Electrode material preparation process control method and system

Abstract

The invention relates to the technical field of electrode material preparation, and discloses an electrode material preparation process control method and system, and the method comprises the following steps: S1, in an electrode material sintering process, collecting multi-modal sensing data in real time, including reaction tail gas component data, material weight data and reactor temperature data; s2, integrating the multi-modal sensing data, and performing combined processing on the collected data and corresponding time nodes to generate a reaction process index; s3, inputting the reaction process index into a pre-trained hybrid prediction model; s4, based on the deviation between the prediction result and the target value, generating a control instruction through a multi-target optimization algorithm; and S5, monitoring the performance index of the adjusted electrode material in real time. According to the electrode material preparation process control method and system, the stability and consistency of electrode material preparation forming are improved, various interferences are sensed and compensated in real time, and it is ensured that key indexes of each batch of products are stabilized within the qualified range.

Description

Technical Field

[0001] This invention relates to the field of electrode material preparation technology, specifically to a method and system for controlling the electrode material preparation process. Background Technology

[0002] With the development of the new energy industry, the application of new energy batteries is becoming more and more widespread, especially with the increasing demand in the existing electric vehicle field and energy storage equipment. The performance difference of new energy batteries largely depends on the quality difference of electrode materials. For example, the difference in thickness uniformity of electrode materials during the preparation process and the effect of electrode thickness control during sintering and shaping will directly affect the charging and discharging efficiency and safety of the battery in actual use.

[0003] In existing electrode material preparation and processing technologies, the preparation and control of electrode materials are generally achieved through a rolling process, i.e., adjusting the gap between adjacent rolling mills. The electrode material is then shaped through sintering or other treatments. Most existing control systems use preset, fixed time-temperature curves for control. This "open-loop" or "weakly closed-loop" control mode relies on the core assumption that production process conditions and material properties remain absolutely stable. However, in actual production, factors such as slight differences between batches of raw materials, the natural non-uniformity of the heat field distribution within the sintering furnace, the slow performance degradation of heating elements and insulation materials, and fluctuations in protective gas flow are unavoidable. A fixed control program... The inability to perceive and respond to these inherent dynamic changes leads to significant fluctuations in product quality (such as the specific capacity and rate performance of materials) between different batches, making it difficult to guarantee consistency in dynamic control. As production time accumulates, equipment inevitably experiences performance degradation (such as a decrease in heater efficiency), or physical property parameters may change when raw material suppliers are changed. At this point, the original control parameters may no longer be optimal, requiring engineers to perform tedious manual readjustment and process verification. This limitation greatly affects the stability of the formed electrode materials, making it difficult to guarantee the consistency of electrode materials produced on the electrode processing production line, and affecting the continuous and efficient preparation and processing of electrode materials.

[0004] To address the aforementioned issues, there is an urgent need for innovative designs based on existing methods for controlling the electrode material preparation process. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for controlling the electrode material preparation process, in order to solve the problem mentioned in the background art that the existing electrode material preparation process control adopts manual-assisted static adjustment control, which cannot effectively guarantee the stability of the produced electrode materials and the consistency of the finished products during long-term use, resulting in low electrode material production capacity in continuous high-intensity working environments and affecting the stability of the finished product quality on the electrode material production line.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for controlling the preparation process of electrode materials, the method comprising the following steps: S1: During the sintering process of electrode materials, multimodal sensing data is collected in real time, including reaction tail gas composition data, material weight data and reactor temperature data; S2: Integrate the multimodal sensing data, combine the collected data with the corresponding time nodes, and generate a reaction process index; S3: Input the reaction process index into a pre-trained hybrid prediction model to predict the product quality indicators and process energy efficiency indicators of the product formation. S4: Based on the deviation between the predicted results and the target values, control commands are generated through a multi-objective optimization algorithm to dynamically adjust the sintering process parameters; S5: Monitor the performance indicators of the adjusted electrode material in real time. If the deviation between the performance indicators and the target range exceeds the preset threshold, start working again from S1 to perform a new round of optimization cycle until the performance indicators are qualified and output the final control parameter set.

[0007] Preferably, in the S1 stage, the concentration, weight, and temperature of the exhaust gas during electrode material sintering are time-series aligned and fused to form a feature vector of multimodal sensing data, thereby determining the changes in the reaction stages during electrode material sintering. The specific multimodal sensing data includes the following: Real-time monitoring of the composition data of the reaction tail gas, and changes in the concentration of characteristic gases during the sintering reaction of electrode materials. Material weight data is used to monitor the weight loss changes of electrode materials during the sintering process in real time. The reactor temperature data is used to monitor the comprehensive thermal field distribution information at different points inside the reactor during the sintering reaction of the detection electrode material in real time.

[0008] Preferably, step S2 includes: S21: Use a temporal convolutional network to extract the local spatiotemporal features of the multimodal sensing data, and process the industrial time-series data of the changes in multimodal sensing data during the sintering reaction stage of the electrode material; S22: Use Transformer The encoder models the global dependencies of the local spatiotemporal features and outputs a global feature vector representing the current state of the process. This vector can be used as input for subsequent models (such as prediction models) to infer future trends. S23: Weighted fusion of local spatiotemporal features and global features to generate a comprehensive reaction process index.

[0009] Preferably, the hybrid prediction model in S3 is set as follows: TCN-Transformer Hybrid network architecture, specifically including the following: TCN Components used to capture short-term local fluctuations in sensor data; Transformer Components used to learn long-term dependencies in the process; The hybrid prediction model simultaneously predicts product quality indicators, process energy efficiency indicators, and process safety indicators, enabling a single model to perform multiple prediction tasks at the same time.

[0010] Preferably, the setting of the hybrid prediction model in S3 further includes a management scheme for model versions, as follows: S301: First, create an independently stored model version for each model update and record version metadata. The initial model consists of experiments under multiple typical process conditions, collected multimodal process data, and performance test data of the final product. S302: When the prediction error continues to exceed the standard, that is, when it exceeds the initial preset threshold, the current working condition data will be matched and analyzed with the historical model version. S303: Select the best historical model as the basis for incremental learning to accelerate model convergence.

[0011] Preferably, in S4: A multi-objective optimization function is constructed with the goals of achieving optimal product quality and minimum energy consumption. In the optimization process, the multi-objective optimization function formalizes the competing or even conflicting optimization objectives, providing clear guidance for the algorithm to solve the problem. The Pareto optimization algorithm is used to solve for the optimal combination of sintering process parameters; under the premise of meeting product quality constraints, a multi-objective balance between energy efficiency and quality is achieved, and the final executable solution is selected from the Pareto solution set.

[0012] Preferably, the monitoring and optimization cycle steps for the electrode material performance indicators in S5 are as follows: S51: Perform online performance testing on the produced electrode materials, including specific surface area and electrochemical capacity indicators; S52: Establish a mapping model between online performance testing indicators and process parameters to determine product performance under the current process parameters, providing a mathematical basis for reverse optimization; S53: Based on the mapping relationship model, reverse optimization realizes closed-loop control from product quality to process parameters. When the detected product performance deviates from the target, the process parameters that need to be adjusted are automatically calculated in reverse to achieve accurate correction.

[0013] The present invention also provides a control system for electrode material preparation process, which includes the following: The multimodal data acquisition module is used to collect real-time data on the composition of the reaction tail gas, the weight of the materials, and the temperature of the reactor. The feature fusion module is used to fuse multimodal data with corresponding acquisition time nodes to generate a reaction process index; The quality prediction module is used to use a hybrid prediction model to predict product quality and energy efficiency indicators. The intelligent decision-making module is used to generate control commands through multi-objective optimization and dynamically adjust process parameters; The closed-loop control module is used for product performance feedback to achieve continuous optimization of process parameters.

[0014] Furthermore, the system also includes a model management module for the hybrid prediction model, which includes the following: A model version repository is used to independently store different versions of production models; A version comparison engine is used to match and analyze real-time data with historical models; The intelligent rollback mechanism facilitates switching to a stable version when model performance degrades, and allows for the use of different models for system control processing in electrode fabrication.

[0015] Compared with the prior art, the beneficial effects of the present invention are: the electrode material preparation process control method and system improve the stability and consistency of electrode material preparation and molding, especially in electrode material preparation under the conditions of long-term equipment wear and tear and changes in material application varieties. It can sense and compensate for various interferences in real time, ensuring that the key indicators of each batch of products are stable within the qualified range. By employing this systematic approach and employing multi-objective optimization, the system proactively seeks the optimal balance between quality and energy consumption, achieving energy-saving production while meeting standards. This directly reduces the production cost per batch. Simultaneously, it integrates operator experience into the intelligent algorithm. Even if raw material sources change or equipment performance slowly declines, the system can automatically adjust process parameters to adapt to new conditions through model self-learning and version management, maintaining optimal production results. This reduces reliance on skilled engineers in specific positions, alleviating personnel burden. Furthermore, it constructs a closed loop from production to quality inspection, continuously improving process levels and forming a continuous improvement cycle of "production-inspection-optimization-reproduction." The system feeds back the performance test results of the final product to the control center, using this data to continuously revise and optimize predictive models and control strategies. This endows the entire production process with self-learning and evolutionary capabilities, improving equipment efficiency and adaptability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the operation flow of the method of the present invention; Figure 2 This is a schematic diagram of the method flow for model management according to the present invention; Figure 3 This is a schematic diagram of the monitoring and optimization cycle of the electrode material performance indicators of the present invention; Figure 4 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: This invention provides a technical solution: a method for controlling the electrode material preparation process, comprising the following steps: S1: During the sintering process of electrode materials, multimodal sensing data is collected in real time, including reaction tail gas composition data, material weight data and reactor temperature data; The acquisition of reaction tail gas composition data is crucial, as tail gas composition is one of the most direct indicators reflecting the chemical reaction process within the sintering furnace. Specific gaseous products are released at different reaction stages, and their concentration and timing are key to determining the reaction progress. The acquisition equipment includes a line mass spectrometer or a miniature gas chromatograph installed on the exhaust pipe of the sintering furnace. Both of these devices can achieve real-time, continuous gas composition analysis. The target of acquisition is the characteristic gases related to key reactions such as the decomposition, carbonization, and crystallization of electrode material precursors; for example, in lithium-ion battery cathode materials (such as…). During the carbothermic reduction sintering process, CO and The concentration changes directly reflect the reaction status of the carbon source and the degree of oxygen removal. In the preparation of porous carbon materials, monitoring the escape rate of organic cracking products such as methane and hydrogen, the inflection point of the characteristic gas concentration often marks the beginning or end of a specific reaction stage, thus directly determining the reaction stage in which the electrode material sintering is performed. The acquisition of reaction tail gas composition data is simultaneously output, with the equipment outputting time-series data, including the concentration value of each target gas (usually in units of...). ppm or % The system uses a set sampling frequency (e.g., per 1000 m / s) and the device's own status parameters (such as signal-to-noise ratio) to sample the data. 10s or each 1min (A data point) records these data.

[0019] Meanwhile, the aforementioned material weight data collection, due to the weight changes of the electrode material during sintering, directly reflects the material losses caused by physicochemical processes such as solvent evaporation, binder decomposition, and precursor transformation, serving as a macroscopic quantitative indicator of reaction completion. Therefore, simultaneously with the electrode material sintering equipment, a high-temperature resistant, high-precision weighing sensor module is used. Existing systems typically isolate the weighing sensor module from the high-temperature zone using special mechanical structures (such as support rods and cooling systems made of high-temperature resistant materials), transmitting only mechanical force to the weighing sensor module. During continuous sintering, the weighing sensor is monitored at a preset constant frequency (e.g., every...). 1s The system reads material weight data in real time. The acquisition of this weight data requires data processing. The raw weight signal contains high-frequency noise (such as noise caused by airflow disturbances or mechanical vibrations within the furnace). Therefore, digital filtering (such as using a low-pass filter or moving average algorithm) is necessary to smooth the data and extract the true weight change trend. Furthermore, the system calculates the weight change rate (i.e., weight loss per unit time) in real time. This derivative signal more sensitively indicates the inflection point of the reaction rate and is more instructive than the absolute weight value.

[0020] Regarding the acquisition of reactor temperature data, temperature is the most critical control parameter in the sintering process, and its field distribution uniformity has a decisive impact on the consistency of the final product. However, relying solely on a single temperature measurement point cannot fully reflect the true thermal environment of the material. Therefore, a temperature field monitoring array (at least three measurement points) can be formed by arranging multiple thermocouples at key locations inside the reactor (sintering furnace). For example, thermocouples can be installed on the reactor inner wall to monitor the furnace set temperature; thermocouples can be installed on the material surface to monitor the ambient temperature directly felt by the material, as close as possible to the material surface without contact; and insertion thermocouples can be installed inside random materials from the same batch for detection, directly monitoring the internal ambient temperature of the material during sintering using random sampling. All thermocouples are connected to the control system through a temperature acquisition module at a high frequency (e.g., every...). 1s Temperature data is collected synchronously multiple times, and it is necessary to maintain the synchronization of multiple temperature data. The synchronization requires the system to ensure that all temperature data points have a unified and accurate timestamp so that accurate correlation analysis can be performed later.

[0021] In the above scheme, the concentration, weight, and temperature of the tail gas during electrode material sintering are time-series aligned and fused to form a feature vector of multimodal sensing data, which determines the changes in the reaction stage during electrode material sintering. The specific multimodal sensing data includes the following: reaction tail gas composition data, which detects the changes in the concentration of characteristic gases during the electrode material sintering reaction stage in real time; material weight data, which monitors the changes in weight loss during the electrode material sintering process in real time; and reactor temperature data, which monitors the comprehensive thermal field distribution information at different points in the reactor during the electrode material sintering reaction in real time. The preprocessing and synchronous fusion of this multimodal data, after obtaining the three original data streams (reaction tail gas composition data, material weight data, and reactor temperature data), must be preprocessed and synchronized to form effective fusion features. This involves data preprocessing, outlier handling, and using statistical methods (such as the Laida criterion) or rules based on the rate of change to identify and remove abnormal data points caused by sensor malfunctions or interference. Since tail gas concentration, weight, and temperature data have different dimensions and numerical ranges, they need to be normalized to scale them to similar numerical ranges (such as [0,1] or a mean of 0 and a variance of 1) for easier subsequent model processing. Time synchronization and alignment are also crucial; since the sampling frequencies and response times of different sensors may vary, the system establishes a data buffer and alignment mechanism. Using a unified clock as a reference, the system... All sensor data are interpolated (or aggregated) to the same time point to generate a multimodal data frame with a unified timestamp; for example, a data frame is generated every 10 seconds, which includes the exhaust gas concentration, material weight, and temperature values ​​at multiple locations at that moment. Preliminary feature extraction and data fusion can be performed on the aligned data frames to prepare for subsequent steps, such as: calculating the real-time temperature difference between the core temperature of the material and the set temperature of the furnace; calculating the variance of temperature readings at multiple locations to quantify the uniformity of the thermal field inside the furnace; combining weight data with exhaust gas concentration data to calculate the amount of exhaust gas released per unit weight loss. The originally isolated multi-sensor signals are integrated into a high-dimensional data stream that can comprehensively, in real time, and accurately characterize the internal physicochemical state of the electrode material sintering process, providing a guarantee for the operation of subsequent system methods.

[0022] S2: Integrate the multimodal sensing data, combine the collected data with the corresponding time nodes, and generate a reaction process index; The above scheme combines the pre-processed multimodal sensing data (exhaust gas concentration, material weight, temperature field) with the corresponding time points of reaction changes to generate a reaction progress index. This allows for direct and effective judgment outside the sintering furnace to determine the current stage of motor material sintering preparation and to clarify the progress of the chemical reaction.

[0023] S21: Use a temporal convolutional network to extract the local spatiotemporal features of the multimodal sensing data, and process the industrial time-series data of the changes in multimodal sensing data during the sintering reaction stage of the electrode material; In the above, the Temporal Convolutional Network (TCN) component acts as a "high-sensitivity sensor" in the entire feature extraction process, specifically designed to capture localized, coordinated change patterns that occur within a short timeframe during the sintering process, triggered by specific physicochemical reactions. Since the sintering process of electrode materials is essentially a continuous, phased evolution, this evolution is not uniform and stable, but rather consists of a series of continuous, overlapping local reaction events (such as solvent evaporation, binder decomposition, precursor carbonization and crystallization). Each local event leaves a unique "distinctive" signal on multiple sensing channels. For example, the occurrence of an exothermic reaction may simultaneously cause a sudden surge in the reactor core temperature, accompanied by specific gases (such as...). CO The release peak of the signal and the accelerated loss of material weight; these signals are highly synchronized in time and spatially correlated. Temporal convolutional networks are chosen here because of their unique architectural advantages, making them very suitable for processing this type of industrial time-series data, as follows: TCN The core of this is dilated causal convolution, which gives the network the ability to capture dependencies at different time scales, even for a one-dimensional input sequence. (in T It is the sequence length. d It is the feature dimension, for example d=3 (representing three characteristics: exhaust gas concentration, weight, and temperature) at time step t At this point, the output of the dilated causal convolution yt The calculation is as follows: ,in yt : at time step t The output characteristics; wi Trainable convolutional kernel weights, size k ; At time step Input data; d : Inflation factor, when d=1 At this time, it's just a regular convolution, with the receptive field being... k ;when d by 2 The power of the increase (e.g.) 1,2,4,8... At this time, the receptive field increases exponentially. (k×d) This allows for the capture of longer-term dependencies without significantly increasing parameters; indexing Ensured calculation yt At that time, only the current and past time step data are used. （i≥0） Never use future data (t + 1, t + 2,...) This is a fundamental requirement for real-time performance in industrial process control; Specifically, when calculating the corresponding local spatiotemporal features, for example, assuming we collect data at a frequency of 1 point per second, we have a 3-dimensional input sequence (exhaust gas concentration, weight, temperature) with a length of 10 seconds. （T=10） We use a k=3 The convolution kernel; Capturing instantaneous events （d=1） The convolution kernel scans data from the past 3 seconds. For example, calculating... t=5 The characteristics of time: y5 =W 0 X 5 +W 1 X 4 +W 2 X 3 This allows TCN to learn things such as "in t=3 arrive t=5 Within those three seconds, does the exhaust gas concentration and temperature rise sharply simultaneously, exhibiting a localized coordinated pattern? Capturing short-term trends ( d=2 With an inflation factor of 2, the receptive field of the convolution kernel expands to 5 time points. Calculate the features at time t=5: y5 = W 0 X 5 +W 1 X 3 +W 2 X 1 This makes TCN It can learn patterns spanning longer periods (e.g., from t=1 to t=5), such as short-term trends like "weight steadily decreasing from the first second while exhaust gas concentration begins to increase after the third second"; By stacking multiple objects with different expansion factors d of TCN Layers, the network can eventually achieve each time step t Output a local spatiotemporal feature vector h t TCN This vector encodes the... t The collaborative change patterns of multi-sensor data at different time scales, centered on the moment.

[0024] S22: Use TransformerThe encoder models the global dependencies of the local spatiotemporal features and outputs a global feature vector representing the current state of the process. This vector can be used as input for subsequent models (such as hybrid prediction models) to infer future trends; specifically as follows: Transformer The core of it is the self-attention mechanism, which can calculate the strength of the relationship between any two time points in a sequence; its core formula is scaled dot product attention. first, TCN Output feature sequence H TCN Linearly projected as a query (Query) ,key (Key) ,value( Value) Three matrices: Q = H TCN W Q ; K = H TCN W K ; V = H TCN W V ;in W Q , W K , W V It is a trainable weight matrix, then the self-attention is calculated: ; QK T : Calculate one T×T Attention score matrix; each element of this matrix score ij Indicates time step i Query (Query) With time step j The key (Key) The similarity directly quantifies the time step. i With time step j The strength of the dependency relationship; A scaling factor is used to prevent the gradient from vanishing due to an excessively large dot product. softmax Attention scores are normalized into weights, representing the degree to which each other time point in the sequence should be "attended" when calculating the output at a given time point; the final output is a weighted average of these weights. V The sum of matrices.

[0025] How to calculate global dependencies, as follows: Assume our TCN The output sequence contains features from 4 time steps.[h 1 ,h 2 ,h 3 ,h 4 ] These correspond to the four stages of the sintering process; we need to understand the global dependency between the fourth time step (the current state) and all previous time steps; in the calculation h 4 When the corresponding output is received, the self-attention mechanism will calculate... h 4 of Query and h 1 ,h 2 ,h 3 ,h 4 of Key Similarity; Suppose we obtain the attention weights as [0.1,0.2,0.1,0.6] This indicates an understanding of the current state. h 4 Most importantly, it has its own information (weight). 0.6 Secondly, the most important thing is... h2 Information at any given time (weight) 0.2 This means that in the early stages of sintering (corresponding to...) h 2 A specific warming event, for the current （h 4 ） The crystallization state of the material has a crucial influence; this causal relationship spanning long time scales has been captured. Multi-head attention mechanisms execute the above process in parallel multiple times (multiple "heads"), with each head using a different projection moment. Matrix W This allows the model to simultaneously learn different types of dependencies from different subspaces (such as the "temperature-exhaust gas" relationship subspace and the "weight-energy consumption" relationship subspace). Finally, the outputs of all heads are concatenated and passed through a feedforward neural network to generate the final global feature vector. h t Trans .

[0026] S23: Weighted fusion of local spatiotemporal features and global features to generate a comprehensive reaction process index.

[0027] In this step, TCN Dilated causal convolution; focusing on capturing the coordinated change patterns among multiple sensor signals in a short period of time to identify specific physicochemical reaction events; Transformer The self-attention mechanism, based on TCN A series of extracted local features are analyzed to determine the long-term dependencies and causal relationships between different events throughout the process history, thus forming a feedback understanding of the process status. in TCN The input data is the original multimodal time series; an example is 10 consecutive time points ( t 1 arrive t 10 The sampled data contains measurements in three dimensions at each time point: CO 2 Concentration: Reflects the release of gaseous products in a specific chemical reaction (such as decomposition); Material weight: Directly reflects the mass loss of material due to volatilization, decomposition, and other reactions; Material temperature: Reflects the thermodynamic conditions under which the reaction proceeds; Calculation process and output: Sliding window scan: TCN Like a fixed-size "time window," it slides over the input sequence; for example, the window might cover... t 5 arrive t 7 Data at any given time; Local pattern recognition: The model learns and identifies the collaborative patterns of multi-dimensional data changes within a window; such as when... t 5 arrive t 7 The window recognized " When the concentration curve shows a sharp peak and the weight curve shows a steep drop, TCN can abstract this into a feature; Feature abstraction: TCN The output is no longer the original data, but a local spatiotemporal feature vector (e.g., );vector It was encoded as "in" t 5 The information is that "a strong venting and weightlessness reaction event was detected near the time"; the output of this step is to convert the raw sensor readings into a series of characteristic "process event" features.

[0028] Transformer, Self-attention mechanism; this mechanism allows the model to calculate the correlation strength between any two "event" features in a sequence, regardless of how far apart they are in time; Input data: TCN The output local feature sequence, i.e. Each vector in this sequence represents an abstract feature of a local reaction event occurring at that moment; Calculation process and output: Global correlation analysis: TransformerIt receives the entire feature sequence and calculates the correlation between all feature pairs using a self-attention mechanism; for example, it analyzes "reaction peak events". Compared to earlier "warming rate events" The correlation between them; key factor tracing: the model automatically determines which historical events have the greatest impact on the current or final state by calculating "attention weights"; in analyzing the final state At that time, the model may find (Initial temperature rise) and (Reaction peak) received the highest attention weight; establishing global dependency: this means the model learned a process knowledge: the quality of the final product ( t 10 The state is not determined solely by the parameters at the final moment, but is critically influenced by the early warming regime. t 2 ) and mid-term core reaction intensity ( t 7 The combined influence and constraints of ) Transformer The output is a deeper feature representation that incorporates this global contextual information; TCN Level: It solves the problem of "what is happening", transforming continuous sensor signals into discrete, meaningful event markers; Transformer Level: It addresses the questions of "why this is happening" and "what will it lead to," revealing the causal and dependent relationships between events at different process stages.

[0029] S3: Input the reaction process index into a pre-trained hybrid prediction model to predict the product quality indicators and process energy efficiency indicators of the product formation; the hybrid prediction model in S3 is set as follows: TCN-Transformer Hybrid network architecture, specifically including the following: TCN Components used to capture short-term local fluctuations in sensor data; Transformer Components used to learn long-term dependencies in the process.

[0030] The setup of the hybrid forecasting model in S3 also includes a scheme for managing model versions, as follows: S301: First, create an independently stored model version for each model update and record version metadata. The initial model consists of experiments under multiple typical process conditions, collected multimodal process data, and performance test data of the final product. S302: When the prediction error continues to exceed the standard, that is, when it exceeds the initial preset threshold, the current working condition data will be matched and analyzed with the historical model version. S303: Select the best historical model as the basis for incremental learning to accelerate model convergence; The above solution establishes a "version archive" for the model and a "smart backtracking learning" workflow. Its core objective is to achieve smooth model iteration and safe rollback, ensuring the stability and reliability of the production system. In S301, each time the model is updated (e.g., after retraining with a new batch of data), the system does not simply overwrite the old model. Instead, it saves the new model as an independent file package and assigns it a unique version number (e.g., v1.0, v1.1, v2.0). Metadata is also recorded: a "model archive" is created, recording key metadata for that version, typically including: training data range: the time period of production data used for training; model performance metrics: key metrics for that version on the test set, such as prediction accuracy and root mean square error; applicable operating condition description: the range of process conditions corresponding to the model's optimal performance (e.g., "applicable to brand A raw materials, ambient temperature 25±5°C"); and the creation time and creator are also included for easy traceability. Its traceability: If problems occur after the new model goes live, it can be quickly and accurately rolled back to any historical stable version; auditability: It meets the requirements of strict process recording and quality auditing in industrial production; flexibility: It allows different dedicated models to be retained and called for different production formulas or raw material batches.

[0031] The S302 prediction error exceeding the limit triggers historical version matching analysis, which includes error monitoring: the system monitors the prediction error of the online model for the current production batch data in real time; when the error exceeds a preset threshold for multiple consecutive periods or the cumulative value (e.g., 15% higher than the historical average), an early warning is triggered; operating condition matching analysis: the system does not immediately use the currently problematic data to train the model, but instead initiates an analysis process: quickly matching the current real-time multimodal data features (e.g., the current temperature curve pattern, exhaust gas release pattern) with the operating condition features of all historical model training data in the version library (e.g., calculating the cosine similarity between feature vectors); finding the optimal historical version: the goal of the matching analysis is to find a historical model version that is most similar to the current operating condition; for example, the system... The system might discover that the volatility characteristics of the current raw material are very similar to a batch produced three months ago, on which the v1.2 model performed exceptionally well. This model version management setting avoids "catastrophic forgetting": if the model is retrained directly on the full dataset using the current data, new knowledge might overwrite old knowledge, causing the model to forget how to handle other previously learned operating conditions. Matching analysis ensures that the new learning process is conducted on a "most relevant experience" baseline. (This is repeated in the original text.)

[0032] S303 enables "safe and efficient evolution" of the model, seamlessly integrating new knowledge into existing experience. The system can load the "optimal historical model" selected in step S302 from the version control library. Instead of retraining with all historical data, it incrementally learns or fine-tunes the base model using only the current batch of data. This process updates only some of the model's parameters, allowing it to quickly adapt to new data distributions while maintaining its core knowledge. After incremental learning, a new model candidate version is generated. The new model is not immediately deployed but is first validated in a parallel "shadow environment" or offline test set. Only when the validation results show that its performance is significantly better than the current online version and not lower than other major historical versions will it be approved for formal deployment. This greatly shortens the time required for the model to adapt to new operating conditions. Incremental learning is a gentle update method, reducing the probability of introducing unknown risks during model updates.

[0033] S4: Based on the deviation between the predicted results and the target values, a multi-objective optimization algorithm is used to generate control commands and dynamically adjust the sintering process parameters. This resolves the inherent contradictions among multiple control objectives during sintering and automatically identifies the process parameters that achieve the optimal balance. In the preparation of electrode materials, there are generally two objectives: optimal product quality and minimum process energy consumption. However, these two objectives are often conflicting. To obtain a higher specific surface area, a longer holding time or a higher final temperature may be required, which inevitably leads to increased energy consumption. Conversely, shortening the process to save energy may sacrifice product quality. Therefore, S4 is used to mathematically formalize these competing objectives and construct a multi-objective optimization problem, as follows: A multi-objective optimization function is constructed with the goals of optimal product quality and minimum energy consumption. In the optimization process, the multi-objective optimization function formalizes the competing and even conflicting optimization objectives, providing guidance for the algorithm solution. The Pareto optimization algorithm is used to solve for the optimal combination of sintering process parameters. Under the premise of meeting product quality constraints, the multi-objective balance between energy efficiency and quality is achieved, and the final executable solution is selected from the Pareto solution set. The sintering process parameters that need to be adjusted are taken as decision variables and denoted as a vector. x ;For example, x It can include the heating rate, holding temperature, and holding time for each stage; quality objective function. f1(x) This is a model whose input is process parameters. x The output is a predicted product quality metric (such as specific surface area); the objective is to maximize... f1(x) Energy consumption objective function f2(x) It's the same model, but the input... x The output is the predicted total energy consumption of the process; the objective is to minimize...f2(x) Therefore, the optimization problem can be expressed as: (Note: Maximize) f1 Equivalent to minimizing - f1 ); Setting constraints and optimizing the process must be done within reasonable limits, that is, meeting the constraints of process feasibility and safety; for example, the temperature must not exceed the limit of the kiln refractory material. (Tmax) The key indicators of the final product must meet the requirements (e.g., purity > 99.5%); the process parameters themselves have upper and lower limits, which are preset manually. The Pareto optimization algorithm is used to find the optimal combination. For the multi-objective problem mentioned above, there is usually no single "unique solution" that is optimal for all objectives; instead, there exists a set of Pareto optimal solutions. The concept of Pareto optimality is: a solution... It is called Pareto optimal, meaning that there is no other solution. x It is superior to without compromising at least one other objective. The set of all Pareto optimal solutions is called the Pareto front. Each point on the front represents a specific trade-off between quality and energy consumption. This achieves multi-objective balance and final solution selection. After obtaining the Pareto solution set, the final step in step S4 is to select a final executable solution from multiple "optimal" solutions based on actual production needs. This is a decision-making process. Engineers can manually select the solution that best meets the requirements from the Pareto front as the final set point based on the current production strategy (such as "quality first," "energy efficiency first," or "balanced mode"). For example, to produce high-end products, the point with the highest quality but also slightly higher energy consumption can be selected; to fulfill ordinary orders, a point with lower energy consumption can be selected. Alternatively, automatic decision-making can be set, and the system can also preset decision rules; for example, defining a comprehensive utility function. ,in w 1 and w 2 The weights are set according to the production strategy; the system automatically calculates the comprehensive utility value of each solution on the Pareto front and selects the solution with the highest utility value as the final process parameter; step S4, by formalizing the multi-objective optimization problem, using advanced algorithms to solve the Pareto optimal solution set, and combining human-machine collaborative decision-making, successfully found the best balance between energy efficiency and quality under the hard constraint of ensuring product quality, and realized intelligent and refined control of the sintering process.

[0034] S5: Real-time monitoring of the adjusted electrode material performance indicators. If the deviation of the performance indicators from the target range exceeds the preset threshold, the process restarts from S1, performing a new round of optimization loops until the performance indicators are qualified, and outputs the final control parameter set. The monitoring and optimization loop steps of the electrode material performance indicators in S5 are as follows: S51: Online performance testing of the produced electrode materials, including specific surface area and electrochemical capacity indicators; S52: Establishing a mapping relationship model between online performance testing indicators and process parameters to determine the product performance under the current process parameters, providing a mathematical basis for reverse optimization; S53: Based on the reverse optimization of the mapping relationship model, closed-loop control from product quality to process parameters is achieved, so that when the detected product performance deviates from the target, the process parameters that need to be adjusted are automatically calculated in reverse to achieve precise correction. The final product quality inspection results are transformed into direct instructions to drive process optimization. At the sintering furnace outlet of the electrode material, a small amount (a few grams) of representative finished material is obtained through an automatic sampling device. The sample is automatically transported to an integrated rapid detection unit, which includes a physical adsorption analyzer and an electrochemical testing station. The detection data is automatically collected, processed, and stamped with batch numbers and timestamps, corresponding one-to-one with the process parameters in the production process database. Then, a mapping relationship model between online performance detection indicators and process parameters is established. The model inputs and outputs key process parameter sets, such as the heating rate of the sintering program, the holding temperature and time of each stage, and the gas flow rate of the reaction atmosphere. The outputs are the product quality indicators obtained by the S51 online detection, such as specific surface area and electrochemical capacity.

[0035] For modeling, since the process-performance relationship is usually nonlinear and high-dimensional, a data-driven machine learning algorithm is used to establish the mapping model F: Product Quality = F (process parameters). Alternatively, Gaussian process regression can be used. The algorithm can effectively handle complex nonlinear relationships and provide an estimate of the uncertainty of prediction. The model learns a mapping function from the process space to the performance space by training on a large amount of historical production data (containing the correspondence between "process parameters and product quality"). The model, given a set of process parameters, can predict the expected product quality. More importantly, it provides the mathematical foundation for S53 back-optimization, enabling the system to answer the question of how to set process parameters to achieve the target quality. Specifically, for back-optimization based on a mapping relationship model, closed-loop control is achieved, realizing a back-optimization problem where quality results drive automatic correction of process parameters: given a target product quality value (e.g., desired specific surface area)... The task of the optimization model is to find a set of optimal process parameters. This makes the predicted value F( of the mapping model) As close as possible Mathematically, this can be expressed as: The goal is to find the combination of process parameters that minimizes the variance between the predicted performance value and the target value, while satisfying all process feasibility constraints. Optimization algorithms such as gradient descent and genetic algorithms are used to solve this problem. The specific methods for implementing closed-loop control are as follows: Detection and Judgment: The system acquires online detection quality data of the current batch of products and compares it with the target value; Trigger optimization: As shown in S51-S53, if the quality indicator is not up to standard (deviation exceeds the threshold), the reverse optimization process will be triggered immediately. Calculation and Execution: The optimization algorithm, based on the mapping model F, quickly calculates the process parameters that should be adjusted. ); Application and Validation: The new set of process parameters is automatically downloaded to the sintering furnace control system to guide the next production batch; Continuous cycle: The next batch of output will be inspected again, thus forming a complete closed loop of "inspection-judgment-optimization-execution-re-inspection"; the system can continuously adapt, so that the product quality dynamically converges to the target value. Example 2: The present invention further provides a control system for the electrode material preparation process; the details are as follows: The multimodal data acquisition module is used to collect real-time data on the composition of reaction tail gas, material weight, and reactor temperature; the feature fusion module is used to fuse multimodal data with corresponding acquisition time points to generate a reaction progress index; the quality prediction module is used to use a hybrid prediction model to predict product quality and energy efficiency indicators; the intelligent decision-making module is used to generate control commands through multi-objective optimization and dynamically adjust process parameters; and the closed-loop control module is used for product performance feedback to achieve continuous optimization of process parameters.

[0036] The system also includes a model management module for hybrid prediction models, which includes the following: A model version repository is used to independently store different versions of production models; A version comparison engine is used to match and analyze real-time data with historical models; The intelligent rollback mechanism facilitates switching to a stable version when model performance degrades, and allows for the use of different models for system control processing in electrode fabrication.

[0037] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for controlling the preparation process of electrode materials, characterized in that, The method includes the following steps: S1: During the sintering process of electrode materials, multimodal sensing data is collected in real time, including reaction tail gas composition data, material weight data and reactor temperature data; S2: Integrate the multimodal sensing data, combine the collected data with the corresponding time nodes, and generate a reaction process index; S3: Input the reaction process index into a pre-trained hybrid prediction model to predict the product quality indicators and process energy efficiency indicators of the product formation. S4: Based on the deviation between the predicted results and the target values, control commands are generated through a multi-objective optimization algorithm to dynamically adjust the sintering process parameters; S5: Monitor the performance indicators of the adjusted electrode material in real time. If the deviation between the performance indicators and the target range exceeds the preset threshold, start working again from S1 to carry out a new round of optimization cycle until the performance indicators are qualified and output the final control parameter set.

2. The method for controlling the electrode material preparation process according to claim 1, characterized in that: In stage S1, the concentration, weight, and temperature of the exhaust gas during electrode material sintering are time-series aligned and fused to form a feature vector of multimodal sensing data, which determines the changes in the reaction stages during electrode material sintering. The specific multimodal sensing data includes the following: Real-time monitoring of the composition data of the reaction tail gas, and changes in the concentration of characteristic gases during the sintering reaction of electrode materials. Material weight data is used to monitor the weight loss changes of electrode materials during the sintering process in real time. The reactor temperature data is used to monitor the comprehensive thermal field distribution information at different points inside the reactor during the sintering reaction of the detection electrode material in real time.

3. The method for controlling the electrode material preparation process according to claim 1, characterized in that: Step S2 includes: S21: Use a temporal convolutional network to extract the local spatiotemporal features of the multimodal sensing data, and process the industrial time-series data of the changes in multimodal sensing data during the sintering reaction stage of the electrode material; S22: Use Transformer The encoder models the global dependencies of the local spatiotemporal features and outputs a global feature vector representing the current state of the process. This vector is used as input for subsequent hybrid prediction models to infer the future trend of electrode material forming. S23: Weighted fusion of local spatiotemporal features and global features to generate a comprehensive reaction process index.

4. The method for controlling the electrode material preparation process according to claim 1, characterized in that: The hybrid prediction model in S3 is set as follows: TCN-Transformer Hybrid network architecture, specifically including the following: TCN Components used to capture short-term local fluctuations in sensor data; Transformer Components used to learn long-term dependencies in the process.

5. The method for controlling the electrode material preparation process according to claim 1, characterized in that: The configuration of the hybrid prediction model in S3 further includes a management scheme for model versions, as follows: S301: First, create an independently stored model version for each model update and record version metadata. The initial model consists of experiments under multiple typical process conditions, collected multimodal process data, and performance test data of the final product. S302: When the prediction error continues to exceed the standard, that is, when it exceeds the initial preset threshold, the current working condition data will be matched and analyzed with the historical model version. S303: Select the best historical model as the basis for incremental learning to accelerate model convergence.

6. The method for controlling the electrode material preparation process according to claim 1, characterized in that: In S4: A multi-objective optimization function is constructed with the goals of optimal product quality and minimum energy consumption. In the optimization process, the multi-objective optimization function formalizes the competing or even conflicting optimization objectives, providing guidance for the algorithm solution. The Pareto optimization algorithm is used to solve for the optimal combination of sintering process parameters; under the premise of meeting product quality constraints, a multi-objective balance between energy efficiency and quality is achieved, and the final executable solution is selected from the Pareto solution set.

7. The method for controlling the electrode material preparation process according to claim 1, characterized in that: The monitoring and optimization cycle steps for the electrode material performance indicators in S5 are as follows: S51: Perform online performance testing on the produced electrode materials, including specific surface area and electrochemical capacity indicators; S52: Establish a mapping model between online performance testing indicators and process parameters to determine product performance under the current process parameters, providing a mathematical basis for reverse optimization; S53: Based on the mapping relationship model, reverse optimization realizes closed-loop control from product quality to process parameters. When the detected product performance deviates from the target, the process parameters that need to be adjusted are automatically calculated in reverse to achieve accurate correction.

8. A control system for an electrode material preparation process, applied to the electrode material preparation process control method according to any one of claims 1-7, characterized in that, The system includes the following: The multimodal data acquisition module is used to collect real-time data on the composition of the reaction tail gas, the weight of the materials, and the temperature of the reactor. The feature fusion module is used to fuse multimodal data with corresponding acquisition time nodes to generate a reaction process index; The quality prediction module is used to use a hybrid prediction model to predict product quality and energy efficiency indicators. The intelligent decision-making module is used to generate control commands through multi-objective optimization and dynamically adjust process parameters; The closed-loop control module is used for product performance feedback to achieve continuous optimization of process parameters.

9. The electrode material preparation process control system according to claim 8, characterized in that: The system also includes a model management module for the hybrid prediction model, which includes the following: A model version repository is used to independently store different versions of production models; A version comparison engine is used to match and analyze real-time data with historical models; The intelligent rollback mechanism facilitates switching to a stable version when model performance degrades, and allows for the use of different models for system control processing in electrode fabrication.

Images