section in the lithium carbonate production line process configuration file, the control perturbation within the preset micro-cycle feedback interval is extracted to generate a feedback modulation factor group;

[0046] The feedback modulation factor group is matched with the process section structure table through channel mapping, and the control modulation response matrix is output.

[0047] Optionally, step S5 is specifically:

[0048] Step S51: Calculating the actual state change amplitude based on the response offset of each control channel in the control modulation response matrix and the process section field state factor set, calibrating the real-time control response offset value of the process section, and thereby obtaining a process control response offset mapping table;

[0049] Step S52: Compare the process control response offset mapping table with the physical quantity-process section relationship table, locate the process section where the offset occurs and the associated physical quantity according to the process section number, and calculate the offset cause weight to obtain a process offset correlation factor group;

[0050] Step S53: Based on the process offset correlation factor group, reverse tracing of the path is performed in the graph structure of the physical quantity-process section relationship table to obtain the offset impact path;

[0051] Step S54: structurally adjust the offset impact path and generate a process map reverse adjustment index table;

[0052] Step S55: Apply the process map reverse adjustment index table to the process section field status factor set, update the structured binding relationship, mark the abnormal response segment, and generate a process abnormality feedback body.

[0053] Optionally, step S54 is specifically as follows:

[0054] Step S541: extracting the control correlation, process effect intensity, and conduction direction of each physical quantity node in the offset impact path, and constructing an offset path structure factor matrix;

[0055] Step S542: analyzing the influence intensity of the control response offset on each physical quantity node according to the offset path structure factor matrix, and marking the path segments with an influence intensity exceeding [0.15, 0.25] to obtain a path abnormal segment identification set;

[0056] Step S543: performing node replacement, weight reconstruction, and direction optimization on the path segments in the path abnormality segment identification set, and recording the structure change operations to obtain a path structure change record;

[0057] Step S544: index and hierarchically encode the path structure change records according to the process section number and physical quantity identifier to obtain a process map reverse adjustment index table.

[0058] The present invention aims to solve the problems existing in traditional lithium carbonate production, such as delayed control response, large process fluctuations, frequent manual intervention, and low process intelligence level, and proposes a multi-stage collaborative closed-loop intelligent control method, which has a significant technical improvement effect. By structurally binding the physical quantities of the process sections of various sensor data of the lithium carbonate production line in step S1, the operating status of each process section can be accurately restored, forming a set of field status factors that can be used for control modeling, avoiding the problem of "physical quantities being separated from process semantics" in traditional data acquisition; wherein in S13, a graph structure mapping relationship table is constructed through process section number, physical quantity type, sensor identification and other contents, which can enhance the structural clarity of the data and the accuracy of subsequent process mapping, and is conducive to multi-region segmented modeling in the downstream model. In step S2, time series slicing is performed by setting the window length T = 60 seconds and the sliding step size 10 seconds to ensure that the state trend can cover the key process change cycle, and both real-time performance and trend integrity can be taken into account in the prediction vector extraction; the combination of multi-channel trend deconstruction processing and multi-layer sliding convolution can effectively extract short-term trend changes and enhance the local accuracy of the prediction, avoiding the omission of inter-segment linkage features in traditional single-variable modeling; the time series similarity calculation mechanism and upstream and downstream embedding strategy introduced in the fine-tuning of embedded time state migration improve the adaptability of the prediction results on the process path, making each step of the prediction more consistent with the actual control response path; the low-dimensional control representation flow after the compressed mapping of the state migration vector takes into account both feature expression ability and system computational efficiency. Step S3 connects to the digital twin simulation platform and, based on the multi-dimensional control representation flow, calls the simulation model corresponding to the process section. This allows for pre-simulation of the response effects of various control instructions on production line operation, avoiding the waste of raw materials and increased equipment load caused by actual trial and error. Control simulation performance evaluation ensures that the final issued instruction trajectory is the most performant, improving the effectiveness and target consistency of the control instructions. Strict reference to the process flow configuration file when splitting the control channel helps ensure process consistency during the instruction issuance process. The actuator control log and sensor data alignment mechanism introduced in step S4 ensures that the feedback source is authentic and traceable. The load parameter recalculation matrix can be used to quantify control deviations in real time. The inertia compensation time window parameters in the process flow file are combined with the dynamic response characteristics of the equipment to extract the perturbation amount and generate a feedback modulation factor group, thereby achieving small dynamic adjustments within a relatively short period of time, ensuring that the system has fine-grained, high-frequency fine-tuning capabilities, and improving response sensitivity and working condition adaptability.Step S5 can accurately locate the source area of the offset by mapping the response offset of the control channel, and use the process map structure to perform reverse path tracking and structural adjustment to achieve the cause analysis of the offset and restore the impact range; in the structural adjustment, the path segment influence intensity is screened in the range of [0.15, 0.25], which is conducive to identifying the key path of medium-intensity disturbances, avoiding both overreaction to weak disturbances and loss of control due to inertia neglect of strong disturbances, making the control more precise and reasonable; path structure change operations such as node replacement and weight reconstruction can dynamically update the process model structure, providing input conditions with adaptability to the current production scenario for subsequent feedback body generation. Ultimately, this solution forms a data-driven, self-correcting, and continuously optimized closed-loop control architecture by looping the process abnormality feedback body into the control modeling process. It has strong robustness and high adaptability, significantly enhancing the real-time response capability of the lithium carbonate production line to raw liquid fluctuations, equipment status changes, and abnormal operating conditions, and effectively improving product purity, yield, and production line operation stability.

[0059] Optionally, this specification also provides an automatic control system for producing lithium carbonate, which is used to execute the automatic control method for producing lithium carbonate as described above, and the automatic control system for producing lithium carbonate includes:

[0060] The field status recognition module is used to obtain the sensor data of the lithium carbonate production line and perform structured binding of the physical quantities of the process sections of the lithium carbonate production line sensor data to obtain the field status factor set of the process sections;

[0061] The multi-step prediction module is used to extract multi-step prediction vectors based on the process section field state factor set, and fine-tune the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow;

[0062] The control simulation module is used to perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions of the digital twin control simulation results to obtain a multi-channel control factor set;

[0063] The production line control module is used to transmit the multi-channel control factor set to the production line management platform to execute the production line control task; it combines the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain the control modulation response matrix;

[0064] The process deviation optimization module is used to trace the process deviation based on the control modulation response matrix, and to reversely adjust the process diagram structure of the process section field status factor set according to the process deviation traceability results to obtain the process abnormality feedback body;

[0065] The iterative parameter adjustment module is used to use the process abnormality feedback body as the input of the multi-step prediction module for a new round, and repeatedly iterate the multi-step prediction module to the process offset optimization module until the process offset is less than the preset process offset threshold.

[0066] The automatic control system for producing lithium carbonate of the present invention can implement any automatic control method for producing lithium carbonate of the present invention, and is used to combine the operation and signal transmission medium between various modules to complete the automatic control method for producing lithium carbonate. The modules within the system cooperate with each other, thereby significantly enhancing the real-time response capability of the lithium carbonate production line to raw liquid fluctuations, equipment status changes and abnormal operating conditions, and effectively improving product purity, yield and production line operation stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0068] Figure 1 Schematic diagram of the process flow of the automatic control method for producing lithium carbonate of the present invention;

[0069] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0070] Figure 3 Detailed flowchart of step S13 in the present invention;

[0071] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0072] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0073] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0074] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0075] To achieve this, please refer to Figures 1 to 3 The present invention provides an automatic control method for producing lithium carbonate, the method comprising the following steps:

[0076] Step S1: Acquire sensor data of the lithium carbonate production line, and perform structured binding of physical quantities of process sections on the sensor data of the lithium carbonate production line to obtain a set of field status factors of the process sections;

[0077] In this embodiment, industrial sensor terminals (including conductivity sensors, pH probes, thermocouples, flow meters, etc.) deployed in each key process section of the lithium carbonate production line are used to collect various types of physical quantity data with a sampling period of 2 seconds, and are connected to the edge computing module for preliminary caching. Using the data classification service deployed on the edge side, the original sensor data is labeled and classified according to the "process section number-physical quantity type" according to the correspondence between the physical quantities of the sections defined in the process flow configuration file. Subsequently, after a process of structured data cleaning and time alignment, including linear interpolation filling of missing values and median replacement of mutation values, it is uniformly processed into a continuous state time series with a step length of 5 seconds. Finally, through the binding mapping rules of physical quantities and process sections, the classified physical quantity data is encapsulated into a structured state factor vector to form a process section field state factor set covering process sections such as lithium precipitation, concentration, filtration, and drying.

[0078] Step S2: extracting multi-step prediction vectors based on the process section field state factor set, and fine-tuning the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow;

[0079] In this embodiment, the process section on-site state factor set is used as input, and the trend value and change direction within the next 5 time steps (i.e., 50 seconds) are extracted from the time series of each process section respectively, and combined into a multi-step prediction vector. During the extraction process, the prediction sequence is updated with a step of 20 seconds through a sliding window mechanism. After generating the initial prediction vector, it is compared with the current state factor item by item, and the dynamic Euclidean distance is calculated as the time series similarity index. According to the similarity, the different prediction vectors are given adjustment weights, and linear fusion is performed to generate a readjusted prediction vector group. Then, according to the action path between the upstream and downstream sections in the process map, the prediction results are embedded in the state prediction of the upstream and downstream sections in a synchronous manner to obtain a time migration sequence covering multiple process sections. After that, the gradient change rate, number of mutations, steady-state maintenance time and other features in the sequence are extracted and uniformly encoded as a state migration vector. Finally, the state vector is reduced in dimension by the principal component compression method, and a low-dimensional representation is output. After each control item is labeled with a physical quantity, it is spliced to generate a multi-dimensional control representation stream.

[0080] Step S3: Perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions for the digital twin control simulation results to obtain a multi-channel control factor set;

[0081] In the present embodiment, after receiving the multidimensional control characterization flow, a digital mirror system of a lithium carbonate production line reflecting the actual physical structure is built in a virtual environment. The system loads the actual process flow configuration file, including the number, spatial position, medium type, action mechanism and dynamic response time of each process section. For example, the precipitation section response delay is set to 12 seconds, and the drying section is 8 seconds. The multidimensional control characterization flow is input into the simulation system one by one, and the state evolution path after the control behavior is triggered is simulated to evaluate its impact on the target production capacity, purity and energy consumption indicators. By comparing the control input and the simulated feedback hour by hour, the control action set with the best response is extracted, and encoded by physical quantity, action point number and time delay to form a transmittable multi-channel control factor set. Each control factor contains five fields: control variable name, application range, control intensity, target process section number and application time point.

[0082] Step S4: transmitting the multi-channel control factor set to the production line management platform to execute the production line control task; combining the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain a control modulation response matrix;

[0083] In this embodiment, the multi-channel control factor set is sent to the production line management platform via a standard control interface, and the platform maps it to the corresponding execution unit according to the preset control protocol. Whenever the control action is completed, real-time sensor data is collected 10 seconds before and after the control, and the control execution log is obtained. The target change of the control factor is compared with the change amplitude of the actual sensor data, and the response offset value of the control variable is extracted. For example, the flow regulation target is to increase by 15%, but the actual increase is only 10%, and the offset value is 5%. The offset values of all control channels are sorted out, and a load parameter recalculation matrix is constructed. Combined with the inertia compensation time window set in the process flow configuration file (such as 6 seconds for the sedimentation section) and the dynamic response characteristics of the equipment (such as the valve adjustment speed is 0.2 seconds / unit flow), the control perturbation amount is extracted within the 5-second micro-cycle feedback interval to form a feedback modulation factor group. This factor group is mapped to the process section structure to generate a structured control modulation response matrix, which serves as the basic data for subsequent traceability and adjustment.

[0084] Step S5: tracing the process deviation based on the control modulation response matrix, and reversely adjusting the process diagram structure of the process section field status factor set according to the process deviation tracing result to obtain a process abnormality feedback body;

[0085] In this embodiment, based on the control modulation response matrix, the relationship between the response offset of each control channel and the state change of the process segment is analyzed item by item. This creates a control response offset mapping table, which records the process segment number and the corresponding offset magnitude. This mapping table is then compared against the physical quantity-process segment binding table to identify the specific process segment and its associated physical quantity where the offset occurred. By comparing the influence of various factors at different path nodes, the physical quantities likely responsible for each offset are calculated and the offset causal weight is assigned. For example, if pH value has a high correlation with the offset magnitude, this physical quantity is given a higher weight. Subsequently, a reverse path tracing is performed within the process map, searching backward from the offset point along the map edge for the possible source physical node. The path structure is extracted, and the impact direction and intensity of each segment are recorded. These paths are then analyzed to identify path segments with an impact intensity greater than 20%, and node replacement and path structure adjustments are performed. Finally, these adjustment operations are encoded into a structured table by process segment number, which is used to update the state factor set and generate a new process anomaly feedback body for the next round of regulation input.

[0086] Step S6: Using the process abnormality feedback body as the input of a new round of step S2, iterate steps S2 to S5 repeatedly until the process offset is less than a preset process offset threshold.

[0087] In this embodiment, the generated process abnormality feedback body is used as a new round of prediction input, replacing the original state factor set, and re-entering the prediction process. Each time a complete cycle from prediction to control to feedback adjustment is completed, the mapping relationship between the state factor and the control factor is updated, and the maximum offset value of the current process offset is recorded. The basis for judging whether to enter the next round of loop is whether the current maximum offset value is lower than the preset threshold, 0.5%. If this requirement is not met, continue to enter the next round of iteration with the adjusted state feedback as input. During the cycle, the state stability, response consistency and prediction accuracy of multiple key process sections are monitored simultaneously to ensure that the adjustment effect is optimized while maintaining process stability. Finally, when all offset values are within the threshold, the control iteration process ends, and the final control result is used as the basis for updating the real-time control instructions.

[0088] Optionally, step S1 specifically includes:

[0089] Step S11: collecting physical sensor data of each process section of the lithium carbonate production line and summarizing them according to a preset sampling period to obtain sensor data of the lithium carbonate production line;

[0090] In this embodiment, in the lithium carbonate production line, six types of physical sensors, namely temperature, pH, conductivity, flow, pressure and liquid level, are respectively set in key process sections such as extraction, precipitation, neutralization, centrifugation and drying. The industrial-grade RS485 communication protocol is selected to connect with the PLC control terminal, and the data acquisition task is performed with a fixed sampling period of 5 seconds. The numbers of various sensors are combined according to the process section number plus the equipment type code. For example, "D-03-TEMP" represents the third temperature sensor in the drying section. The acquisition task is triggered regularly by the central control system. After each batch of data is collected, it is uniformly summarized in the production data center server. The storage structure adopts a process section-timestamp dual index method. In order to ensure the synchronization of each segment of data, the acquisition program adopts a data caching mechanism. When a segment of data is not reported due to sensor delay, it is temporarily not written to the database. It is not until all data of the current sampling period are ready that the data is stored in batches.

[0091] Step S12: performing data cleaning and time series alignment on the sensor data of the lithium carbonate production line, removing missing values and outliers, and completing the unified time step to generate a preprocessed data set;

[0092] In this embodiment, the sensor data of the lithium carbonate production line summarized in the previous step is used as input, and the timestamp alignment operation is first performed. Based on the unified time step of 5 seconds, a complete timeline sequence is constructed. The missing time points on the time axis are filled by linear interpolation. If a sensor does not return valid data for three consecutive times, it is marked with "NA" in the data set and recorded as a missing segment. During the data cleaning process, a threshold interval is set, such as a pH value setting range of 3.5 to 8.5 and a temperature of 15°C to 95°C. When the monitoring data exceeds the physical reasonable range, it is removed. In order to avoid accidental deletion of instantaneous fluctuations, the abnormal value is required to be maintained within 3 consecutive cycles before it is cleared. The final generated preprocessed data set consists of a three-field structure composed of sensor ID, timestamp and numerical value, and the original data identifier is retained for subsequent comparison.

[0093] Step S13: performing physical quantity-process section mapping and binding according to the pre-processed data set to obtain a physical quantity-process section relationship table;

[0094] In this embodiment, after the preprocessing is completed, the sensor and process segment binding table established in the configuration file is used to map and match the physical quantity and process segment for each data. The binding table is constructed in the form of three columns, namely "sensor number", "process segment number" and "physical quantity type". For example, the sensor numbered "C-04-PH" can be corresponded to the physical quantities of "precipitation segment" and "pH value" through the table items. During the matching process, if a sensor is not defined in the configuration table, it will be recorded in the unbound list for manual supplementation. In order to improve the mapping efficiency, the system hashes the sensor number to establish a quick index table, and the matching results are directly appended to the original data set to form new data fields "bound process segment number" and "physical quantity name". Through this mapping, unstructured sensor data can be quickly classified into actual process segments, providing a basis for subsequent data structure reorganization.

[0095] Step S14: classifying and integrating the pre-processed data sets based on the physical quantity-process section relationship table, and outputting a structured set by process section;

[0096] In this embodiment, after obtaining a data set with physical quantities and process segment identifiers, all data are partitioned and integrated according to the process segment number. For example, after extracting all the data identified as "drying segment", they are sorted in ascending order by timestamp, classified and organized by the "temperature", "flow", and "pressure" fields to construct a structured data set for the drying segment. The data set for each process segment contains a time column, multiple physical quantity columns, and a data source identification column. During the integration process, a memory buffer partitioning strategy is adopted. First, a cache area is created according to the process segment number, and then the corresponding data is filled in sequentially. After filling, the result structure is written uniformly. For cross-segment data (such as temperature sensors arranged across two segments at the same time), they are classified into process segments with a relatively high proportion according to the physical center point. The final output result is a structured set containing all process segments, and each set encapsulates the data details within all valid time ranges of each segment.

[0097] Step S15: convert the structured set into a state factor vector, and splice them in sequence according to the data collection time to form a process section field state factor set.

[0098] In this embodiment, the physical quantities in the output structured set are spliced in chronological order to form a time-continuous state factor vector stream. Taking the drying section as an example, the temperature, flow, and pressure data are arranged in sequence by field at a data sampling frequency of once every 5 seconds to construct a state factor vector for a single time point. The vectors of all time points are arranged in ascending order of time to form a state factor sequence for the drying section. In order to facilitate the processing of asynchronous sampling in different sections, interpolation and normalization are performed on each state factor sequence to ensure that each time point has a complete factor vector. The final result is a field state factor set that includes all process sections. The structure is a two-layer structure of process section number and corresponding time series. Each sequence node contains a set of state factor data for subsequent predictive modeling and control response analysis.

[0099] Optionally, step S13 is specifically as follows:

[0100] Step S131: extracting the sensor ID, sensor installation location number, and data collection time corresponding to each type of physical quantity in the preprocessed data set, thereby constructing a basic metadata index table;

[0101] In this embodiment, before implementing structured management on the preprocessed dataset, the basic key information fields are first extracted. These include the sensor ID, physical quantity type, acquisition time, and the specific location number where the sensor is installed, corresponding to each piece of data. Taking actual data as an example, if a temperature data record is "TEMP-06-123.4-20240508 08:15:10," where "TEMP-06" is the sensor number, "123.4" is the measured value, and "20240508 08:15:10" is the timestamp, the system will interpret this record as having a sensor ID of "TEMP-06," a location number parsed as "D-04-Z3" (representing the third block of the fourth station in the drying section) according to the configuration file, and a data time of "20240508 08:15:10." Through batch processing, the system structurally deconstructs the records of all different physical quantities, sorts them by triples of time, location, and physical quantity type, and finally forms a metadata index table for subsequent binding and comparison, which has the ability of efficient retrieval and spatial segment positioning.

[0102] Step S132: Connecting to the production line management platform, loading the lithium carbonate production line process flow configuration file, extracting the spatial location range, physical function label and corresponding process section number of each process section, and constructing a process section structure table;

[0103] In this embodiment, by establishing communication with the management platform of the lithium carbonate production line, the REST interface is called to load the "lithium carbonate process flow configuration file" stored in the back end. The configuration file adopts the JSON format, and the content includes the number of each process segment, the physical action category, the geographical distribution boundary (defined by the plane coordinate block) and the upstream and downstream connection information. For example, the definition of the "precipitation segment" includes its spatial range as "[X1=12.0, Y1=4.5, X2=16.0, Y2=7.5]", the action label as "precipitation reaction", and the segment number as "SEC-02". After parsing all the paragraphs, a process segment structure table is constructed, arranged in ascending order by number, and accompanied by the action category and spatial boundary fields. The table is structured into a five-field data structure of "segment number-spatial range-physical function-connection upstream segment number-connection downstream segment number", which provides a spatial and logical matching basis for subsequent mapping.

[0104] Step S133: Compare the basic metadata index table with the process section structure table, establish a mapping relationship between the sensor installation position number and the spatial position range of the process section, and mark the corresponding process section number of each physical quantity to obtain a physical quantity-process section mapping relationship;

[0105] In this embodiment, with the sensor installation position number as the core, the established metadata index table is spatially matched with the constructed process section structure table. The specific process is to parse the installation position number of each sensor into the corresponding coordinate area, and then make a regional inclusion judgment with the spatial boundary defined by each process section. If the position number of a temperature sensor is "D-03-Z4", the coordinate point [X=14.5, Y=6.0] is obtained through spatial transcoding, and the inclusion is verified with the boundary of the process section "SEC-02". Once it is confirmed to fall within its defined range, it is marked as belonging to "SEC-02". This process processes all sensors one by one, and records the physical quantity type and belonging section number corresponding to each sensor, thereby generating a complete "physical quantity-process section mapping relationship". The result is subsequently used for the directional binding of process data and functional linkage judgment.

[0106] Step S134: The physical quantity-process section mapping relationship is structured and expressed according to a preset process section number-physical quantity type-sensor identifier-data acquisition frequency framework to form a structured mapping relationship table;

[0107] In this embodiment, the obtained mapping results are organized into a structured expression format. The specific framework is set as a four-level structure, namely "process section number-physical quantity type-sensor identification-acquisition frequency". Taking the drying section (number SEC-05) as an example, its subordinate temperature sensors include "TEMP-05-A" and "TEMP-05-B", and their acquisition frequency is once every 5 seconds, which are mapped into structural units: "SEC-05-temperature-TEMP-05-A-5s", "SEC-05-temperature-TEMP-05-B-5s". Similarly, a full structured mapping is constructed for all physical quantities and their corresponding sensors, and the output is in a tabular form. Each record is a unique identification combination with high indexability and engineering traceability. The structured mapping relationship table can also be exported as a CSV or database table for use in field control systems or data modeling.

[0108] Step S135: Represent the structured mapping relationship table in a graph structure, where nodes represent process sections, edges represent binding paths of physical quantities, and node attributes are physical quantity sets, to generate a physical quantity-process section relationship table.

[0109] In this embodiment, the structured mapping relationship is further abstracted into a graph structure. In the process of graph composition, the process segment number is used as the node in the graph, and the sensor mapping path of each type of physical quantity is used as the edge in the graph. The direction of the edge is defined according to the material flow direction of the production line, and the node attribute is set to the set of physical quantity types in the current segment, such as "{temperature, pressure, pH}". If there is a binding relationship between the flow of physical quantities between two process segments, such as the pH sensor data of a precipitation segment also affects the downstream neutralization segment, then an edge of "SEC-02→SEC-03" is drawn in the graph, and "pH" is marked as the edge attribute. The physical quantity-process segment relationship table constructed in this way is presented as a directed attribute graph structure, which has good visualization capabilities and cross-segment data dependency expression capabilities, and can be used to track data sources and flow paths in engineering analysis.

[0110] Optionally, the multi-step prediction vector extraction in step S2 is specifically as follows:

[0111] The window length is set to T = 60 seconds, the sliding step is set to 10 seconds, and the process section field status factor set is sliced in time series to generate a set of time series fragments;

[0112] In this embodiment, before constructing the multi-step prediction input data, the collected process section field state factor set is trimmed according to a unified time window. The length of each time window is set to 60 seconds, the sliding step is 10 seconds, and the entire time series data set is traversed in a non-overlapping manner. Taking a certain temperature factor as an example, if its sampling frequency within one minute is 1Hz, each window will contain 60 data points. By sliding forward 10 seconds each time from the starting point, the system forms multiple continuous but partially overlapping time periods, and the number of state factors retained in each segment remains consistent, thereby ensuring the coverage continuity between segments on the time axis. The final set of time series segments not only retains the complete temporal characteristics, but also ensures that there is sufficient contextual information support for subsequent trend extraction.

[0113] Perform multi-channel trend deconstruction on the time series fragment set to obtain a set of standardized state trend trajectories;

[0114] In this embodiment, based on the previously generated set of time series segments, the data fluctuation trend within each segment is normalized and trend deconstructed according to the physical quantity channel. Each type of physical quantity, such as temperature, pressure, conductivity, etc., may have amplitude differences in different time segments. It is necessary to first perform maximum-minimum normalization and uniformly map the data of each channel to the interval [0,1] to eliminate the influence of dimension and amplitude. Subsequently, local extreme points are extracted and the slope of change is calculated for each channel, and a trend trajectory code is generated based on this structure. For example, after extracting 5 main rising segments and 4 stable segments in the temperature trend, a "U5-S4" coding structure is formed as the basis of the trend trajectory. The trend trajectories of all physical quantity channels are combined by time axis alignment to eventually form a standardized state trend trajectory set for use in the next stage.

[0115] Perform multi-layer sliding convolution based on the standardized state trend trajectory set to predict the trend prediction vector of the next 6 sliding steps and generate the initial multi-step prediction sequence set;

[0116] In this embodiment, continuous sliding window processing is performed on the standardized trend trajectory set, and the short-term dynamic pattern and periodic fluctuation characteristics are extracted by performing the moving convolution operation layer by layer, and these characteristics are used to estimate the future trend direction. Based on the parameters of a sliding step of 10 seconds and a prediction length of 6 steps (i.e., the next 60 seconds), the system starts the prediction process at the end position of each segment. In the adopted structure, the length of each convolution kernel is set to 5, the number of convolution layers is 3, and the stride is 1. The edge information is retained during the convolution process to prevent trend truncation. Under each slide, the system outputs 6 groups of continuous trend estimation vectors, each group contains trend values of multiple physical quantity channels. All output vectors are combined in chronological order to obtain an initial multi-step prediction sequence set, which reflects the trend change direction and intensity of each physical quantity in the future.

[0117] The initial multi-step prediction sequence set is cross-adjusted based on the data correlation between segments, low confidence vectors are eliminated, and a refined prediction vector set is obtained;

[0118] In this embodiment, after the preliminary prediction vector is generated, it will be subjected to cross-segment correlation analysis and confidence screening. First, the prediction results between the upstream and downstream segments are compared in units of process segments. If there is a prediction result that significantly deviates from the coordination of upstream and downstream trends, the prediction vector is marked as a potentially unreliable item. For example, if the pH is predicted to drop sharply in the precipitation segment and the downstream neutralization segment does not show a corresponding pH change trend, the confidence of the vector is assessed to be lower than the threshold value of 0.6. All vectors with insufficient confidence will be eliminated, and the remaining part will enter the re-sorting stage, and the priority will be re-determined based on the continuity of the trend and the internal consistency within the segment, and finally a set of refined prediction vectors with unified structure and clear direction will be formed, providing a high-credibility data foundation for subsequent packaging processing.

[0119] According to the formatting framework of process section number, physical quantity type and time series, the refined prediction vector set is formatted and packaged to obtain a multi-step prediction vector set.

[0120] In this embodiment, after completing the confidence filtering and structural arrangement, the refined prediction results need to be packaged in a unified format for subsequent analysis and control calls. The package structure adopts a three-layer naming format, namely "process section number-physical quantity type-time label". The time label is encoded according to the prediction start time and prediction step, such as "T+10s", "T+20s", etc., to ensure that the sequence is traceable. Taking the example vector "SEC-04-temperature-T+30s" as an example, the corresponding data record is "0.78", which represents the normalized trend value of the physical quantity. All packaged data is stored in a dictionary structure, which supports calling by segment or expanding by time series.

[0121] Optionally, the fine-tuning embedding time state transition in step S2 is specifically as follows:

[0122] Calculate the time series similarity between the multi-step prediction vector set and the current state factor in the process section field state factor set, and perform reweighted fine-tuning fusion on the multi-step prediction vector set based on the time series similarity to generate a re-adjusted prediction vector group;

[0123] In this embodiment, in order to improve the adaptability of the prediction results to the actual state, the change characteristics of each group of multi-step prediction vectors and the current on-site state factors within the same time range are compared. The comparison window is set to 60 seconds, and the matching criteria include the similarity of the change slope, the similarity of the fluctuation frequency, and the coincidence of the extreme point positions. Taking the temperature factor as an example, in a certain process section, its prediction vector shows a continuous upward trend, while the on-site state factor shows a slight fluctuation upward trend. The similarity score is calculated based on the three indicators, and 0.85 is set as the reweighted threshold. For prediction vectors with a score higher than the threshold, a high weight is directly assigned; for those with a score lower than the threshold but not lower than 0.6, the correlation between adjacent physical quantities will be referred to for weight correction. Finally, all prediction vectors are fine-tuned according to the corresponding weights to generate a re-adjusted prediction vector group that is more in line with the current state.

[0124] According to the process section structure table, the readjusted prediction vector group is synchronously embedded in the upstream and downstream sections to generate a section-level time migration sequence;

[0125] In this embodiment, after completing the reweighted fine-tuning, in order to establish a complete dynamic process expression, the re-adjusted prediction vector groups of each segment are sequentially integrated according to the upstream and downstream topological relationship defined in the process segment structure table. In the specific operation, the prediction vector of the prediction start time point is first extracted from the upstream segment, and then gradually extended and embedded downstream in time to ensure that the time label is continuous and the physical quantity coverage is complete. For example, if the precipitation segment (segment number SEC-03) and the filtration segment (segment number SEC-04) form a sequential relationship, the temperature, pH value, conductivity, etc. corresponding to "T+30s" in SEC-03 will be embedded in the prediction input of SEC-04 at "T+40s" in each group of predictions to achieve cross-segment time-lapse feature fusion. Ultimately, each process segment has a segment-level time migration sequence containing its upstream and downstream historical states, which enhances the contextual understanding of future change trends.

[0126] Extract the change rate, mutation frequency and steady-state interval characteristics in the segment-level time migration sequence and encode them into state migration vectors;

[0127] In this embodiment, after obtaining the complete segment-level time migration sequence, in order to further refine the core dynamic features, multiple index analyses are performed on the physical quantity curves in each sequence. The rate of change is expressed as the ratio of the increase or decrease of the value per unit time, and the frequency of mutation is counted as the number of times the difference between the values of two consecutive time points exceeds a threshold (such as 10%). The steady-state interval is defined as the length of a continuous time period with a change rate lower than 0.01. Taking the conductivity of the filter section as an example, if three mutations occur between "T+10s" and "T+60s", the average change rate is 0.03 / s, and the cumulative steady-state duration is 30 seconds, the corresponding indicator code is (R=0.03, J=3, S=30). All indicators are arranged in order according to the physical quantity dimension to form a state migration vector, which serves as an important input feature set for subsequent dimensionality reduction processing.

[0128] Compress and map the state transition vector and output a low-dimensional control representation code;

[0129] In the present embodiment, in order to facilitate the rapid use of state transition information in the control process, the multidimensional state transition vector is projected to a fixed low-dimensional space. A plurality of reference control samples are used to form a projection base in the mapping process, and the dimension after compression is set to 8 dimensions to ensure that both the main trend information is retained and a sufficient compression ratio is provided. The reference control sample refers to a process state data set with a clear process label, a stable state identifier or a known control effect screened out from the historical lithium carbonate production process. These samples are used as a control or base in the compression mapping process to establish the projection relationship of the state transition vector in the low-dimensional space. Taking the precipitation section as an example, its state transition vector length is 45 dimensions. After mapping processing, it is converted into a control coding sequence comprising 8 floating-point numbers, such as (0.12, 0.45, 0.03, ..., 0.68). The coding vector will be used as a simplified description representing the current and predicted future process state for integrated analysis with the control data of other process sections to ensure that control accuracy is maintained on the basis of information concentration.

[0130] The low-dimensional control representation codes of each segment are spliced in time sequence and labeled with physical action labels to obtain a multi-dimensional control representation stream.

[0131] In this embodiment, in order to complete the control input structure, the low-dimensional control codes of each process section are spliced in chronological order to form a complete control characterization flow. An identifier is added before each set of codes, including a segment number (such as SEC-05), a physical action label (such as "separation", "heating"), and a time point label (such as "T+60s") to ensure that the data has source traceability and application scalability. The splicing process is unfolded along the acquisition time axis to form a two-dimensional matrix structure, where each row represents the control state of all process sections at a certain time point, and each column corresponds to a certain type of physical action dimension. The final output result can be used for digital twin simulation input or control decision-making judgment basis, and constitutes the input core of the entire dynamic control system-the multi-dimensional control characterization flow.

[0132] What is particularly important is that the reweighted fine-tuning fusion is specifically:

[0133] Extract the current state factor sequence from the field state factor set of the process section, set the sliding interval according to the length of the time window, and construct the current state benchmark sequence set;

[0134] In this example, during the on-site state factor extraction phase, all state factor data within a continuous time window of T = 60 seconds from the current time point forward is first selected from the process segment on-site state factor set, with a sliding step size of 10 seconds, to form a set of sequences with overlapping time periods. Each sequence is primarily indexed by time and includes multiple physical quantity channels such as temperature, pressure, flow rate, and conductivity, ensuring coverage of the complete process response dimension. During the construction process, data is grouped according to the process segment number to avoid cross-interference between different segment data. This forms a current state benchmark sequence set with segment characteristics, providing a structurally unified reference object for subsequent similarity measurement.

[0135] Each prediction vector in the multi-step prediction vector set is compared with the current state benchmark sequence set item by item, and the prediction-state similarity matrix is output;

[0136] In this embodiment, during the similarity calculation phase, each sequence in the benchmark sequence set obtained in the previous step is sequentially aligned with the predicted trajectory from the multi-step prediction vector set. The sum of squared differences for each physical dimension at the same time position is calculated, ultimately outputting a prediction-state similarity matrix. To control the interference level of abnormal channels, a channel weight filtering threshold is set to 0.05, masking low-contribution channels. Furthermore, to avoid mismatches caused by too short predicted trajectories, only prediction vectors with a length of at least window T are retained in the matching process, ensuring that the similarity matrix is physically comparable.

[0137] According to the prediction-state similarity matrix and the preset time offset weight function, a similarity weighting factor is assigned to each prediction vector to generate a reweighted coefficient vector;

[0138] In this embodiment, during the weighting coefficient construction process, the similarity score is extracted based on each row value of the prediction-state similarity matrix and processed in combination with a set of predefined time offset attenuation weight functions. The time offset weight function adopts a Gaussian decay form, with the center aligned to the current time point and the standard deviation set to 15 seconds to control the degree of influence of the similarity away from the current time window. Ultimately, each prediction vector is assigned a corresponding reweighting factor and aggregated to form a reweighting coefficient vector group with a length consistent with the number of prediction vectors. The entire weight generation process enhances the temporal adaptability of the prediction value while maintaining structural alignment.

[0139] The reweighted coefficient vector is applied to the multi-step prediction vector set, a weighted reconstruction operation is performed, the prediction vectors are fine-tuned and fused, and a re-tuned prediction vector group is generated.

[0140] In this embodiment, during the weighted fine-tuning operation, the reweighted coefficient vector group output from the previous stage is applied to the original multi-step prediction vector set. Weighted reconstruction is then performed by multiplying each term at the prediction time point. Multiple weighted prediction values for the same segment and time position are then normalized and averaged to eliminate the effects of abnormal disturbances. This operation ensures that the fine-tuned prediction results retain the original prediction trend while enhancing their fit with the current process state. The resulting re-adjusted prediction vector group is output as a structured matrix, maintaining the original channel dimensions and time span while enhancing its practical reference value in process control scenarios.

[0141] Optionally, step S3 specifically includes:

[0142] Step S31: Connecting to the digital twin physical simulation platform; loading the corresponding digital twin simulation model from the digital twin physical simulation platform according to the physical action label in the multi-dimensional control representation flow, initializing the simulation input parameters, and generating a simulation model configuration set;

[0143] In this example, a digital twin physical simulation environment for a lithium carbonate production line deployed on a private industrial cloud platform is connected via the HTTPS protocol to verify identity information and obtain relevant model resources. The multidimensional control representation stream contains multiple physical action tags, such as "heat exchange," "ion migration," and "gas-liquid phase transfer." These tags are extracted using pre-defined tag parsing rules. Each physical action tag in the multidimensional control representation stream is parsed to first identify its corresponding process characteristics. For example, heat exchange is typically associated with sections such as the reactor and cooling system, while ion migration is associated with equipment such as the electrolytic cell and reactor. Based on the nature and actual application of these physical actions, these tags are compared with pre-configured digital twin simulation models in the simulation platform. The simulation model registration information includes information such as the physical action type, the process section to which the model applies, and the parameter setting range. Based on the tag parsing results, the system queries the simulation platform's model library and loads the simulation model that matches the tag. For example, if the tag is "heat exchange," the heat conduction model associated with the reactor and cooling system is queried and loaded; if the tag is "ion migration," the electrolytic cell and its corresponding material transport model are loaded. This process ensures that the loaded simulation model is consistent with the physical action labels extracted from the multi-dimensional control characterization flow and can accurately reflect the required control object. Next, the simulation platform automatically initializes the input parameters of each simulation model according to the control objectives and the input parameter requirements of the simulation model. For example, for the "heat exchange" label, the platform will set the temperature, fluid rate and other parameters of the reactor and obtain their values from the control system. For the "ion migration" label, the platform loads relevant data such as the solution concentration and electrolysis voltage in the electrolytic cell as initial input parameters. Through these preset and initialization operations, the simulation model can perform accurate simulations based on real-time data and provide reliable data support for subsequent control decisions.

[0144] Step S32: inputting the simulation model configuration set into the digital twin physical simulation platform, running multi-segment parallel process response simulation, and obtaining simulation response data;

[0145] In this embodiment, in the multi-segment process simulation stage, the simulation model configuration set generated in the previous step is input into the task scheduling interface of the simulation platform, and the parallel operation module is activated. The simulation task is split according to the process segment, and each segment is executed in an independent container. The simulation time span is set to 180 seconds and the time step is 2 seconds to ensure that the response details are fully recorded. The output parameters recorded in each simulation process include real-time reaction temperature changes, material conversion rate, electrolytic cell voltage response, etc. The simulation platform uses pre-loaded physical equations and historical measured data for real-time calculations, and finally outputs a set of simulation response data covering all segments with time alignment features, and packages them into a structured data set as a basis for subsequent analysis.

[0146] Step S33: performing control target index matching on the simulation response data, and performing control simulation performance evaluation based on the control target index, selecting the control input trajectory with the highest performance evaluation, and obtaining the optimal control trajectory set;

[0147] In this embodiment, in the simulation result analysis link, the key indicator values related to the control target in each simulation response sequence are extracted, such as temperature control stability (temperature variance), target concentration achievement rate, voltage curve fluctuation rate, etc., and compared with the established control target interval. Then, the response results that meet the error tolerance range (±3%) and have the highest comprehensive score on the control target indicator are selected from all simulation tasks, and their corresponding control input trajectories are regarded as the preferred paths. The control trajectory retains the curve of the control variable changing over time, such as the solution pump speed change, the voltage gain adjustment curve, etc., forming a set of optimal control trajectory sets including segment numbers, time labels, and control variable values, which provide input basis for actual control execution.

[0148] Step S34: discretizing the control variables in the optimal control trajectory set, reorganizing the format and performing time synchronization to generate a standard control instruction stream;

[0149] In this embodiment, for the selected optimal control trajectory, each control variable (such as pump speed, voltage, and current intensity) is discretized separately, with the unit precision unified to two decimal places, and they are classified and packaged according to the segment number. All control points are uniformly aligned with the timeline based on the simulation time step to ensure that the control instructions can be recognized by the PLC device. Then, all parameter items in the control trajectory are reorganized into a fixed field format, such as [control variable type, target value, timestamp], and finally a set of standard control instruction streams with a clear structure that is easy to schedule and execute are output. This instruction stream is used for the subsequent dispatch and execution of the automatic control system.

[0150] Step S35: According to the control instruction execution relationship in the lithium carbonate production line process configuration file, the standard control instruction flow is split into control channels and concurrently mapped to generate a multi-channel control factor set.

[0151] In this embodiment, in the configuration link of sending control instructions, the process flow configuration file of the lithium carbonate production line is read, and the execution channel information corresponding to each control instruction is extracted, such as the binding relationship with the PLC controller port, the synchronization identifier of the feedback sensor, etc. According to the target control point, physical action type and execution priority of the instruction, the standard control instruction stream is split into multiple control channels. For sections with concurrent control requirements, such as the stirring rate and heating temperature in the reactor, the instructions will be delivered in parallel within a minimum time difference of 0.1 seconds to form the final multi-channel control factor set. Each control factor is clearly marked with fields such as channel number, execution delay time, and instruction effective period to ensure accurate response of downstream equipment.

[0152] Optionally, the on-site load parameter recalculation and micro-cycle feedback modulation in step S4 are specifically as follows:

[0153] Receive real-time lithium carbonate production line sensor data and actuator control logs of executed control instructions after production line control;

[0154] In this embodiment, after the production line is controlled, the real-time sensor data stream and actuator control system connected to the lithium carbonate production line are connected. Sensor data (such as temperature, pressure, concentration, etc.) from each process section of the production line and the control log of the actuator are received through the industrial protocol. The sensor data is transmitted in real time at a set frequency (for example, updated once per second), and the actuator control log contains the control instruction execution time, control parameters and actual execution results of the equipment. Data transmission is guaranteed to be secure and real-time by an encryption protocol to ensure that the status of the production line can be accurately reflected.

[0155] Based on the real-time lithium carbonate production line sensor data and the actuator control log, logical judgment is performed to extract the corresponding control variable execution feedback value to obtain the actuator operation status data set;

[0156] In this embodiment, after receiving real-time sensor data and control logs, data synchronization is first performed, and predetermined rules (such as timestamp alignment) are used to ensure that the various types of data match. Then, a logical judgment is made on the execution of the instructions in the control log. For example, for a reactor heating control system, if the control log indicates "heating instruction executed," the feedback data actually executed by the actuator, such as the heating temperature change value, is extracted and corrected to ensure that the feedback value matches the actual change. Finally, based on this extracted feedback data, a state data set containing the operating status of each actuator (such as the motor operating status, heating device temperature, etc.) is constructed.

[0157] Align the actuator operating status data set with the multi-channel control factor set before production line control, calculate the response offset of each control channel, perform leveling, and generate a load parameter recalculation matrix.

[0158] In this embodiment, the actuator operating status dataset is field-aligned with the multi-channel control factor set. Fields for each control channel (such as pressure, temperature, and flow) are matched with the corresponding historical control data. All data fields are standardized according to the control factor format of the multi-channel control factor set (for example, temperature values are in degrees Celsius and flow rates are in liters per minute) to ensure consistent data format. After data alignment, the actuator's actual feedback data is compared with the expected target value in the optimal control trajectory set to calculate the response offset for each control channel. Specifically, the offset is the difference between the current feedback value and the target set value. For example, if the reactor temperature is set to 90°C and the actual feedback value is 88°C, the response offset is -2°C. These offsets are then leveled to eliminate long-term deviations caused by equipment variations or operational errors. This leveling process utilizes a dynamic adjustment method based on a time window. For example, within a 15-minute time window, the temperature fluctuation range is observed and the control variable is adjusted accordingly. The leveling process gradually reduces the offset and adjusts it to a value close to the target value through linear interpolation or proportional adjustment. Finally, the adjustment parameters of each control channel are summarized to form a load parameter recalculation matrix.

[0159] Based on the load parameter recalculation matrix and using the inertia compensation time window and equipment dynamic response characteristics of each process section in the lithium carbonate production line process configuration file, the control perturbation within the preset micro-cycle feedback interval is extracted to generate a feedback modulation factor group;

[0160] In this embodiment, the matrix is recalculated based on the load parameters, and the inertia compensation time window and the dynamic response characteristics of the equipment in each process section in the process flow configuration file are combined to extract the control perturbation amount within the preset micro-cycle feedback interval. For example, if the equipment response of a certain process section (such as a concentration tank) is slow, a time window (for example, 30 seconds) will be set, and the dynamic response of the equipment will be predicted based on historical data, and the current control strategy will be adjusted in combination with the feedback perturbation amount. For each control channel, these perturbations (such as pressure fine-tuning, temperature fine-tuning, etc.) will be extracted to form a feedback modulation factor group. These modulation factors will be used to accurately adjust the various control factors in the production line to achieve precise control of the production process.

[0161] The feedback modulation factor group is matched with the process section structure table through channel mapping, and the control modulation response matrix is output.

[0162] In this embodiment, after obtaining the feedback modulation factor set, these modulation factors are mapped to the corresponding process segments by consulting the process segment structure table. For example, if a segment (such as a reactor) contains two control channels: temperature control and pressure control, the perturbation values in the feedback modulation factor set are assigned to the corresponding control channels, and an independent modulation response value is generated for each channel. This process ensures that each control channel is appropriately adjusted, thereby accurately fine-tuning the production line. Ultimately, a control modulation response matrix is output, where each row represents the control response of a process segment and each column represents the corresponding feedback modulation factor.

[0163] What is particularly important is that the logical judgment is:

[0164] Extract the time series sampling timestamps, physical quantity types, and corresponding values from the real-time lithium carbonate production line sensor data to generate a sampling state mapping table;

[0165] In this embodiment, each physical quantity type (for example: reactor outlet temperature, sedimentation tank liquid level, filter press system pressure) and its corresponding sequential sampling timestamp and actual measurement value are extracted from the data stream collected by the lithium carbonate production line sensor. The data of each sensor type adopts a unified sampling period (set here as once every 2 seconds) and is accompanied by timestamp information in the ISO-8601 standard format. In order to ensure the uniformity of the data structure, all sampled values are converted to standard units (such as temperature is unified as ℃, pressure is unified as kPa), and abnormal mutation data caused by sensor drift are eliminated. The extracted content generates a structured sampling state mapping table according to fields such as process section number, channel number, timestamp, physical quantity type and value, which is convenient for subsequent matching and analysis.

[0166] Extract the target setting value, execution time point and channel number of each control channel in the actuator control log to generate a control instruction execution table;

[0167] In this embodiment, the control log from the actuator is parsed one by one to extract the control channel number, control target set value, instruction issuance time point and corresponding control type (such as speed control, temperature setting, liquid inlet valve opening, etc.). The control log is uploaded to the control master station using the industrial Ethernet protocol. All records carry high-precision time tags (less than 1 millisecond error) to ensure time synchronization with the sensor data. Each control instruction is encoded in the format of "channel number-time point-set value" and included in the control instruction execution table. This table uses the logical order of the control instructions as an index to establish a clear data structure, laying a data foundation for subsequent actuator response tracking and control closed-loop feedback.

[0168] According to the sampling state mapping table and the control instruction execution table, the channel number is used as the key value and the execution time point is used as the matching window. The alignment matching is performed within the preset maximum response delay range to obtain the initial control feedback pair;

[0169] In this embodiment, the sampling state mapping table and the control instruction execution table are combined, and the control channel number is used as the index key, and the maximum response delay is set to 20 seconds as the time matching window. Starting from the execution time of each control instruction, the continuous sampling records of the channel within 20 seconds after the time point are searched, and the data whose change rate reaches the set threshold for the first time (such as the temperature change rate is greater than 0.2°C / s) is selected as the response point, and paired with the instruction to form an initial control feedback pair. If there is no valid response within 20 seconds, it is recorded as an empty pair. This alignment process strictly ensures that the response sampling time is no earlier than the control instruction time, ensuring that the feedback is indeed caused by the control action.

[0170] Perform logical consistency judgment on the initial control feedback pairs, eliminate logically abnormal pairs, and obtain effective control feedback pairs;

[0171] In this embodiment, a logical consistency check is performed on the initial control feedback pair formed, and logical abnormal data pairs are eliminated. Specifically, the following rules are included: if the control type is temperature increase, and the feedback value continues to decrease within the response time window, it is marked as a reverse response; if the physical quantity units of the response sampling point are inconsistent or NaN or null values appear, it is marked as a data abnormality; if the control instruction set value changes less than the set tolerance (such as less than 0.1%) and there is no obvious physical quantity change, it is considered an invalid response. All the above-mentioned marked items will be eliminated from the feedback pair, and the remaining parts will be retained as valid control feedback pairs. Logical consistency judgment also includes synchronous judgment of the state of the process section. For example, the feedback temperature increase of the reactor should be valid in the "reacting" state, otherwise it will be eliminated.

[0172] The actual response value of each control variable is extracted from the effective control feedback pair, classified and aggregated by control channel, and the actuator operation status dataset is output.

[0173] In this embodiment, the actual response value of each control variable is extracted from the verified effective control feedback pair. Taking the control channel as the aggregation primary key, the feedback values under the same channel are aggregated and statistically analyzed in chronological order, and the average response level, maximum response amplitude and the time required to actually reach the set value (i.e., response delay) are calculated. For example, for the speed control channel of the stirring motor, the feedback speed sequence is extracted and its change trend is calculated to obtain key indicators such as the startup and stabilization time of the equipment after receiving the instruction. Finally, an actuator operation status data set is constructed and stored by channel number, time series, control variable type, response value and response status for subsequent load adjustment and feedback modulation calculation.

[0174] Optionally, step S5 is specifically:

[0175] Step S51: Calculating the actual state change amplitude based on the response offset of each control channel in the control modulation response matrix and the process section field state factor set, calibrating the real-time control response offset value of the process section, and thereby obtaining a process control response offset mapping table;

[0176] In this embodiment, the response offset of each control channel is extracted from the control modulation response matrix. Combined with the process segment number and the corresponding physical quantity set mapped in the established "physical quantity-process segment relationship table", the actual state change amplitude is calculated channel by channel by comparing the currently collected field state factor data with the standard fluctuation range of the process segment field state factor set in the previous 24 hours. Taking the drying section temperature channel TEMP-05-A as an example, if the standard fluctuation is ±0.8°C and the current temperature offset is +2.1°C, the response offset of this channel in this process segment is calibrated as "high temperature deviation threshold range". The offset result is bound to the segment to which it belongs (SEC-05) and recorded in the real-time control response offset table. A unified format process control response offset mapping table is constructed, with fields including channel ID, process segment number, physical quantity type, offset value and calibration label.

[0177] Step S52: Compare the process control response offset mapping table with the physical quantity-process section relationship table, locate the process section where the offset occurs and the associated physical quantity according to the process section number, and calculate the offset cause weight to obtain a process offset correlation factor group;

[0178] In this embodiment, the generated "Process Control Response Deviation Mapping Table" is compared with the "Physical Quantity-Process Section Relationship Table" at the field level, focusing on matching the physical quantity type and process section number, to find all process sections with deviations and their associated physical quantities. After matching, the causal weight of the influencing factor is calculated for each deviation record. This calculation is based on the frequency of similar deviation events in historical data, the stability coefficient of the associated physical quantity, and the calibration status of the field sensor. The specific form is: Where W i is the weight value of the i-th offset cause (range 0 to 1); F i is the average daily frequency of historical excursion events of the ith excursion cause (unit: times / day); F max is the highest offset frequency among all causes within the period (used for normalization); S i is the stability coefficient of the physical quantity corresponding to the i-th offset cause (range 0 to 1, the higher the value, the more stable); C i is the actual calibration period of the equipment related to the i-th offset cause (unit: day); C max is the maximum allowable calibration cycle (used for normalization); α, β, and γ are weight coefficients for each influencing factor, with default values of α = 0.4, β = 0.3, and γ = 0.3. For example, when a pH excursion occurs in a neutralization section (SEC-03), it is compared with historical pH excursion events and scored based on a frequency of 6 events per day, a maximum excursion frequency of 10 events per day among all causes within the period, a stability coefficient of 0.82, a calibration cycle of 9 days, and a maximum allowable calibration cycle of 14. The resulting weight for this excursion cause is 0.487, which can be compared horizontally with other causes to construct a "process excursion correlation factor group." Statistical analysis indicates that this cause is primarily due to a combination of sensor drift and raw material fluctuations. Each cause is assigned a corresponding weight, ultimately forming a structured "process excursion correlation factor group."

[0179] Step S53: Based on the process offset correlation factor group, reverse tracing of the path is performed in the graph structure of the physical quantity-process section relationship table to obtain the offset impact path;

[0180] In this embodiment, the source of the offset influence is retrieved in reverse in the graph structure of the "physical quantity-process section relationship table" based on the generated "process offset correlation factor group". Each edge in the graph structure represents the flow path of the physical quantity, and the node represents the process section. The input path with the same physical quantity label is traced upstream from the node where the offset occurs. For example, for the pH offset of the neutralization section, if its upstream is the precipitation section (SEC-02) and there is a connection path with an edge attribute of "pH", the data dependency source will be traced back along the path. The maximum tracing depth is set to 3 segments, and the path credibility is scored according to the timestamp and physical quantity synchronization. Finally, an offset influence path from SEC-02 to SEC-03 is output, and its path node sequence, edge attributes and time window matching are recorded for subsequent structural analysis.

[0181] Step S54: structurally adjust the offset impact path and generate a process map reverse adjustment index table;

[0182] In this embodiment, structural adjustments are made to the identified offset influencing paths, that is, the association relationship of the nodes or the path strength of the edges in the graph structure are adjusted to generate a "process map reverse adjustment index table". During implementation, the current physical quantity fluctuation state of all nodes on the path is first compared with the historical stable value interval. If the state fluctuation of a node exceeds 3 times the standard deviation, the priority of the node in the index table is increased and marked as a "first-level adjustment candidate". Then, based on the type of physical quantity in the edge attribute, the edges with connection strength lower than the threshold (such as signal correlation coefficient <0.6) are downgraded or shielded. The final reverse adjustment index table structure includes: priority adjustment segment number, corresponding physical quantity, adjustment strength level and recommended processing actions (such as sensor re-inspection, control logic delay adjustment, etc.), which are used to guide the operation and maintenance or automatic control system to respond quickly.

[0183] Step S55: Apply the process map reverse adjustment index table to the process section field status factor set, update the structured binding relationship, mark the abnormal response segment, and generate a process abnormality feedback body.

[0184] In this embodiment, the generated "process map reverse adjustment index table" is applied one by one to the currently collected "process section field status factor set". The application process includes updating the binding relationship in the original state factor data structure to replace or strengthen the dependent path between the offset physical quantity and the process section. For example, if the original binding path of a process section points to three physical quantity nodes, only two main physical quantity paths are retained after the index table is adjusted, and the rest are marked as abnormal data input. The updated binding structure will be visualized in a graph structure, and the abnormal response segment will be highlighted in the graph (such as highlighted edges, red nodes), and a "process abnormality feedback body" will be generated, which includes the updated binding structure, the retained main physical quantity paths, the abnormal process section number, the abnormal physical quantity, the response intensity score and other fields.

[0185] Optionally, step S54 is specifically as follows:

[0186] Step S541: extracting the control correlation, process effect intensity, and conduction direction of each physical quantity node in the offset impact path, and constructing an offset path structure factor matrix;

[0187] In this embodiment, all physical quantity nodes that appear in the offset impact path are first extracted from the previously obtained physical quantity-process segment relationship diagram. For each node, the control correlation parameter (set as a decimal between 0 and 1, such as "0.32" indicating moderate correlation) is extracted using the control linkage configuration table built into the field control system. Simultaneously, the process effect intensity corresponding to the node is read in conjunction with the process segment configuration file (e.g., the effect coefficient of temperature control in the drying segment is "0.58," while the effect coefficient of pH control in the precipitation segment is "0.81"). Furthermore, offset conduction direction information is extracted from the process map based on the installation sequence of the physical sensors deployed on site and the direction of material flow, encoded as two directional indicators: "→" or "←." These three pieces of information are structurally combined and populated into a two-dimensional matrix structure in node order. Each row of the matrix represents an offset path node, and each column represents "control correlation," "process effect intensity," and "conduction direction encoding," forming the offset path structure factor matrix. This matrix is used to subsequently identify which segments of the path may play a key role in offset propagation.

[0188] Step S542: analyzing the influence intensity of the control response offset on each physical quantity node according to the offset path structure factor matrix, and marking the path segments with an influence intensity exceeding [0.15, 0.25] to obtain a path abnormal segment identification set;

[0189] In this embodiment, the generated offset path structure factor matrix is used to analyze the impact strength of each path node. First, a weighted average of the control correlation and process action strength of each node is calculated, defining the impact strength as 0.6 × control correlation + 0.4 × process action strength. The resulting strength value is then compared with the predefined impact threshold range [0.15, 0.25]. If the strength value of the node's corresponding path segment falls within this range, it indicates a moderate offset diffusion risk and is therefore marked as a "potentially abnormal segment." For example, if a temperature node in a drying section has a control correlation of 0.20 and an action strength of 0.35, its impact strength is 0.6 × 0.20 + 0.4 × 0.35 = 0.26, exceeding the upper limit and thus marked as an "abnormally high-risk segment." This method is then used to traverse all segments in the offset path, extracting the segment numbers and physical quantity identifiers that meet the criteria. Finally, an "abnormal path segment identification set" is output and stored as a structured table. The record fields include the segment number, physical quantity name, impact strength value, and a label (e.g., "high risk" or "medium risk").

[0190] Step S543: performing node replacement, weight reconstruction, and direction optimization on the path segments in the path abnormality segment identification set, and recording the structure change operations to obtain a path structure change record;

[0191] After identifying the "path anomaly segment identification set", this embodiment performs a structural reconstruction operation on the path segments involved one by one. First, for the physical quantity nodes in each abnormal path segment, the adjacent node with the weakest correlation strength is selected as the replacement node, and the node replacement operation is performed. For example, in SEC-04, if the influence intensity of the pH node is too high, the nearby conductivity node with an intensity below 0.05 will be automatically selected as the replacement item, and the conduction link will be updated. Secondly, after completing the node replacement, the value of all edges on the path segment is weighted and reconstructed, that is, the control weight is recalculated, and a new weight is generated in the form of "new node control correlation × action intensity", and the old value in the original path is replaced. Finally, the direction adjustment is performed on the path segment with contradictory direction (that is, the original direction is "→", but the upstream and downstream segment numbers are changed to form "←"), the connection direction in the graph structure is updated, and its conduction vector is rewritten in the graph data structure. Every replacement, reconstruction or direction adjustment operation is recorded, with fields such as the original segment number, modification time, change content, and executor identification marked to form a path structure change record, which serves as a basis for subsequent backtracking or reverse deduction.

[0192] Step S544: index and hierarchically encode the path structure change records according to the process section number and physical quantity identifier to obtain a process map reverse adjustment index table.

[0193] In this embodiment, the path structure change records generated in the above steps are indexed so as to be used for the subsequent reverse correction operation of the process map. The specific approach is: first, the process segment number and physical quantity identifier in each change record are double-field indexed and encoded, and "segment number + physical quantity identifier" is set as the first-level index key, for example, SEC-04 and "pH" generate "SEC-04-pH"; then the records are sorted according to the change time sequence, and the timestamp and change sequence number are attached as the second-level index. Secondly, in order to ensure the traceability of multi-level mapping, the type of change operation (such as "node replacement" and "direction adjustment") is used as the index classification label to establish a three-level index structure of "operation type-process segment-physical quantity". Finally, the system exports the index structure as a reverse adjustment index table. The fields in the table include index key, operation type, original path information, new structure description, execution time, responsible person identification, etc., and support JSON and database table dual format output. This table can be used by the upstream map management engine to dynamically correct the process segment control logic and visually render the path map.

[0194] Optionally, this specification also provides an automatic control system for producing lithium carbonate, which is used to execute the automatic control method for producing lithium carbonate as described above, and the automatic control system for producing lithium carbonate includes:

[0195] The field status recognition module is used to obtain the sensor data of the lithium carbonate production line and perform structured binding of the physical quantities of the process sections of the lithium carbonate production line sensor data to obtain the field status factor set of the process sections;

[0196] The multi-step prediction module is used to extract multi-step prediction vectors based on the process section field state factor set, and fine-tune the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow;

[0197] The control simulation module is used to perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions of the digital twin control simulation results to obtain a multi-channel control factor set;

[0198] The production line control module is used to transmit the multi-channel control factor set to the production line management platform to execute the production line control task; it combines the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain the control modulation response matrix;

[0199] The process deviation optimization module is used to trace the process deviation based on the control modulation response matrix, and to reversely adjust the process diagram structure of the process section field status factor set according to the process deviation traceability results to obtain the process abnormality feedback body;

[0200] The iterative parameter adjustment module is used to use the process abnormality feedback body as the input of the multi-step prediction module for a new round, and repeatedly iterate the multi-step prediction module to the process offset optimization module until the process offset is less than the preset process offset threshold.

[0201] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0202] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. An automatic control method for producing lithium carbonate, characterized in that, The following steps are involved: Step S1: Acquire sensor data of the lithium carbonate production line, and perform structured binding of physical quantities of process sections on the sensor data of the lithium carbonate production line to obtain a set of field status factors of the process sections; Step S2: extracting multi-step prediction vectors based on the process section field state factor set, and fine-tuning the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow; Step S3: Perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions for the digital twin control simulation results to obtain a multi-channel control factor set; Step S4: transmitting the multi-channel control factor set to the production line management platform to execute the production line control task; combining the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain a control modulation response matrix; Step S5: tracing the process deviation based on the control modulation response matrix, and reversely adjusting the process diagram structure of the process section field status factor set according to the process deviation tracing result to obtain a process abnormality feedback body; Step S6: Using the process abnormality feedback body as the input of a new round of step S2, iterate steps S2 to S5 repeatedly until the process offset is less than a preset process offset threshold.

2. The automatic control method for producing lithium carbonate according to claim 1, wherein Step S1 is specifically as follows: Step S11: collecting physical sensor data of each process section of the lithium carbonate production line and summarizing them according to a preset sampling period to obtain sensor data of the lithium carbonate production line; Step S12: performing data cleaning and time series alignment on the sensor data of the lithium carbonate production line, removing missing values and outliers, and completing the unified time step to generate a preprocessed data set; Step S13: performing physical quantity-process section mapping and binding according to the pre-processed data set to obtain a physical quantity-process section relationship table; Step S14: classifying and integrating the pre-processed data sets based on the physical quantity-process section relationship table, and outputting a structured set by process section; Step S15: convert the structured set into a state factor vector, and splice them in sequence according to the data collection time to form a process section field state factor set.

3. The automatic control method for producing lithium carbonate according to claim 2, wherein Step S13 is specifically as follows: Step S131: extracting the sensor ID, sensor installation location number, and data collection time corresponding to each type of physical quantity in the preprocessed data set, thereby constructing a basic metadata index table; Step S132: Connecting to the production line management platform, loading the lithium carbonate production line process flow configuration file, extracting the spatial location range, physical function label and corresponding process section number of each process section, and constructing a process section structure table; Step S133: Compare the basic metadata index table with the process section structure table, establish a mapping relationship between the sensor installation position number and the spatial position range of the process section, and mark the corresponding process section number of each physical quantity to obtain a physical quantity-process section mapping relationship; Step S134: The physical quantity-process section mapping relationship is structured and expressed according to a preset process section number-physical quantity type-sensor identifier-data acquisition frequency framework to form a structured mapping relationship table; Step S135: Represent the structured mapping relationship table in a graph structure, where nodes represent process sections, edges represent binding paths of physical quantities, and node attributes are physical quantity sets, to generate a physical quantity-process section relationship table.

4. The automatic control method for producing lithium carbonate according to claim 2, wherein The multi-step prediction vector extraction in step S2 is specifically as follows: The window length is set to T = 60 seconds, the sliding step is set to 10 seconds, and the process section field status factor set is sliced in time series to generate a set of time series fragments; Perform multi-channel trend deconstruction on the time series fragment set to obtain a set of standardized state trend trajectories; Perform multi-layer sliding convolution based on the standardized state trend trajectory set to predict the trend prediction vector of the next 6 sliding steps and generate the initial multi-step prediction sequence set; The initial multi-step prediction sequence set is cross-adjusted based on the data correlation between segments, low confidence vectors are eliminated, and a refined prediction vector set is obtained; According to the formatting framework of process section number, physical quantity type and time series, the refined prediction vector set is formatted and packaged to obtain a multi-step prediction vector set.

5. The automatic control method for producing lithium carbonate according to claim 1, wherein The fine-tuning embedding time state transition in step S2 is specifically as follows: Calculate the time series similarity between the multi-step prediction vector set and the current state factor in the process section field state factor set, and perform reweighted fine-tuning fusion on the multi-step prediction vector set based on the time series similarity to generate a re-adjusted prediction vector group; According to the process section structure table, the readjusted prediction vector group is synchronously embedded in the upstream and downstream sections to generate a section-level time migration sequence; Extract the change rate, mutation frequency and steady-state interval characteristics in the segment-level time migration sequence and encode them into state migration vectors; Compress and map the state transition vector and output a low-dimensional control representation code; The low-dimensional control representation codes of each segment are spliced in time sequence and labeled with physical action labels to obtain a multi-dimensional control representation stream.

6. The automatic control method for producing lithium carbonate according to claim 1, wherein Step S3 is specifically as follows: Step S31: Connecting to the digital twin physical simulation platform; loading the corresponding digital twin simulation model from the digital twin physical simulation platform according to the physical action label in the multi-dimensional control representation flow, initializing the simulation input parameters, and generating a simulation model configuration set; Step S32: inputting the simulation model configuration set into the digital twin physical simulation platform, running multi-segment parallel process response simulation, and obtaining simulation response data; Step S33: performing control target index matching on the simulation response data, and performing control simulation performance evaluation based on the control target index, selecting the control input trajectory with the highest performance evaluation, and obtaining the optimal control trajectory set; Step S34: discretizing the control variables in the optimal control trajectory set, reorganizing the format and performing time synchronization to generate a standard control instruction stream; Step S35: According to the control instruction execution relationship in the lithium carbonate production line process configuration file, the standard control instruction flow is split into control channels and concurrently mapped to generate a multi-channel control factor set.

7. The automatic control method for producing lithium carbonate according to claim 1, characterized in that, The on-site load parameter recalculation and micro-cycle feedback modulation in step S4 are specifically as follows: Receive real-time lithium carbonate production line sensor data and actuator control logs of executed control instructions after production line control; Based on the real-time lithium carbonate production line sensor data and the actuator control log, logical judgment is performed to extract the corresponding control variable execution feedback value to obtain the actuator operation status data set; Align the actuator operating status data set with the multi-channel control factor set before production line control, calculate the response offset of each control channel, perform leveling, and generate a load parameter recalculation matrix. Based on the load parameter recalculation matrix and using the inertia compensation time window and equipment dynamic response characteristics of each process section in the lithium carbonate production line process configuration file, the control perturbation within the preset micro-cycle feedback interval is extracted to generate a feedback modulation factor group; The feedback modulation factor group is matched with the process section structure table through channel mapping, and the control modulation response matrix is output.

8. The automatic control method for producing lithium carbonate according to claim 1, wherein Step S5 is specifically as follows: Step S51: Calculate the actual state change amplitude based on the response offset of each control channel in the control modulation response matrix and the process section field state factor set, calibrate the real-time control response offset value of the process section, and thus obtain a process control response offset mapping table; Step S52: Compare the process control response offset mapping table with the physical quantity-process section relationship table, locate the process section where the offset occurs and the associated physical quantity according to the process section number, and calculate the offset cause weight to obtain a process offset correlation factor group; Step S53: Based on the process offset correlation factor group, reverse tracing of the path is performed in the graph structure of the physical quantity-process section relationship table to obtain the offset impact path; Step S54: structurally adjust the offset impact path and generate a process map reverse adjustment index table; Step S55: Apply the process map reverse adjustment index table to the process section field status factor set, update the structured binding relationship, mark the abnormal response segment, and generate a process abnormality feedback body.

9. The automatic control method for producing lithium carbonate according to claim 8, characterized in that Step S54 is specifically as follows: Step S541: extracting the control correlation, process effect intensity, and conduction direction of each physical quantity node in the offset impact path, and constructing an offset path structure factor matrix; Step S542: analyzing the influence intensity of the control response offset on each physical quantity node according to the offset path structure factor matrix, and marking the path segments with an influence intensity exceeding [0.15, 0.25] to obtain a path abnormal segment identification set; Step S543: performing node replacement, weight reconstruction, and direction optimization on the path segments in the path abnormality segment identification set, and recording the structure change operations to obtain a path structure change record; Step S544: index and hierarchically encode the path structure change records according to the process section number and physical quantity identifier to obtain a process map reverse adjustment index table.

10. An automatic control system for producing lithium carbonate, characterized in that: For executing the automatic control method for producing lithium carbonate according to claim 1, the automatic control system for producing lithium carbonate comprises: The field status recognition module is used to obtain the sensor data of the lithium carbonate production line and perform structured binding of the physical quantities of the process sections of the lithium carbonate production line sensor data to obtain the field status factor set of the process sections; The multi-step prediction module is used to extract multi-step prediction vectors based on the process section field state factor set, and fine-tune the multi-step prediction vectors to embed time state migration to obtain a multi-dimensional control representation flow; The control simulation module is used to perform digital twin control simulation based on the multi-dimensional control representation flow, and encode the control instructions of the digital twin control simulation results to obtain a multi-channel control factor set; The production line control module is used to transmit the multi-channel control factor set to the production line management platform to execute the production line control task; it combines the acquired real-time lithium carbonate production line sensor data to perform on-site load parameter recalculation and micro-cycle feedback modulation to obtain the control modulation response matrix; The process deviation optimization module is used to trace the process deviation based on the control modulation response matrix, and to reversely adjust the process diagram structure of the process section field status factor set according to the process deviation traceability results to obtain the process abnormality feedback body; The iterative parameter adjustment module is used to use the process abnormality feedback body as the input of the multi-step prediction module for a new round, and repeatedly iterate the multi-step prediction module to the process offset optimization module until the process offset is less than the preset process offset threshold.

Images