Method for homogenizing the coating surface density of NCM-Li battery positive electrode sheet

Abstract

The application provides a kind of NCM-Li battery positive pole piece coating surface density equalization method, belongs to NCM-Li battery positive pole piece preparation technical field.The application improves the control accuracy of coating surface density by surface density detection correction and double-parameter collaborative regulation, extracts short-term pulsation characteristics and long-term drift characteristics by different expansion rate time series convolution network, so that the system can accurately distinguish the nature and source of disturbance;On this basis, the response strength of each feature dimension is dynamically adjusted according to the current slurry viscosity state;In addition, the application also establishes a double-parameter collaborative regulation mechanism, finally combines the revised surface density deviation information to make a joint decision, outputs the collaborative parameter adjustment instruction, overcomes the effect of mutual cancellation or nonlinear amplification caused by single-parameter independent adjustment, so that the control system can maintain stable control accuracy under the scene of working condition switching, slurry batch replacement and the like.

Description

Technical Field

[0001] This invention belongs to the field of NCM-Li battery positive electrode preparation technology, specifically relating to a method for homogenizing the coating surface density of NCM-Li battery positive electrode. Background Technology

[0002] NCM-Li (nickel-cobalt-manganese) batteries are among the most widely used high-energy-density lithium-ion battery systems in the new energy vehicle field. Their cathode active material specific capacity can reach 180-220 mAh / g, higher than conventional cathode systems such as lithium iron phosphate. Therefore, given the greater capacity contribution per unit areal density, even small deviations in areal density can cause fluctuations in single-cell capacity. Furthermore, NCM cathodes typically employ a high compaction density design of 3.0-3.6 g / cm³. Uneven areal density further amplifies differences in electrode thickness and internal resistance after compaction, ultimately affecting the cycle life and safety performance of the battery pack. Therefore, achieving precise and uniform control of coating areal density is of paramount importance for ensuring the consistency and quality stability of NCM battery products.

[0003] The coating preparation of positive electrode sheets typically employs a slot extrusion coating process: after stirring and preparing the slurry, it is pressurized and delivered to the coating nozzle by a pump. The amount of slurry extruded is controlled by the gap between the nozzle and the substrate, forming a continuous wet coating on the current collector substrate. This coating is then dried and rolled to form the electrode sheet. In this process, pump speed and nozzle gap are two core control parameters determining the coating surface density: pump speed directly determines the amount of slurry delivered per unit time, while nozzle gap determines the thickness distribution of the extruded slurry. Ideally, precise control of surface density can be achieved by adjusting these two parameters. However, in actual production scenarios, due to fluctuations in slurry rheological properties, differences in equipment dynamic characteristics, and the hydrodynamic coupling effect between the two parameters, closed-loop control of surface density faces numerous technical challenges.

[0004] There is a systematic deviation between the feedback value of the β-ray areal density detection system and the actual dry matter areal density, a problem particularly prominent in NCM cathode systems. NCM active materials are mostly secondary spherical aggregates, formed by the aggregation of nanoscale primary particles. After coating, these secondary particles undergo varying degrees of deformation and localized breakage during rolling, resulting in a more complex distribution and variation of coating porosity compared to single-crystal or regularly shaped cathode materials. Furthermore, residual lithium compounds (LiOH, Li2CO3, etc.) on the surface of high-nickel NCM particles readily react with PVDF binders, causing the slurry viscosity to continuously increase and even gel during storage. Simultaneously, NCM slurries are high-solids-content non-Newtonian fluids with significant shear-thinning behavior. Changes in pump speed during coating alter the shear rate at the nozzle, leading to nonlinear changes in the apparent viscosity of the slurry. This viscosity fluctuation's impact on areal density is even more pronounced in NCM systems.

[0005] To address the aforementioned issues, recent studies have introduced deep learning technology, utilizing neural networks to model historical process data and predict areal density trends. However, existing methods still suffer from the following shortcomings: First, most methods directly use the equivalent value measured by β-rays as the true value to train the model, without considering the mapping error caused by porosity fluctuations, thus introducing systematic bias into the model itself. Second, there is a lack of effective separation and identification mechanisms for the multi-timescale disturbances aliased in the areal density time series signal, making it difficult to distinguish between short-term random fluctuations and long-term trend drifts, resulting in insufficient stability of model prediction accuracy under complex operating conditions. Third, control strategies often employ single-parameter optimization or simple parameter lookup methods, failing to establish a quantitative description of the hydrodynamic coupling relationship between pump speed and nozzle gap, thus failing to achieve coordinated control under multi-parameter coupling conditions, resulting in poor control performance in scenarios such as operating condition switching or slurry batch replacement. Summary of the Invention

[0006] To address the problems existing in the background technology, the purpose of this invention is to provide a method for homogenizing the surface density of NCM-Li battery positive electrode coating, so as to solve the problem of poor surface density control of NCM-Li battery positive electrode coating caused by the distortion of equivalent measurement values ​​and the strong hydrodynamic coupling relationship between pump speed and nozzle gap in traditional deep learning methods.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for homogenizing the surface density of coating on the positive electrode sheet of an NCM-Li battery includes the following steps:

[0009] Step 1: Collect slurry viscosity data; collect time-series data of the areal density equivalent value of the battery positive electrode sheet and real-time process parameter data of the coating machine, and then perform preprocessing to obtain the preprocessed areal density equivalent value time-series data and the preprocessed process parameter time-series data.

[0010] Step 2: Extract multi-scale fused temporal features based on the preprocessed areal density equivalent value time series data; extract viscosity-modulated features based on slurry viscosity data through a viscosity modulation network; and then correct the areal density equivalent value through a porosity compensation neural network to obtain the corrected areal density.

[0011] Step 3: Based on the preprocessed process parameter time series data, extract the pump speed time series features and nozzle gap time series features respectively, and then fuse them to obtain the joint coding features of process parameters; calculate the two-parameter interaction weight matrix through the cross attention mechanism, and then calculate the parameter-area density coupling features.

[0012] Step 4: Based on the parameter-area density coupling characteristics, extract the dual-parameter joint characteristics, and then combine them with the corrected area density to calculate the dual-parameter joint adjustment characteristics, pump speed command value, and nozzle gap command value in sequence.

[0013] Furthermore, the specific process of step 1 is as follows:

[0014] Time-series data of the areal density equivalent value of the positive electrode sheet of the battery were collected using a beta-ray thickness gauge. The data type is one-dimensional time-series numerical data, which includes the real-time equivalent areal density values ​​of multiple measurement points uniformly distributed along the coating direction of the electrode sheet and their corresponding timestamps. Outliers were removed from the areal density equivalent value time-series data using the three-standard-deviation principle to obtain the noise-reduced areal density equivalent value time-series data.

[0015] Collect real-time process parameter data of the coating machine; the data type is multidimensional time series numerical data.

[0016] The viscosity data of the slurry was measured using a rotational viscometer, and the data type was numerical.

[0017] The Z-score standardization method was used to eliminate dimensional differences in the time series data of the denoised areal density equivalent value and the real-time process parameter data of the coating machine, so as to obtain the pre-processed areal density equivalent value time series data and the pre-processed process parameter time series data.

[0018] Furthermore, the real-time process parameters of the coating machine include pump speed data and nozzle gap data; each parameter corresponds to a timestamp.

[0019] Furthermore, the specific process of step 2 is as follows:

[0020] Step 2.1: Based on the preprocessed areal density equivalent value time series data, extract multi-scale fused time series features. The calculation method is as follows:

[0021] ,

[0022] ,

[0023] ,

[0024] in, This represents the short-term fluctuation characteristics at time t. For a temporal convolutional network with an inflation rate of 1, The time series data represents the preprocessed areal density equivalent values ​​at time t. This represents the long-term trend characteristics at time t. For a temporal convolutional network with an inflation rate of 4, The multi-scale fused temporal features at time t are given. For feature splicing operations;

[0025] Step 2.2: Based on the slurry viscosity data, extract the slurry rheological property encoding features through a viscosity modulation network, and perform viscosity modulation to obtain the viscosity-modulated features. The calculation method of the viscosity modulation network is as follows:

[0026] ,

[0027] ,

[0028] ,

[0029] in, Encoding features for the rheological properties of slurry, It is a multilayer perceptron. The viscosity data of the slurry obtained in step 1.4, Let be the viscosity modulation gate weight vector at time t. The viscosity modulated characteristics at time t. For Hadama accumulation, for function;

[0030] Step 2.3: Based on the viscosity modulation characteristics and the slurry rheological property encoding characteristics, the porosity compensation weight coefficient is calculated using a porosity compensation neural network, and the equivalent value of the areal density is corrected to obtain the corrected areal density. The calculation method of the porosity compensation neural network is as follows:

[0031] ,

[0032] ,

[0033] ,

[0034] in, The porosity compensation intermediate characteristics at time t, For layer normalization, It is a fully connected layer. For the Softmax function, For residual connection blocks, The porosity compensation weighting coefficient at time t. For gated loop unit, Let be the corrected areal density at time t. This is the compensation strength coefficient.

[0035] Furthermore, the specific process of step 3 is as follows:

[0036] Step 3.1: Based on the preprocessed process parameter time series data, extract the pump speed time series features and nozzle gap time series features respectively, and then fuse them to obtain the joint coding features of the process parameters. The calculation method is as follows:

[0037] ,

[0038] ,

[0039] ,

[0040] in, The pump speed time-series characteristic at time t, For a causal convolutional network with an inflation rate of 2, The pump speed data is from the preprocessed time-series data of process parameters at time t. Let be the time-series characteristic of the nozzle gap at time t. For a causal convolutional network with an inflation rate of 4, The nozzle gap data is the preprocessed process parameter time series data at time t. The joint encoding features of the process parameters at time t;

[0041] Step 3.2: Based on the pump speed time-series characteristics and nozzle gap time-series characteristics, calculate the two-parameter interactive weight matrix using a cross-attention mechanism. The calculation method of the cross-attention mechanism is as follows:

[0042] ,

[0043] ,

[0044] ,

[0045] in, Let be the initial weight matrix for the two-parameter interaction at time t. For attention mechanisms, For a multilayer perceptron with gating weights, Let be the coupling weight matrix at time t. It is a one-dimensional convolutional layer. Let be the two-parameter interaction weight matrix at time t. This is an element-wise addition;

[0046] Step 3.3: Based on the corrected areal density, process parameter joint encoding features, and two-parameter interaction weight matrix, generate parameter-areal density coupled intermediate features through an attention mechanism. The calculation method is as follows:

[0047] ,

[0048] ,

[0049] ,

[0050] in, The surface density embedding features at time t, Let be the attention fusion feature at time t. The parameter-area density coupled intermediate feature at time t. It is a residual fully connected layer;

[0051] Step 3.4: Generate parameter-surface density coupled features based on the parameter-surface density coupled intermediate features and the two-parameter interaction weight matrix. The calculation method is as follows:

[0052] ,

[0053] in, The parameter-area density coupling characteristic at time t.

[0054] Furthermore, the specific process of step 4 is as follows:

[0055] Step 4.1: Based on the parameter-area density coupling features, extract the two-parameter joint features through feature concatenation and a one-dimensional convolutional neural network. The calculation method is as follows:

[0056] ,

[0057] ,

[0058] ,

[0059] in, The nozzle gap control characteristics at time t. For GeLU functions, The pump speed regulation characteristics at time t, The two-parameter joint feature at time t;

[0060] Step 4.2: Based on the two-parameter joint features and the corrected areal density, calculate the two-parameter joint adjustment features using a multilayer perceptron. The calculation method is as follows:

[0061] ,

[0062] ,

[0063] in, The target surface density deviation at time t. The target value for standard surface density. The two-parameter joint adjustment feature at time t, It is the hyperbolic tangent function;

[0064] Step 4.3: Based on the dual-parameter joint adjustment features, the pump speed command value and nozzle gap command value are obtained through the feature separation layer and truncation mapping function. The calculation method is as follows:

[0065] ,

[0066] ,

[0067] in, The value of the pump speed adjustment command at time t. This is a truncation function. For linear layers, Adjust the gain coefficient to adjust the pump speed. This is the lower limit for adjusting the pump speed. To adjust the upper limit of pump speed, The nozzle gap adjustment command value at time t. Adjust the gain coefficient for nozzle gap. Adjust the lower limit of the nozzle gap. Adjust the upper limit of the nozzle gap.

[0068] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0069] (1) This invention addresses the challenge of areal density control in NCM-Li battery cathode coating caused by complex fluctuations in the porosity of secondary particles and significant nonlinear changes in slurry viscosity. It improves the control accuracy of coating areal density through areal density detection correction and dual-parameter synergistic regulation. On the one hand, it corrects the equivalent value deviation of β-ray detection through a porosity compensation mechanism to adapt to the complex porosity changes caused by the deformation of NCM secondary particles. On the other hand, it adapts to the parameter coupling effect under the strong shear thinning characteristics of NCM slurry through viscosity modulation network and dual-parameter coupling modeling, avoiding the effect cancellation problem of single-parameter independent control. It significantly improves the areal density control stability in dynamic and multi-condition production scenarios, thereby improving the consistency of NCM battery cell capacity and improving the energy density performance and cycle life of the entire pack.

[0070] (2) To address the problem of feedback distortion caused by porosity and viscosity fluctuations in β-ray equivalent measurements, this invention extracts short-term pulsation features and long-term drift features through temporal convolutional networks with different expansion rates, enabling the system to accurately distinguish the nature and source of disturbances. Based on this, the response intensity of each feature dimension is dynamically adjusted according to the current viscosity state of the slurry—increasing the compensation focus under low-speed coating conditions where viscosity has a significant impact, and suppressing the compensation weight under high-speed conditions where viscosity has a smaller impact to avoid overcorrection. Finally, the influence of porosity changes on the measured value is indirectly inferred from multi-scale temporal features and viscosity state, and superimposed on the original measured value in the form of additive residuals. When the porosity is close to the calibration state, the compensation weight automatically approaches zero, ensuring that no additional interference is introduced, effectively eliminating false components in the detection signal, and obtaining a corrected density that is closer to the true dry matter content, providing a reliable feedback basis for subsequent precise control.

[0071] (3) To address the problem of significant differences in the dynamic characteristics of pump speed and nozzle gap, and the failure of single-parameter adjustment due to the fluid dynamic coupling relationship, this invention establishes a dual-parameter collaborative control mechanism. The pump speed is captured by causal convolution with a narrow receptive field to capture its complex dynamics of millisecond-level electrical response and second-level delivery delay. The nozzle gap is perceived by causal convolution with a wide receptive field to sense its slow mechanical adjustment trend, so that the feature extraction of the two parameters is adapted to their physical response characteristics. Through the cross-attention mechanism and the dual-parameter interactive weight matrix, the influence of pump speed change on pipeline pressure and nozzle shear rate, as well as the reverse constraint of nozzle gap adjustment on actual flow rate, are explicitly quantified, and the mutual constraint relationship between the two parameters is encoded. Finally, the corrected areal density deviation information is combined to make joint decisions and output collaborative parameter adjustment instructions, which overcomes the problem of mutual cancellation or nonlinear amplification of effects caused by independent adjustment of single parameters, so that the control system can maintain stable control accuracy in scenarios such as working condition switching and slurry batch replacement. Attached Figure Description

[0072] Figure 1The image shows the training convergence process of the composite loss function provided in the embodiment of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0074] A method for homogenizing the surface density of coating on the positive electrode sheet of an NCM-Li battery includes the following steps:

[0075] Step 1: Collect the viscosity data of the current batch of slurry; collect the time-series data of the areal density equivalent value of the battery positive electrode sheet and the real-time process parameter data of the coating machine, and then perform preprocessing to obtain the preprocessed areal density equivalent value time-series data and the preprocessed process parameter time-series data; the specific process is as follows:

[0076] Step 1.1: Collect time-series data of the equivalent surface density of the positive electrode sheet of the battery using a β-ray thickness gauge. The data type is one-dimensional time-series numerical data, which includes the real-time equivalent surface density values ​​of multiple measurement points uniformly distributed along the coating direction of the electrode sheet and their corresponding timestamps.

[0077] Step 1.2: For the time series data of surface density equivalent values, outliers are removed using the three-standard-deviation principle to obtain the denoised surface density equivalent value time series data;

[0078] Step 1.3: Collect real-time process parameter data of the coating machine. The data type is multi-dimensional time series numerical data, including pump speed data and nozzle gap data. Each parameter corresponds to a timestamp.

[0079] Step 1.4: Measure the viscosity data of the current batch of slurry using a rotational viscometer. The data type is numerical data.

[0080] Step 1.5: The Z-score standardization method is used to eliminate dimensional differences in the time series data of the denoised areal density equivalent value and the real-time process parameter data of the coating machine to obtain the pre-processed areal density equivalent value time series data and the pre-processed process parameter time series data.

[0081] Step 2: Based on the preprocessed areal density equivalent value time series data, extract multi-scale fused time series features; based on the slurry viscosity data, extract viscosity-modulated features through a viscosity modulation network; then, correct the areal density equivalent value through a porosity compensation neural network to obtain the corrected areal density; the specific process is as follows:

[0082] Step 2.1: Based on the preprocessed areal density equivalent value time series data, extract multi-scale fused time series features. The calculation method is as follows:

[0083] ,

[0084] ,

[0085] ,

[0086] in, This represents the short-term fluctuation characteristics at time t. For a temporal convolutional network with an inflation rate of 1, The time series data represents the preprocessed areal density equivalent values ​​at time t. This represents the long-term trend characteristics at time t. For a temporal convolutional network with an inflation rate of 4, The multi-scale fused temporal features at time t are given. For feature splicing operations;

[0087] Step 2.2: Based on the slurry viscosity data, extract the slurry rheological property encoding features through a viscosity modulation network, and perform viscosity modulation to obtain the viscosity-modulated features. The calculation method of the viscosity modulation network is as follows:

[0088] ,

[0089] ,

[0090] ,

[0091] in, Encoding features for the rheological properties of slurry, It is a multilayer perceptron. This is the baseline data for slurry viscosity. Let be the viscosity modulation gate weight vector at time t. The viscosity modulated characteristics at time t. For Hadama accumulation, for function;

[0092] Step 2.3: Based on the viscosity modulation characteristics and the slurry rheological property encoding characteristics, the porosity compensation weight coefficient is calculated using a porosity compensation neural network, and the equivalent value of the areal density is corrected to obtain the corrected areal density. The calculation method of the porosity compensation neural network is as follows:

[0093] ,

[0094] ,

[0095] ,

[0096] in, The porosity compensation intermediate characteristics at time t, For layer normalization, It is a fully connected layer. For the Softmax function, For residual connection blocks, The porosity compensation weighting coefficient at time t. For gated loop unit, Let be the corrected areal density at time t. This is the compensation strength coefficient.

[0097] It needs further explanation that the online accurate detection and correction of areal density during the NCM-Li battery cathode coating production process faces multiple challenges, and these problems are more prominent than those of conventional cathode systems due to the inherent structure and rheological properties of NCM materials. Firstly, the equivalent areal density value measured by β-rays is itself affected by changes in electrode porosity. Coatings of the same thickness will have different dry matter contents depending on the porosity. Since NCM active materials are mostly secondary spherical aggregates, the deformation and local breakage of these secondary particles during rolling dynamically change with the operating conditions, further exacerbating the porosity fluctuations. Simultaneously, porosity dynamically changes with factors such as slurry viscosity, coating speed, and drying conditions, leading to a deviation between the detection feedback value used by the control system and the actual dry matter areal density, thus causing errors in subsequent control decisions. Secondly, in the areal density equivalent value data, the actual coated... The density deviation on the production line originates from two different time scales: one is short-term rapid fluctuations, such as instantaneous changes like pump pulsation or partial nozzle blockage; the other is long-term slow drift, such as gradual changes in solid content due to sedimentation and stratification of the slurry in the circulation pipeline over time, or overall viscosity shifts caused by slow changes in ambient temperature. These trends may take hundreds of sampling cycles to become apparent. Finally, slight differences between batches of raw materials can cause slurry viscosity fluctuations. For high-nickel NCM systems, lithium compounds remaining on the particle surface are prone to react with the binder, leading to continuous viscosity drift. Moreover, the impact of viscosity changes on surface density is not a fixed linear proportional relationship, but rather dynamically changes with operating conditions such as coating speed and nozzle gap. For example, during high-speed coating, the slurry is in a high-shear state, and the impact of viscosity differences on surface density is relatively small, while the impact is significantly amplified during low-speed coating or roll-changing deceleration.

[0098] To address the aforementioned challenges, this invention achieves the correction of the equivalent surface density value through multi-scale fusion temporal feature extraction, viscosity modulation, and porosity compensation. Firstly, in the multi-scale fusion temporal feature extraction stage, two temporal convolutional networks with different dilation rates are used to perform convolution operations on the preprocessed temporal data of the equivalent surface density value. The convolutional network with a smaller dilation rate has a narrower receptive field, used to capture short-term, rapid fluctuations, while the convolutional network with a larger dilation rate has a larger receptive field, capable of perceiving slow drift trends over longer time spans. Then, feature concatenation operations fuse the features from the two scales into a multi-scale fusion temporal feature, enabling subsequent modules to simultaneously acquire complete information on both the fast-changing and slow-drift components in the surface density signal, thus accurately determining the nature and source of the current deviation. First, it lays the foundation. Second, in the viscosity modulation stage, the slurry viscosity reference data is first mapped to the slurry rheological property encoding features through a multilayer perceptron. Then, the encoded features are jointly input into the gating network with the multi-scale fusion temporal features at the current moment to generate a time-by-time viscosity modulation gating weight vector. The multi-scale fusion temporal features are selectively scaled by the Hadamard product. Its core function is to dynamically adjust the response intensity of each feature dimension according to the current rheological state of the slurry. Under the condition that viscosity has a significant impact, the weight of relevant features is amplified to enhance the attention to the deviation caused by viscosity. Under the condition that viscosity has a small impact, the relevant weights are suppressed to avoid overcompensation. Thus, the influence of viscosity fluctuation is effectively separated from the temporal features, providing a clean feature input after viscosity correction for the next step of porosity compensation.

[0099] Finally, in the porosity compensation stage, since porosity is a microstructural feature of the coating and is difficult to measure directly with a single sensor, the system indirectly infers the influence of porosity on the equivalent surface density from multi-scale fusion temporal features and slurry rheological property coding features. Specifically, changes in porosity will change the mapping relationship between the apparent density and dry matter mass of the coating. When the slurry viscosity remains stable but the equivalent surface density fluctuates abnormally, this abnormality often originates from changes in porosity. Step 2.3: After fusing the slurry rheological property encoding features and multi-scale fusion temporal features through a fully connected layer, the residual connections of the viscosity-modulated features are superimposed and layer normalized to obtain the porosity compensation intermediate features. These intermediate features are essentially an implicit representation of the porosity deviation state under the current operating conditions, integrating information from both the slurry viscosity state and the temporal dynamics after viscosity correction. This allows the network to indirectly infer whether the porosity has increased or decreased compared to the calibration value by comparing the current feature representation with the feature pattern under the calibration conditions. Based on this, the porosity deviation intensity signal generated by the porosity compensation intermediate features through the fully connected layer is multiplied element-wise with the temporal evolution memory signal output by the gated loop unit of the viscosity-modulated features to obtain the porosity compensation weight coefficients. The judgment of porosity deviation is divided into two dimensions: the fully connected layer focuses on judging the instantaneous state of porosity deviation based on the feature fusion result at the current moment, for example, when the slurry viscosity remains unchanged but the areal density equivalent value... In the event of a sudden drop, the network determines whether this is due to an increase in porosity. The gated loop unit tracks the gradual change in porosity over time and reflects it in the compensation weight. The multiplication of the two results in a final compensation weight coefficient that reflects both the current state of the porosity deviation and its temporal evolution characteristics. After calculating the compensation weight coefficient, the correction magnitude and compensation intensity coefficient generated from the intermediate characteristics of porosity compensation are combined and superimposed on the original areal density equivalent value in the form of additive residuals to obtain the corrected areal density. When the network determines that the porosity is close to the calibration state or that the abnormal characteristics are mainly due to other factors, the compensation weight coefficient will automatically approach zero, and the correction amount will approach zero, without introducing additional interference to the original measurement value. However, when the porosity does change significantly, the magnitude and direction of the correction amount are adaptively determined by the network based on the viscosity state and temporal characteristics. This allows the corrected areal density to eliminate the false signals of porosity fluctuations in the measurement value and more accurately reflect the true areal density of the electrode coating.

[0100] Compared to conventional practices in existing technologies, traditional control systems typically use fixed-window sliding averages or single filters to process areal density signals, facing inherent contradictions in window length selection. Regarding viscosity compensation, the traditional approach involves manually measuring viscosity during batch changes and then adjusting the pump speed reference value once according to an experience table. This method of one-time calibration and continuous application is ill-suited to address the dynamic changes in viscosity under varying operating conditions. This invention, through the coordinated operation of the three sub-steps mentioned above, achieves simultaneous perception of rapid fluctuations and slow drifts on a time scale, dynamic adaptive stripping of viscosity influences on a process parameter scale, and compensation and correction of porosity interference on a measurement physical scale. It gradually eliminates deviation factors from different sources, ultimately outputting a corrected areal density that provides a reliable basis for subsequent deviation calculations and coating parameter adjustments, avoiding the problem of incorrect control direction due to distorted detection feedback.

[0101] Step 3: Based on the preprocessed process parameter time-series data, extract the pump speed time-series features and nozzle gap time-series features respectively, and then fuse them to obtain the joint encoding features of the process parameters; calculate the two-parameter interaction weight matrix through the cross-attention mechanism, and then calculate the parameter-area density coupling features; the specific process is as follows:

[0102] Step 3.1: Based on the preprocessed process parameter time series data, extract the pump speed time series features and nozzle gap time series features respectively, and then fuse them to obtain the joint coding features of the process parameters. The calculation method is as follows:

[0103] ,

[0104] ,

[0105] ,

[0106] in, The pump speed time-series characteristic at time t, For a causal convolutional network with an inflation rate of 2, The pump speed data is from the preprocessed time-series data of process parameters at time t. Let be the time-series characteristic of the nozzle gap at time t. For a causal convolutional network with an inflation rate of 4, The nozzle gap data is the preprocessed process parameter time series data at time t. The joint encoding features of the process parameters at time t;

[0107] Step 3.2: Based on the pump speed time-series characteristics and nozzle gap time-series characteristics, calculate the two-parameter interactive weight matrix using a cross-attention mechanism. The calculation method of the cross-attention mechanism is as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] in, Let be the initial weight matrix for the two-parameter interaction at time t. For attention mechanisms, For a multilayer perceptron with gating weights, Let be the coupling weight matrix at time t. It is a one-dimensional convolutional layer. Let be the two-parameter interaction weight matrix at time t. This is an element-wise addition;

[0112] Step 3.3: Based on the corrected areal density, process parameter joint encoding features, and two-parameter interaction weight matrix, generate parameter-areal density coupled intermediate features through an attention mechanism. The calculation method is as follows:

[0113] ,

[0114] ,

[0115] ,

[0116] in, The surface density embedding features at time t, Let be the attention fusion feature at time t. The parameter-area density coupled intermediate feature at time t. It is a residual fully connected layer;

[0117] Step 3.4: Generate parameter-surface density coupled features based on the parameter-surface density coupled intermediate features and the two-parameter interaction weight matrix. The calculation method is as follows:

[0118] ,

[0119] in, The parameter-area density coupling characteristic at time t.

[0120] It is important to further explain that in the NCM positive electrode coating process, pump speed and nozzle gap are two key process control parameters, each with different dynamic characteristics and physical response mechanisms. The electrical adjustment of pump speed has a fast response time, typically in the millisecond range, but its final impact on the coating requires the slurry to be transported from the pump through the pipeline to the nozzle. This transport delay is usually in the second range, resulting in a time lag between changes in pump speed and changes in coating areal density. In contrast, the adjustment of the nozzle gap is a mechanical movement with a relatively slow response, but its physical effect on coating thickness is almost instantaneous. These two parameters differ in their time constants, response mechanisms, and the information patterns contained in their historical trajectories. Furthermore, there is a strong hydrodynamic coupling between the two parameters. Since NCM slurry is a typical high-solids-content non-Newtonian fluid with significant shear-thinning behavior, this coupling effect is more pronounced than that of low-viscosity slurry systems. Specifically, when the pump speed is increased to increase the slurry supply, the increased pipeline pressure will change the shear rate of the slurry at the nozzle, thereby changing the apparent viscosity of the non-Newtonian fluid slurry. Ultimately, the increase in coating thickness is not proportional to the increase in pump speed. When the nozzle gap is reduced, the increased flow resistance will affect the actual flow rate. Even if the pump speed remains unchanged, it will lead to changes in coating thickness. This coupling relationship itself will also change dynamically with the slurry viscosity state and the absolute value of the current operating point.

[0121] To address the aforementioned issues, this invention first performs feature processing on pump speed and nozzle gap data separately in the process parameter independent temporal feature extraction stage. Pump speed features are extracted using a causal convolutional network with a smaller expansion rate and a relatively narrower receptive field, used to capture the dynamic response features of rapid pump speed changes. Nozzle gap features are extracted using a causal convolutional network with a larger expansion rate and a larger receptive field, used to perceive the slow mechanical adjustment trend of the nozzle gap across a longer time window, ensuring that the time scale of feature extraction matches the response time constant of each parameter. The use of causal convolution instead of ordinary convolution ensures that feature extraction strictly adheres to causal constraints, relying only on current and past historical information without involving future data, conforming to the physical constraint of "not using future information" in real-time online control scenarios. Subsequently, the pump speed and nozzle gap features are concatenated and fused through a gated recurrent unit. The internal memory mechanism of the gated recurrent unit bridges the time scale differences between the two features, extracting a joint encoded feature of the process parameters that simultaneously contains the dynamic features of both parameters.

[0122] Secondly, in the calculation of the two-parameter interactive weight matrix, a cross-attention mechanism is used to quantify the mutual influence between pump speed and nozzle gap. Cross-attention first uses the pump speed feature as a query vector to evaluate the correlation strength of each dimension in the nozzle gap feature, generating an initial weight matrix. Then, a gated multilayer perceptron incorporates the information of the jointly encoded process parameters into the weight modulation, avoiding the information bias caused by purely local feature interactions. Finally, a one-dimensional convolutional layer performs a nonlinear transformation on the weight matrix and adds inverse attention weights (from nozzle gap to pump speed). The coupling relationship between the two directions is fused through residual connections to obtain the final two-parameter interactive weight matrix. This matrix encodes the mutual constraint relationship between pump speed and nozzle gap at the current moment. Specifically, when the pump speed increases, the pressure in the pipeline increases accordingly, and the outflow velocity at the nozzle increases. This increase in velocity will change... The shear conditions of the slurry passing through the nozzle slit affect the apparent viscosity of the non-Newtonian fluid slurry and the final increase in coating thickness. This process involves the change in pump speed affecting the effect that the nozzle gap should produce by altering the flow field state. The weight values ​​in the matrix are used to quantify the strength of this effect. Conversely, when the nozzle gap narrows, the flow resistance increases. Even if the pump speed remains constant, the actual flow rate in the pipeline will be affected and decrease. This means that adjusting the nozzle gap will change the actual outflow of the slurry that the pump speed could provide at the nozzle. Another set of weights in the matrix is ​​used to quantify the reverse constraint of the nozzle gap change on the pump speed effect. This allows the control system to understand that the adjustment of each parameter not only produces a direct effect but also an indirect effect caused by another parameter when making subsequent parameter adjustment decisions, fully considering the fluid dynamic coupling relationship between them.

[0123] Furthermore, in the intermediate feature generation stage of parameter-density coupling, the corrected areal density value is embedded as a density feature. Then, an attention mechanism is used to allow the areal density feature to query the process parameter joint encoding feature. The two-parameter interactive weight matrix is ​​then used for modulation, thereby generating an intermediate feature that simultaneously carries three aspects of information: areal density state, process parameter historical trajectory, and the coupling relationship between the two parameters. This intermediate feature is then fused with the process parameter joint encoding through a residual fully connected layer and processed by layer normalization, so that the intermediate feature can stably represent the comprehensive impact of parameter adjustment on areal density under the current operating conditions.

[0124] Finally, in the parameter surface density coupling feature generation stage, the element-wise product of the intermediate features and the two-parameter interaction weight matrix is ​​passed through a hyperbolic tangent nonlinear function and a fully connected layer, respectively. The two outputs are then multiplied element-wise to form a gating mechanism. The hyperbolic tangent function provides a bounded nonlinear mapping, and its product with the fully connected path forms adaptive information selection and amplitude modulation. This gating structure can express the nonlinear saturation response characteristics of "significant parameter adjustment effect in some operating ranges (gating fully open), and saturation and flattening effect in other operating ranges (gating gradually close)". It is used to match the nonlinear change characteristics of fluid mechanics during the coating process. The output coupling features integrate the nonlinear fluid response, the dynamic coupling relationship between parameters, and the adjustment effect of the current surface density state on parameter adjustment, providing a sufficient physical information basis for the final control command generation.

[0125] Compared to conventional approaches in existing technologies, traditional coating control systems typically employ fixed parameter mapping rules. During the design phase, the correspondence between pump speed, nozzle gap, and target areal density under different operating conditions is calibrated once. Then, during actual production operation, parameters are adjusted based on current areal density feedback by consulting tables or using fixed formulas. The limitations of this method are: fixed parameter mapping ignores the dynamic coupling relationship between pump speed and nozzle gap. When both parameters change simultaneously, the assumption of linear superposition or simple multiplication makes it difficult to predict the coating effect. The influence of nonlinear factors such as changes in pipeline pressure, flow resistance, and shear rate varies significantly at different operating points. Finally, traditional methods lack awareness of the operating context. The same areal density deviation should be adjusted differently under different viscosity states, different coating speeds, and different current operating points, but the parameter mapping table cannot adaptively adjust based on multi-dimensional operating information, only using a uniform adjustment range. This leads to overcompensation under some operating conditions and undercompensation under others.

[0126] This invention establishes a dynamic, multidimensional coupled parameter-area density relationship model with working condition awareness. In the extraction of time-scale features of pump speed and nozzle gap, the differences in the response mechanisms of the two parameters are considered. In the calculation of the two-parameter interaction weight matrix, the hydrodynamic coupling strength between the two parameters is dynamically quantified. In the generation of area density coupling features, multidimensional working condition information such as current area density state, slurry viscosity encoding, and porosity compensation is incorporated. This enables the system to adapt to various working condition disturbances such as slurry viscosity fluctuations, changes in operating point, and external condition drift, and automatically adjust the parameter adjustment strategy, overcoming the inherent limitations of static parameter tables.

[0127] Step 4: Based on the parameter-area density coupling characteristics, extract the dual-parameter joint features, and then combine them with the corrected area density to sequentially calculate the dual-parameter joint adjustment features, pump speed command value, and nozzle gap command value; the specific process is as follows:

[0128] Step 4.1: Based on the parameter-area density coupling features, extract the two-parameter joint features through feature concatenation and a one-dimensional convolutional neural network. The calculation method is as follows:

[0129] ,

[0130] ,

[0131] ,

[0132] in, The nozzle gap control characteristics at time t. For GeLU functions, The pump speed regulation characteristics at time t, The two-parameter joint feature at time t;

[0133] Step 4.2: Based on the two-parameter joint features and the corrected areal density, calculate the two-parameter joint adjustment features using a multilayer perceptron. The calculation method is as follows:

[0134] ,

[0135] ,

[0136] in, The target surface density deviation at time t. The target value for standard surface density. The two-parameter joint adjustment feature at time t, It is the hyperbolic tangent function;

[0137] Step 4.3: Based on the dual-parameter joint adjustment features, the pump speed command value and nozzle gap command value are obtained through the feature separation layer and truncation mapping function. The calculation method is as follows:

[0138] ,

[0139] ,

[0140] in, Let t be the pump speed command value at time t. This is a truncation function. For linear layers, Adjust the gain coefficient to adjust the pump speed. This is the lower limit for adjusting the pump speed. To adjust the upper limit of pump speed, Let t be the nozzle gap command value at time t. Adjust the gain coefficient for nozzle gap. Adjust the lower limit of the nozzle gap. Adjust the upper limit of the nozzle gap.

[0141] It should be further explained that there is a complex fluid dynamic coupling between the two key process parameters, pump speed and nozzle gap. Changes in pump speed will change the pipeline pressure and thus affect the flow field distribution at the nozzle, while adjustments to the nozzle gap will change the flow resistance and thus affect the actual slurry flow rate. The effects of the two parameters are mutually restrictive and mutually influential in the production process. However, traditional control systems often treat them as independent controllable objects and adjust them separately, resulting in the adjustment effects canceling each other out or nonlinearly amplifying each other.

[0142] In step 4 of this invention, two parallel one-dimensional convolutions are used to extract control features for nozzle gap and pump speed, respectively. The two convolutions share the same parameter-area density coupling feature as input, ensuring that the control feature of nozzle gap naturally includes the influence information of pump speed on its effect, and the control feature of pump speed naturally includes the influence information of nozzle gap on its effect. This prevents subsequent command calculations from falling into the trap of independent parameter adjustment. In the second step, the system calculates the magnitude and direction of the deviation between the current corrected density and the standard target value. This deviation value is regarded as an important working condition context information. After being concatenated with the two-parameter joint feature, it is input into the multilayer perceptron. Based on the current parameter coupling relationship, the area density deviation amplitude, the slurry viscosity state, and other comprehensive working condition information, a reasonable joint adjustment decision is output. In the third step, the common information of the joint adjustment feature is mapped to the two dimensions of pump speed and nozzle gap through the feature separation layer. Each dimension is multiplied by its own adjustment gain coefficient for amplitude scaling. Finally, the truncated mapping function is constrained to the physical working range of the hardware, and the final pump speed command value and nozzle gap command value are output.

[0143] Compared to conventional practices in existing technologies, traditional coating control systems typically employ empirically calibrated parameter curves or lookup tables. These methods involve directly looking up measured values ​​in a table or outputting corresponding pump speed and nozzle gap settings according to fixed formulas. Alternatively, separate proportional-derivative (PD) controllers can be used to adjust two parameters independently. However, these methods have limitations: empirically calibrated parameter tables cannot adapt to dynamic changes in operating conditions such as slurry viscosity fluctuations, coating speed variations, and environmental condition drift. When actual operating conditions deviate from the calibrated baseline conditions, the original parameter correspondence becomes invalid, requiring recalibration. Furthermore, independent PDC controllers neglect the coupling relationship between the two parameters. Adjusting one parameter alters the actual effect of the other, but the controller cannot detect this mutual influence, leading to mutual interference and cancellation during actual execution of the adjustments.

[0144] This invention establishes a dynamic, multidimensional coupled parameter-area density relationship mapping with working condition awareness. When extracting control features, it integrates the coupling information of the two parameters. When calculating joint adjustment, it incorporates multidimensional working condition contexts such as areal density deviation, slurry viscosity encoding, and porosity correction. This enables the system to adapt to various working condition changes and automatically adjust the direction and intensity of parameter adjustment. It overcomes the inherent limitations of static parameter tables or independent controllers and realizes the collaborative control of the electrode coating process.

[0145] Example 1

[0146] In step 4.1, the one-dimensional convolutional neural network used to extract nozzle gap control features has the same number of input channels as the dimension of the parameter-area density coupled features, set to 64, the number of output channels set to 32, the kernel size set to 3, causal padding to maintain the temporal length, and a stride set to 1. The one-dimensional convolutional neural network used to extract pump speed control features uses the same network structure and parameter configuration, i.e., 64 input channels, 32 output channels, 3 kernels, causal padding, and a stride of 1. Both one-dimensional convolutional neural networks use the GeLU function as their activation function. After feature concatenation, the dimension of the two-parameter joint feature is 64.

[0147] In step 4.2, the target surface density is determined according to the process specifications of the NCM-Li battery cathode sheet. In this embodiment, the target process surface density corresponding to a certain model of battery is set to 200.0 mg / cm². The target surface density deviation is expanded in dimension and then concatenated with the two-parameter joint feature. The input dimension after concatenation is 65. The multilayer perceptron adopts a three-layer fully connected structure. The number of neurons in each layer is set to 65, 128, and 64 respectively. The first layer and the second layer use the ReLU activation function and apply the Dropout operation with a dropout probability of 0.1. The final output is activated by the hyperbolic tangent function Tanh, which constrains the numerical range of the two-parameter joint adjustment feature to the interval [-1, 1], and the output dimension is 64.

[0148] In step 4.3, the feature separation layer contains two independent linear layers, used to generate pump speed adjustment and nozzle gap adjustment, respectively. The linear layer for pump speed has an input dimension of 64 and an output dimension of 1; the linear layer for nozzle gap has an input dimension of 64 and an output dimension of 1. Through multi-condition verification experiments, the pump speed adjustment gain coefficient is set to 0.5, and the nozzle gap adjustment gain coefficient is set to 0.3. This ratio can fully utilize the coupling effect of pump speed and nozzle gap to achieve efficient dual-parameter coordinated control. In the parameter settings of the clip function Clip, the lower limit of pump speed adjustment is set to -5.0 rpm, and the upper limit of pump speed adjustment is set to 5.0 rpm, corresponding to the nominal working speed adjustment range of the coating equipment pump; the lower limit of nozzle gap adjustment is set to -10.0 μm, and the upper limit of nozzle gap adjustment is set to 10.0 μm, corresponding to the precision control range of the nominal working gap of the nozzle. The above clipping range is determined according to the actual adjustable range of the coating equipment and the process safety margin, ensuring that the output pump speed command value and nozzle gap command value are both within the safe adjustment range allowed by the equipment.

[0149] The neural network modules involved in this invention, including temporal convolutional networks, viscosity modulation networks, porosity compensation neural networks, causal convolutional networks, cross-attention mechanism modules, gated recurrent units, multilayer perceptrons, and trainable modules such as fully connected layers, one-dimensional convolutional layers, and linear layers, are all jointly trained using an end-to-end supervised learning approach. Specifically, the training dataset consists of time-series data of areal density equivalent values ​​collected during historical production processes, time-series data of coating machine process parameters, slurry viscosity data, and their corresponding labeled data. The labeled data includes the actual measured areal density values, the calibrated optimal pump speed adjustment, and the nozzle gap adjustment. Before training, the original data is preprocessed according to the method described in step 1. The training set, validation set, and test set are randomly divided in a ratio of 7:1.5:1.5. During training, all trainable parameters are initialized with weights using the He initialization method, and the bias term is initialized to zero. The Adam optimizer is used for parameter updates, with the initial learning rate set to 1×10⁻⁶. -3 The first-order momentum decay coefficient was set to 0.9, the second-order momentum decay coefficient was set to 0.999, and the numerical stability constant was set to 1×10⁻⁶. -8 The learning rate scheduling strategy employs cosine annealing, gradually decaying the learning rate from its initial value to 1×10 during training. -6 The training batch size is set to 64, and the maximum number of training epochs is set to 200. The loss function is a composite loss function, defined as:

[0150]

[0151] in, This is the mean square error loss between the corrected areal density and the labeled areal density. This represents the mean square error loss between the pump speed command value and the labeled optimal pump speed adjustment. This represents the mean square error loss between the commanded nozzle gap value and the optimal nozzle gap adjustment value. The first, second, and third loss weights are set to 1.0, 0.5, and 0.5, respectively. To prevent overfitting, the following regularization strategies are introduced during training: Dropout is applied after each multilayer perceptron and fully connected layer, with a dropout probability of 0.1; L2 weight decay regularization is applied to all trainable weight parameters, with a decay coefficient of 1×10⁻⁻⁶. 4 Simultaneously, an early stopping strategy is employed: if the total loss on the validation set does not decrease for 20 consecutive training epochs, training is terminated early, and the model parameters with the optimal validation set loss are retained as the final model. The training process is as follows: Figure 1 As shown.

[0152] In the coating production of NCM-Li battery positive electrode sheets, the viscosity of a certain batch of slurry increased to 1050 mPa·s due to batch variations in raw materials, a 5% increase compared to the standard 1000 mPa·s. Simultaneously, the coating speed decreased from a stable 50 m / min to 40 m / min for process adjustment. At this time, the areal density equivalent value showed a slow decreasing trend. The detection data sequence is as follows: Time... to The time series data of the areal density equivalent value are At the same time, short-term pulsation fluctuations were detected with an amplitude of approximately Under this condition, the coupling effect of slurry viscosity fluctuation and coating speed reduction leads to a slow decrease in areal density. Furthermore, there is a strong hydrodynamic coupling between the pump speed and nozzle gap parameters, and traditional single-parameter adjustment can easily cause the adjustment effects to cancel each other out.

[0153] This invention's method identifies the mapping deviation caused by the decrease in coating speed through a porosity compensation neural network, identifies the slow decreasing trend of areal density through multi-scale time-series features, and then calculates the coupling adjustment weight between pump speed and nozzle gap through two-parameter joint feature learning, at time... to The generated pump speed command value time sequence is as follows The nozzle gap command value time sequence is as follows The coordinated adjustment of the two parameters increases the pump flow rate while reducing the nozzle gap, creating a hydrodynamic coupling compensation that effectively compensates for the decrease in areal density caused by the reduction in coating speed. This results in subsequent... to The areal density equivalent value is stable as The control error is within ±0.5 mg / cm², which fully demonstrates the superiority and robustness of the method under multi-parameter coupling and multi-operating conditions.

[0154] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A method for homogenizing the surface density of coating on the positive electrode sheet of an NCM-Li battery, characterized in that, Includes the following steps: Step 1: Collect slurry viscosity data; collect time-series data of the areal density equivalent value of the battery positive electrode sheet and real-time process parameter data of the coating machine, and then perform preprocessing to obtain the preprocessed areal density equivalent value time-series data and the preprocessed process parameter time-series data. Step 2: Extract multi-scale fused temporal features based on the preprocessed areal density equivalent value time series data; extract viscosity-modulated features based on slurry viscosity data using a viscosity modulation network; The areal density is then corrected by using a porosity compensation neural network to obtain the equivalent value of the areal density. Step 3: Based on the preprocessed process parameter time series data, extract the pump speed time series features and nozzle gap time series features respectively, and then fuse them to obtain the joint coding features of process parameters; The two-parameter interaction weight matrix is ​​calculated through the cross-attention mechanism, and then the parameter-surface density coupling feature is obtained. Step 4: Based on the parameter-area density coupling characteristics, extract the dual-parameter joint characteristics, and then combine them with the corrected area density to calculate the dual-parameter joint adjustment characteristics, pump speed command value, and nozzle gap command value in sequence.

2. The method for homogenizing the surface density of the coating on the positive electrode sheet of an NCM-Li battery as described in claim 1, characterized in that, The specific process of step 1 is as follows: Time-series data of the areal density equivalent value of the positive electrode sheet of the battery were collected using a beta-ray thickness gauge. The data type is one-dimensional time-series numerical data, which includes the real-time equivalent areal density values ​​of multiple measurement points uniformly distributed along the coating direction of the electrode sheet and their corresponding timestamps. Outliers were removed from the areal density equivalent value time-series data using the three-standard-deviation principle to obtain the noise-reduced areal density equivalent value time-series data. Collect real-time process parameter data of the coating machine; the data type is multidimensional time series numerical data. The viscosity data of the slurry was measured using a rotational viscometer, and the data type was numerical. The Z-score standardization method was used to eliminate dimensional differences in the time series data of the denoised areal density equivalent value and the real-time process parameter data of the coating machine, so as to obtain the pre-processed areal density equivalent value time series data and the pre-processed process parameter time series data.

3. The method for homogenizing the surface density of the coating on the positive electrode sheet of an NCM-Li battery as described in claim 2, characterized in that, The real-time process parameters of the coating machine include pump speed data and nozzle gap data; each parameter corresponds to a timestamp.

4. The method for homogenizing the surface density of the coating on the positive electrode sheet of an NCM-Li battery as described in claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Based on the preprocessed areal density equivalent value time series data, extract multi-scale fused time series features. The calculation method is as follows: ， ， ， in, This represents the short-term fluctuation characteristics at time t. For a temporal convolutional network with an inflation rate of 1, The time series data represents the preprocessed areal density equivalent values ​​at time t. This represents the long-term trend characteristics at time t. For a temporal convolutional network with an inflation rate of 4, The multi-scale fused temporal features at time t are given. For feature splicing operations; Step 2.2: Based on the slurry viscosity data, extract the slurry rheological property encoding features through a viscosity modulation network, and perform viscosity modulation to obtain the viscosity-modulated features. The calculation method of the viscosity modulation network is as follows: ， ， ， in, Encoding features for the rheological properties of slurry, It is a multilayer perceptron. The viscosity data of the slurry obtained in step 1.4, Let be the viscosity modulation gate weight vector at time t. The viscosity modulated characteristics at time t. For Hadama accumulation, for function; Step 2.3: Based on the viscosity modulation characteristics and the slurry rheological property encoding characteristics, the porosity compensation weight coefficient is calculated using a porosity compensation neural network, and the equivalent value of the areal density is corrected to obtain the corrected areal density. The calculation method of the porosity compensation neural network is as follows: ， ， ， in, The porosity compensation intermediate characteristics at time t, For layer normalization, It is a fully connected layer. For the Softmax function, For residual connection blocks, The porosity compensation weighting coefficient at time t. For gated loop unit, Let be the corrected areal density at time t. This is the compensation strength coefficient.

5. The method for homogenizing the surface density of the coating on the positive electrode sheet of an NCM-Li battery as described in claim 4, characterized in that, The specific process of step 3 is as follows: Step 3.1: Based on the preprocessed process parameter time series data, extract the pump speed time series features and nozzle gap time series features respectively, and then fuse them to obtain the joint coding features of the process parameters. The calculation method is as follows: ， ， ， in, The pump speed time-series characteristic at time t, For a causal convolutional network with an inflation rate of 2, The pump speed data is from the preprocessed time-series data of process parameters at time t. Let be the time-series characteristic of the nozzle gap at time t. For a causal convolutional network with an inflation rate of 4, The nozzle gap data is the preprocessed process parameter time series data at time t. The joint encoding features of the process parameters at time t; Step 3.2: Based on the pump speed time-series characteristics and nozzle gap time-series characteristics, calculate the two-parameter interactive weight matrix using a cross-attention mechanism. The calculation method of the cross-attention mechanism is as follows: ， ， ， in, Let be the initial weight matrix for the two-parameter interaction at time t. For attention mechanisms, For a multilayer perceptron with gating weights, Let be the coupling weight matrix at time t. It is a one-dimensional convolutional layer. Let be the two-parameter interaction weight matrix at time t. This is an element-wise addition; Step 3.3: Based on the corrected areal density, process parameter joint encoding features, and two-parameter interaction weight matrix, generate parameter-areal density coupled intermediate features through an attention mechanism. The calculation method is as follows: ， ， ， in, The surface density embedding features at time t, Let be the attention fusion feature at time t. The parameter-area density coupled intermediate feature at time t. It is a residual fully connected layer; Step 3.4: Generate parameter-surface density coupled features based on the parameter-surface density coupled intermediate features and the two-parameter interaction weight matrix. The calculation method is as follows: ， in, The parameter-area density coupling characteristic at time t.

6. The method for homogenizing the surface density of the coating on the positive electrode sheet of an NCM-Li battery as described in claim 5, characterized in that, The specific process of step 4 is as follows: Step 4.1: Based on the parameter-area density coupling features, extract the two-parameter joint features through feature concatenation and a one-dimensional convolutional neural network. The calculation method is as follows: ， ， ， in, The nozzle gap control characteristics at time t. For GeLU functions, The pump speed regulation characteristics at time t, The two-parameter joint feature at time t; Step 4.2: Based on the two-parameter joint features and the corrected areal density, calculate the two-parameter joint adjustment features using a multilayer perceptron. The calculation method is as follows: ， ， in, The target surface density deviation at time t. The target value for standard surface density. The two-parameter joint adjustment feature at time t, It is the hyperbolic tangent function; Step 4.3: Based on the dual-parameter joint adjustment features, the pump speed command value and nozzle gap command value are obtained through the feature separation layer and truncation mapping function. The calculation method is as follows: ， ， in, The value of the pump speed adjustment command at time t. This is a truncation function. For linear layers, Adjust the gain coefficient to adjust the pump speed. This is the lower limit for adjusting the pump speed. To adjust the upper limit of pump speed, The nozzle gap adjustment command value at time t. Adjust the gain coefficient for nozzle gap. Adjust the lower limit of the nozzle gap. Adjust the upper limit of the nozzle gap.

Images