A fluidized bed intelligent monitoring method and system for silicon-carbon negative electrode materials

Abstract

The present application relates to the technical field of fluidized bed intelligent monitoring, and provides a fluidized bed intelligent monitoring method and system for silicon-carbon negative electrode materials, which comprises collecting preset data and process acquisition data of the fluidized bed of the silicon-carbon negative electrode materials and preprocessing, constructing differential features according to the original feature data, performing drift detection and processing on the differential features and the original feature data, obtaining a modeling feature set, obtaining a fault type classification through a fault diagnosis submodule, performing flow loss degree evaluation through a risk evaluation submodule to obtain a risk temperature classification, and intelligently monitoring the fluidized bed for the silicon-carbon negative electrode materials according to the fault type classification and the risk temperature classification. The present application provides reliable technical support for the safe production of silicon-carbon negative electrode materials, provides early warning and buffer time for operation and maintenance personnel, can accurately identify different degrees of flow loss state, and has the advantages of high diagnostic precision, early warning time, strong self-adaptability and the like.

Description

Technical Field

[0001] This invention relates to the field of fluidized bed intelligent monitoring technology, and in particular to a fluidized bed intelligent monitoring method and system for silicon-carbon anode materials. Background Technology

[0002] Fluidized bed reactors possess highly efficient heat and mass transfer capabilities and flexible operation. In the field of lithium-ion battery anode material manufacturing, the preparation of novel silicon-carbon composite materials using fluidized bed chemical vapor deposition technology has become a key research focus. However, this process often faces challenges during operation, such as abrupt changes in fluidization state, abnormal particle aggregation, and uneven distribution of reactant gases. If these conditions are not identified and addressed in a timely manner, they can easily lead to serious consequences such as equipment blockage or even runaway reactions.

[0003] Currently, industrial sites mainly rely on operators to interpret limited process parameters, such as pressure and temperature, and make status judgments based on their personal experience. This method has a significant lag in responding to abnormal operating conditions and is difficult to meet the high safety standards required by modern production processes.

[0004] Early studies mostly used signal analysis methods, such as analyzing acoustic signals through wavelet transform to assess the degree of agglomeration, or using statistical methods to process pressure fluctuation data to construct early warning indicators; in recent years, Gaussian mixture models, random forests and various neural networks have been gradually applied to fluidized bed flow pattern recognition and fault detection tasks.

[0005] Existing technical solutions still have significant limitations when facing the specific scenario of silicon-carbon anode material preparation. Most methods are designed only for a single type of fault and lack the ability to comprehensively diagnose multiple abnormal concurrent situations under complex operating conditions. On the other hand, traditional models are not sensitive enough to identify specific risk modes such as abnormal silane distribution and accelerated acetylene cracking, making it difficult to achieve effective early warning in actual production.

[0006] By systematically comparing the performance differences of various modern machine learning frameworks in processing fluidized bed dynamic data, a state recognition model with high accuracy and strong generalization ability was constructed, providing a reliable technical guarantee for the safe production of silicon-carbon anode materials. Notably, this invention further explores different degrees of current loss states, providing maintenance personnel with early warning and buffer time, allowing them to formulate differentiated handling strategies based on the severity. Summary of the Invention

[0007] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a fluidized bed intelligent monitoring method and system for silicon-carbon anode materials. By systematically comparing the performance differences of various modern machine learning frameworks in processing dynamic fluidized bed data, a fault identification model with high accuracy and strong generalization ability is constructed. This provides a reliable technical guarantee for the safe production of silicon-carbon anode materials, identifies different degrees of current loss states, provides early warning and buffer time for operation and maintenance personnel, and allows them to formulate differentiated handling strategies according to the severity. It can accurately identify different degrees of current loss states and has advantages such as high diagnostic accuracy, early warning time, and strong adaptability.

[0008] This invention provides a fluidized bed intelligent monitoring method for silicon-carbon anode materials, comprising:

[0009] S1: Collect fluidized bed preset data and process acquisition data of silicon-carbon anode materials;

[0010] S2: Preprocess the preset data and process acquisition data to obtain the raw feature data;

[0011] S3: Construct differential features based on the original feature data;

[0012] S4: Use HyperGBM to perform drift detection and processing on the differential features and original feature data to obtain the modeling feature set;

[0013] S5: Input the data from the modeling feature set into the fault diagnosis submodule to obtain the fault type classification;

[0014] S6: The risk assessment submodule assesses the degree of flow loss in the data of the modeling feature set to obtain a risk temperature classification.

[0015] S7: Intelligent monitoring of fluidized beds for silicon-carbon anode materials based on fault type and risk temperature classification.

[0016] Furthermore, the preset data includes silane flow rate, nitrogen flow rate, and set temperature, and the process acquisition data refers to the time-series data reflecting the pressure and temperature of the fluidization process in the reactor under the preset data conditions.

[0017] Furthermore, step S2 includes:

[0018] The pressure signal is converted into a pressure drop signal, and anomalies are detected using box plots combined with expert experience. The outliers obtained are treated as missing values.

[0019] Missing values ​​are filled by selecting a missing value imputation method through cross-validation to obtain preprocessed device data;

[0020] A dataset is constructed by integrating equipment data with loss labels based on expert experience annotation. The dataset is then divided into a modeling set and a cross-time validation set. The modeling set is further divided into a training set and a test set. The cross-time validation set consists of the last sample of the time slice of the modeling set. The equipment data includes preset data and process acquisition data.

[0021] Furthermore, in step S3, differential features are constructed based on the original feature data, and the calculation expression is as follows:

[0022]

[0023] in, For original features in Data at any given time For original features in Data at any given time These are the constructed difference features.

[0024] Furthermore, in step S5, the fault diagnosis submodule includes an AxLSTM model, which includes an attention layer, an xLSTM layer, and a normalization layer stacked sequentially. The xLSTM layer includes sLSTM or mLSTM. sLSTM enhances memory capacity through exponential gating, while mLSTM improves expressive power by utilizing a matrix memory structure.

[0025] Furthermore, in step S5, the optimal combination of hyperparameters for the AxLSTM model is determined using a Bayesian optimization method.

[0026] Furthermore, the risk assessment submodule employs principal component analysis to assess the degree of data loss in the modeling feature set, including:

[0027] Pressure data of the equipment at different preset temperatures are grouped by equipment and temperature;

[0028] PCA decomposition was performed on each set of pressure data to obtain 1 eigenvalue and eigenvector;

[0029] Calculate the risk score for each group based on the eigenvalues ​​and eigenvectors.

[0030] Furthermore, the expression for calculating the risk score is as follows:

[0031]

[0032] in, For the first Group risk score, For the first The original data vector of the group, For the first 1 eigenvalue, For the first Eigenvectors with eigenvalues, This is the transpose of the matrix.

[0033] Furthermore, the risk scores of each group are sorted in descending order to obtain the risk ranking results of each device at different temperatures;

[0034] The natural breakpoint classification method is used to divide the ranked risk scores into three risk levels.

[0035] Safe flow zone: Powder flow is stable with low risk of flow loss;

[0036] Critical monitoring zone: Fluctuations in flow patterns begin to appear, requiring an increase in monitoring frequency;

[0037] High-risk loss of current zone: Loss of current is very likely to occur, and long-term operation in this temperature range should be avoided.

[0038] This invention also provides a fluidized bed intelligent monitoring system for silicon-carbon anode materials, used to execute a fluidized bed intelligent monitoring method for silicon-carbon anode materials, comprising:

[0039] The collection module collects preset data and process acquisition data of the fluidized bed for silicon-carbon anode materials;

[0040] The preprocessing module preprocesses the preset data and the process acquisition data to obtain the original feature data;

[0041] The feature optimization module constructs differential features based on the original feature data;

[0042] The drift detection module uses HyperGBM to perform drift detection and processing on the differential features and original feature data to obtain a modeling feature set.

[0043] The fault diagnosis module inputs data from the modeling feature set into the fault diagnosis submodule to obtain fault type classification;

[0044] The current loss risk assessment module assesses the degree of current loss by evaluating the data in the modeling feature set through the risk assessment submodule, and obtains the risk temperature classification.

[0045] The intelligent monitoring module performs intelligent monitoring of the fluidized bed for silicon-carbon anode materials based on fault type classification and risk temperature classification.

[0046] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0047] This invention addresses the fluidized bed chemical vapor deposition process in the preparation of silicon-carbon anode materials. By dynamically acquiring multi-source sensor data within the reactor, an optimized machine learning fault identification model is constructed to achieve real-time evaluation of fluidization quality and material coating effect within the bed. This model also identifies various abnormal operating conditions, including adhesion loss, at an early stage. This method can effectively prevent safety risks caused by local enrichment of silanes and rapid decomposition of acetylene, significantly improve the safety and reliability of the production process, and enhance adaptability and diagnostic accuracy for complex operating conditions.

[0048] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a schematic flowchart of a fluidized bed intelligent monitoring method for silicon-carbon anode materials provided by the present invention.

[0051] Figure 2 This is a schematic diagram of the AxLSTM model structure provided by the present invention.

[0052] Figure 3 This is a schematic diagram of the structure of a fluidized bed intelligent monitoring system for silicon-carbon anode materials provided by the present invention.

[0053] Figure label:

[0054] 101. Collection module; 102. Preprocessing module; 103. Feature optimization module; 104. Drift detection module; 105. Fault diagnosis module; 106. Loss of current risk assessment module; 107. Intelligent monitoring module. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0056] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0057] The following is combined with Figures 1 to 3 This invention describes a fluidized bed intelligent monitoring method and system for silicon-carbon anode materials.

[0058] like Figure 1 As shown, a fluidized bed intelligent monitoring method for silicon-carbon anode materials includes:

[0059] S1: Collect preset data and process acquisition data of the fluidized bed for silicon-carbon anode materials;

[0060] Data on the reaction of specific gases at specified flow rates and set temperatures is collected, including preset data and process acquisition data. The preset data includes silane flow rate, nitrogen flow rate, and set temperature.

[0061] The process acquisition data refers to the time-series data of pressure drop and temperature reflecting the fluidization process in the reactor under preset conditions. It is obtained by conversion through sensors. High-precision dynamic pressure sensors and thermocouple temperature sensors are installed at the bottom, middle, and top of the powder section of the reactor and at the gas outlet, respectively.

[0062] The pressure sensor has a sampling frequency of 100 Hz, and the thermocouple temperature sensor has a sampling frequency of 10 Hz. Real-time pressure and temperature data are obtained by converting the sensor signals into AD signals using a digital hub.

[0063] If the pressure drop decreases by 20%, it indicates the onset of localized sintering, which is the beginning of current loss (early stage); when the pressure drop decreases by 40%, there is significant internal sintering, which is partial current loss (mid-stage); when the pressure drop decreases by more than 60%, large-scale internal sintering occurs, which is complete current loss (late stage). If the temperature change after the reaction temperature exceeds 40 degrees Celsius, it is defined as a malfunction or current loss.

[0064] The experimental data of this invention comes from the monitoring data of the silicon deposition process. The preset data are silane flow rate 0.20 L / min, nitrogen flow rate 0.60 L / min, and set temperature 420℃. The process data collected are pressure drop at measuring port 1, pressure drop at measuring port 2, pressure drop at measuring port 3, and real-time temperature. The collected data are shown in Table 1. Table 1 only lists part of the data. Data was collected by a single device under the preset operating conditions, and multiple pressure and real-time temperature data were collected under the same preset conditions.

[0065] Table 1 Data Collection

[0066]

[0067] Measuring port 1 represents the pressure drop of all powder particles from bottom to top, measuring port 2 represents the pressure drop of all powder particles from middle to top, measuring port 3 represents the pressure drop of local powder particles at the top, flow rate represents the operating conditions of the reacting gas, and temperature represents the real-time temperature change.

[0068] The pressure drop signal is the primary indicator of fluidization quality. If sintering and flow loss occur, the pressure drop of the powder will decrease. We can define label 1 as follows: when the pressure drop decreases by 20%, there is localized sintering inside, and flow loss begins (early stage); when the pressure drop decreases by 40%, there is a large amount of sintering inside, with localized bubbling flow and partial flow loss (mid-stage); when the pressure drop decreases by 60%, large-scale sintering occurs inside, and complete flow loss occurs (late stage).

[0069] Temperature fluctuations mainly occur after flow loss, when the powder reacts too quickly in certain areas, leading to temperature changes. These fluctuations are generally within ±30 degrees Celsius; fluctuations exceeding ±40 degrees Celsius are considered a malfunction. A label can be defined: a malfunction occurs when the temperature increases or decreases by 40 degrees Celsius; otherwise, normal operation is maintained.

[0070] S2: Preprocess the preset data and process acquisition data to obtain the raw feature data;

[0071] The pressure signal is converted into a pressure drop signal, and outlier detection is performed using box plots combined with expert experience. The outliers are treated as missing values, and the missing value filling method is selected through cross-validation to fill the missing values, thus obtaining the preprocessed equipment data.

[0072] The pressure signal is converted into a pressure drop signal. Specifically, the difference between the powder section signal and the outlet pressure signal can be used to obtain pressure drop data that characterizes the fluidization quality of different powders.

[0073] Outlier detection using box plots combined with expert experience includes:

[0074] Based on the upper quartile Q3 and the lower quartile Q1, the interquartile range IQR is calculated, IQR = Q3 - Q1. Data points that exceed the range [Q1-1.5IQR, Q3+1.5IQR] are initially identified as outliers.

[0075] Based on expert experience, conduct a business rationality assessment of these statistical outliers;

[0076] Samples that are ultimately confirmed as outliers will be treated as missing values;

[0077] By comparing the performance of various missing value imputation methods on preset evaluation metrics (such as fault detection rate, recall rate, etc.), the mean imputation method with the best model performance is selected based on the model effect to ensure that it can effectively improve the model's predictive performance.

[0078] Various methods for filling missing values ​​include fixed value filling, mean filling, filling with the previous data, and interpolation filling.

[0079] A dataset is constructed by integrating equipment data with loss labels based on expert experience annotation. The dataset is divided into a modeling set and a cross-time validation set. The modeling set is further divided into a training set and a test set with a ratio of 8:2. The cross-time validation set includes the last sample segment of the time slice of the modeling set. The equipment data includes preset data and process acquisition data.

[0080] S3: Construct differential features based on the original feature data;

[0081] To extract richer feature information, differential features are constructed based on the original feature data, and the calculation expression is as follows:

[0082]

[0083] in, For original features in Data at any given time For original features in Data at any given time For the constructed difference features, by adjusting The size can be used to construct different difference features.

[0084] The specific meanings of the data characteristics under the same preset temperature and equipment are shown in Table 2.

[0085] Table 2 Data characteristics and their meanings

[0086]

[0087] S4: Use HyperGBM to perform drift detection and processing on the differential features and original feature data to obtain the modeling feature set;

[0088] Based on the idea of ​​adversarial verification, HyperGBM can automatically detect and process data drift. If a feature drifts, it will directly delete the feature and update the candidate feature set to obtain the modeling feature set.

[0089] S5: Input the data from the modeling feature set into the fault diagnosis submodule to obtain the fault type classification;

[0090] The fault diagnosis submodule includes an AxLSTM model, which comprises sequentially stacked attention layers, xLSTM layers, and normalization layers. The xLSTM layers can be either sLSTM or mLSTM. sLSTM enhances memory capacity through exponential gating, while mLSTM utilizes a matrix memory structure to improve expressive power. sLSTM enhances memory capacity by replacing the sigmoid activation functions of the input and forget gates with exponential functions, expanding the output range and gradient magnitude, thereby mitigating gradient vanishing and accelerating the capture of input dynamics. It also improves stability through normalization and additional states. mLSTM expands the states and weights into a matrix form, significantly increasing memory capacity, and enhances expressive efficiency through parallelizable layer normalization.

[0091] sLSTM and mLSTM improve the model's memory capacity and training efficiency by focusing on gating mechanisms and memory structures, respectively.

[0092] Inspired by the successful application of LSTM models in time series anomaly detection, their gating mechanisms (such as input gates, forget gates, and output gates) can effectively capture temporal dependencies. The state update of a traditional LSTM unit can be represented as:

[0093]

[0094]

[0095]

[0096]

[0097] in, The activation value for the forget gate. It is the sigmoid activation function. For the weight of the forget gate, This is the hidden state from the previous moment. Enter the current time. To offset the forget gate, This is the activation value of the input gate. For the input gate weights, For input gate bias, Candidate memory cells for the current moment. It is the tangent function. As candidate gate weights, For candidate gate bias, For current memory cells, This refers to the memory cells from the previous time step. However, traditional LSTM has limitations in handling long-range dependencies and scalability. Furthermore, xLSTM introduces sLSTM and mLSTM blocks. sLSTM enhances memory capacity through exponential gating, while mLSTM utilizes a matrix memory structure to improve expressive power and significantly improves training efficiency through parallelizable layer normalization. Due to the improved matrix memory structure and parallelizable layer normalization, both model capacity and training efficiency are significantly improved.

[0098] To further enhance the model's ability to perceive global contextual information, this invention introduces an attention mechanism on top of xLSTM to construct a fault diagnosis model, AxLSTM (AttentionxLSTM), capable of collaboratively capturing long-term dynamics and key features. AxLSTM calculates the global context vector through attention weights, as expressed in the following expression:

[0099]

[0100]

[0101] in, In order to be in Time for the first Attention weights for each position, The target location is in a hidden state. For the first The hidden state of each source location, For the first The hidden state of each source location, For the generated context vector, For attention scoring functions.

[0102] By effectively integrating the long-term dynamic features extracted by xLSTM with key information focused by attention, a robust fault diagnosis model can be constructed that can collaboratively capture long-term dependency and discriminative features.

[0103] Bayesian optimization uses Bayes' theorem to estimate the posterior distribution of the objective function, and then selects the next sample of hyperparameter combinations based on this distribution. By fully utilizing information from the previous sampling point, it can better adjust the current parameters and quickly find the parameters that maximize the global objective function. Compared to grid search, Bayesian optimization requires fewer iterations and runs faster. Given a specific range for the parameters, multiple parameters can be adjusted at once, thus preventing dimensionality explosion even with a large number of parameters.

[0104] To determine the optimal hyperparameter combination for the AxLSTM model, this invention employs a Bayesian optimization method for automated hyperparameter tuning. The Bayesian optimization method formalizes the hyperparameter optimization problem into finding the optimal combination on the validation set. Make the objective function To maximize the hyperparameter configuration, the calculation expression is:

[0105]

[0106] in, For optimal hyperparameter configuration, For the hyperparameter search space, For model in hyperparameter configuration The AUC (area under the receiver operating characteristic curve) performance metric is used to measure the performance of the receiver operating characteristic curve. To maximize the value of the independent variable, Bayesian optimization models the objective function using a Gaussian process:

[0107]

[0108] in, It is a mean function. Let covariance function be used. The Gaussian modeling process is based on the observed dataset. , , For the first The hyperparameter configuration of each sample is used to update the posterior distribution using Bayes' theorem:

[0109]

[0110] in, In order to observe the dataset Under the conditions, the model The probability of it being true. Proportional to, In the model Under the condition that it is valid, the observed dataset The probability, To model before any data is observed The probability of it being true;

[0111] Using acquisition functions Select the next evaluation point and calculate the expression:

[0112]

[0113] in, for Hyperparameter configuration at any time.

[0114] Bayesian optimization methods can efficiently locate high-performance hyperparameter regions with a small number of iterations. The Bayesian optimization parameter tuning range and results of the AxLSTM model of this invention are shown in Table 3.

[0115] Table 3. Bayesian optimization parameter tuning range and results

[0116]

[0117] like Figure 2 As shown, the data in the modeling feature set Input the AxLSTM model to obtain the fault type. Predicted output .

[0118] S6: The risk assessment submodule assesses the degree of flow loss in the data of the modeling feature set to obtain a risk temperature classification.

[0119] The risk assessment submodule employs principal component analysis (PCA) to evaluate the degree of flow loss in the data within the modeling feature set. The PCA-based pre-assessment algorithm aims to establish a correlation model between temperature and powder flow stability. This model is built by analyzing pressure sequences at different temperatures. The algorithm maps flow data at various temperatures to a low-dimensional space and groups pressure data from the equipment at different preset temperatures according to both the equipment and temperature.

[0120] PCA decomposition was performed on each set of pressure data to obtain 1 eigenvalue and eigenvector;

[0121] The risk score for each group is calculated based on the eigenvalues ​​and eigenvectors, using the following formula:

[0122]

[0123] in, For the first Group risk score, For the first Group 1 Risk contribution of each principal component For the first The original data vector of the group, For the first 1 eigenvalue, For the first Eigenvectors with eigenvalues, This is the transpose of the matrix.

[0124] S7: Intelligent monitoring of fluidized beds for silicon-carbon anode materials based on fault type and risk temperature classification.

[0125] A higher risk score indicates a greater risk of flow loss. After ranking the risk scores of all temperatures, the risk levels were divided into three categories using the natural breakpoint classification method: "Safe Flow Zone" (low risk); "Critical Monitoring Zone" (requiring enhanced monitoring); and "High-Risk Flow Loss Zone" (should avoid long-term operation). This method provides effective data support for setting safe process temperatures.

[0126] The risk scores of each group are sorted in descending order to obtain the risk ranking results of each device at different temperatures.

[0127] The ranked risk scores were divided into three risk levels using Jenks Natural Breaks classification.

[0128] Safe flow zone: Powder flow is stable with low risk of flow loss;

[0129] Critical monitoring zone: Fluctuations in flow patterns begin to appear, requiring an increase in monitoring frequency;

[0130] High-risk loss of current zone: Loss of current is very likely to occur, and long-term operation in this temperature range should be avoided.

[0131] It can systematically evaluate the overall impact of different temperature conditions on the stability of powder flow, thereby providing clear data support and decision-making guidance for operation and maintenance personnel in setting safe process temperatures and optimizing operating windows.

[0132] Fault types include:

[0133] When the pressure drop decreases by 20%, localized sintering occurs inside, and flow loss begins.

[0134] When the pressure drop decreases by 40%, there is a large amount of sintering inside, local bubbling flow, and partial loss of flow.

[0135] When the pressure drop decreases by 60%, large-scale sintering occurs internally, resulting in complete loss of flow.

[0136] By comparing different algorithms and combining system evaluation indicators, the system's discrimination ability, prediction ability, generalization ability, and stability are comprehensively evaluated.

[0137] The system evaluation metrics are fault detection rate, recall rate, false alarm rate, and F0.5 score. As shown in Table 4, in industrial fault scenarios, reducing the false alarm rate is more critical than capturing potential faults. Therefore, a higher F0.5 value indicates that the system maintains a high fault detection capability while having a lower false alarm rate, resulting in better overall performance and higher practical value for fault prediction. System predictive ability refers to the system's prediction accuracy; better predictive ability indicates stronger discriminative ability. System generalization ability refers to the system's predictive ability on new datasets with similar patterns. System stability refers to the fluctuation of prediction results under different random sampling outcomes.

[0138] Table 4 Evaluation Indicators

[0139]

[0140] To verify the effectiveness of this invention, performance was compared with several baseline models, as shown in Table 5. The performance comparison results are shown in Table 6. The AxLSTM model of this invention achieved an F0.5 score of 89.64%, significantly outperforming all baseline models, thus verifying the effectiveness of its hybrid architecture designed for the temporal characteristics of powder flow. Specific analysis is as follows: Traditional strong baseline models LightGBM and LSTM achieved F0.5 scores of 80.25% and 83.12%, respectively, indicating that they can capture discrimination patterns in the data to some extent, but may be limited in their ability to model long-range, complex dynamic dependencies. Other modern time series models, such as Time-LLM (53.56%), ModernTCN (46.12%), and SensitiveHUE (56.41%), all failed to reach 60%. Their poor performance may be due to the following reasons: Time-LLM, as a large language model, may focus too much on global semantic associations and be insensitive to local subtle fault precursor oscillations; ModernTCN relies on dilated convolution to capture long-term dependencies and may have difficulty adaptively learning multi-scale time-varying periodic patterns in powder flow; SensitiveHUE's spectral analysis method may lack stability under strong noise and non-stationary conditions.

[0141] Table 5 Baseline Model

[0142]

[0143] The xLSTM module enhances the model's ability to remember and generalize long-term dynamics by utilizing the scalar memory gating of sLSTM and the matrix memory structure of mLSTM. Meanwhile, the introduced multi-head attention mechanism adaptively focuses on key time segments strongly correlated with faults. This synergistic work enables the model to more accurately capture transient fluctuations and gradual trends in fluidized beds that foreshadow faults, thus achieving a lower false alarm rate while maintaining high recall, meeting the stringent reliability requirements of industrial applications.

[0144] Table 6 Performance Comparison of AxLSTM Model and Baseline Model

[0145]

[0146] like Figure 3 As shown, a fluidized bed intelligent monitoring system for silicon-carbon anode materials is used to execute a fluidized bed intelligent monitoring method for silicon-carbon anode materials, including:

[0147] The collection module 101 collects fluidized bed preset data and process acquisition data for silicon-carbon anode materials;

[0148] The preprocessing module 102 preprocesses the preset data and the process acquisition data to obtain the raw feature data;

[0149] Feature optimization module 103 constructs differential features based on the original feature data;

[0150] The drift detection module 104 uses HyperGBM to perform drift detection and processing on the differential features and the original feature data to obtain the modeling feature set;

[0151] The fault diagnosis module 105 inputs the data from the modeling feature set into the fault diagnosis submodule to obtain the fault type classification;

[0152] The current loss risk assessment module 106 assesses the degree of current loss in the data in the modeling feature set through the risk assessment submodule, and obtains the risk temperature classification.

[0153] The intelligent monitoring module 107 performs intelligent monitoring of the fluidized bed for silicon-carbon anode materials based on fault type classification and risk temperature classification.

[0154] Through the collaborative work of the above modules, and by dynamically acquiring multi-source sensor data within the reactor, an optimized machine learning fault identification model is constructed. This enables real-time evaluation of fluidization quality and material coating effect within the bed, and early identification of various abnormal operating conditions, including adhesion loss. This method can effectively prevent safety risks caused by local enrichment of silane and rapid decomposition of acetylene, significantly improve the safety and reliability of the production process, and enhance adaptability and diagnostic accuracy to complex operating conditions.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fluidized bed intelligent monitoring method for silicon-carbon anode materials, characterized in that, include: S1: Collect fluidized bed preset data and process acquisition data for silicon-carbon anode materials; S2: Preprocess the preset data and process acquisition data to obtain the raw feature data; S3: Construct differential features based on the original feature data, and determine the optimal feature set based on the differential features; S4: Use HyperGBM to perform drift detection and processing on the feature data in the better feature set to obtain the optimal feature set; S5: Input the data from the optimal feature set into the AxLSTM model to obtain the fault type classification; S6: Based on principal component analysis, assess the degree of flow loss in the data of the optimal feature set to obtain risk temperature classification; Pressure data of the equipment at different preset temperatures are grouped by equipment and temperature; PCA decomposition was performed on each set of pressure data to obtain Each eigenvalue and eigenvector; Calculate the risk score for each group based on eigenvalues ​​and eigenvectors; The formula for calculating the risk score is: in, For the first Group risk score, For the first The original data vector of the group, For the first 1 eigenvalue, For the first Eigenvectors with eigenvalues, This is the transpose of the matrix; The risk scores of each group are sorted in descending order to obtain the risk ranking results of each device at different temperatures. The natural breakpoint classification method is used to divide the ranked risk scores into three risk levels. Safe flow zone: Powder flow is stable with low risk of flow loss; Critical monitoring zone: Fluctuations in flow patterns begin to appear, requiring an increase in monitoring frequency; High-risk loss-of-current zone: Loss of current is very likely to occur, and long-term operation in this temperature range should be avoided; S7: Intelligent monitoring of fluidized beds for silicon-carbon anode materials based on fault type and risk temperature classification.

2. The fluidized bed intelligent monitoring method for silicon-carbon anode materials according to claim 1, characterized in that, The preset data includes silane flow rate, nitrogen flow rate, and set temperature. The process acquisition data refers to the time-series data reflecting the pressure and temperature of the fluidization process in the reactor under the preset data conditions.

3. The fluidized bed intelligent monitoring method for silicon-carbon anode materials according to claim 1, characterized in that, Step S2 includes: The pressure signal is converted into a pressure drop signal, and anomalies are detected using box plots combined with expert experience. The outliers obtained are treated as missing values. Missing values ​​are filled by selecting a missing value imputation method through cross-validation to obtain preprocessed device data; A dataset is constructed by integrating equipment data with loss labels based on expert experience annotation. The dataset is divided into a modeling set and a cross-time validation set. The modeling set is divided into a training set and a test set. The cross-time validation set is the last sample of the time slice of the modeling set. The equipment data includes preset data and process acquisition data.

4. The fluidized bed intelligent monitoring method for silicon-carbon anode materials according to claim 1, characterized in that, In step S3, differential features are constructed based on the original feature data, and the calculation expression is as follows: in, For original features in Data at any given time For original features in Data at any given time These are the constructed difference features; The preferred feature set includes differential features, real-time temperature, and real-time pressure drop.

5. The fluidized bed intelligent monitoring method for silicon-carbon anode materials according to claim 1, characterized in that, In step S5, the AxLSTM model includes an attention layer, an xLSTM layer, and a normalization layer stacked in sequence. The xLSTM layer includes sLSTM or mLSTM. sLSTM enhances memory capacity through exponential gating, while mLSTM improves expressive power by utilizing a matrix memory structure.

6. The fluidized bed intelligent monitoring method for silicon-carbon anode materials according to claim 1, characterized in that, In step S5, the optimal combination of hyperparameters for the AxLSTM model is determined using the Bayesian optimization method.

7. A fluidized bed intelligent monitoring system for silicon-carbon anode materials, characterized in that, A fluidized bed intelligent monitoring method for silicon-carbon anode materials as described in any one of claims 1 to 6, comprising: The collection module collects preset data and process acquisition data of the fluidized bed for silicon-carbon anode materials; The preprocessing module preprocesses the preset data and the process acquisition data to obtain the original feature data; The feature optimization module constructs differential features based on the original feature data and determines a better feature set based on the differential features. A drift detection module uses HyperGBM to perform drift detection and processing on the feature data in the optimal feature set to obtain the optimal feature set. The fault diagnosis module inputs the data from the optimal feature set into the AxLSTM model to obtain fault type classification. The loss of current risk assessment module assesses the degree of loss of current based on principal component analysis of the data in the optimal feature set, and obtains a risk temperature classification. The intelligent monitoring module performs intelligent monitoring of the fluidized bed for silicon-carbon anode materials based on fault type classification and risk temperature classification.

Images