Graphite component online detection system and method for industrial production line

Abstract

The invention provides an industrial production line-oriented graphite component online detection system and method, and belongs to the technical field of material component detection. The scheme aims at solving the problems that in the prior art, online detection of the graphite purity is low in precision and cannot be automated. According to the scheme, graphite samples are collected in real time on an industrial production line, and physical state adjustment pretreatment is carried out on the graphite samples to eliminate interference of sample heterogeneity; performing spectrum scanning on the pretreated sample to obtain spectrum data; a deep learning model comprising a feature extraction unit and an attention unit is adopted to process the spectral data, the feature extraction unit is used for extracting spectral local features, and the attention unit is used for weighting the local features to focus key features, so that a graphite purity value is obtained through calculation; and feeding back the purity value to a production control system to realize closed-loop control. According to the invention, high-precision and high-efficiency online detection and automatic control of the graphite purity are realized, and the robustness is strong.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material composition detection, in particular, to an industrial production line-oriented graphite composition online detection system and method based on spectral analysis and deep learning. BACKGROUND

[0002] As a key industrial raw material, the purity of graphite directly affects the final performance of downstream products such as lithium battery negative electrodes and special graphite products. Therefore, it is crucial to accurately monitor the purity during the production of graphite.

[0003] In the prior art, traditional graphite purity detection methods, such as chemical titration or laboratory spectral analysis, can provide high detection accuracy, but generally have problems such as long detection period, complex operation process, need to take samples and destroy samples, etc., which cannot meet the needs of real-time and continuous quality monitoring of modern industrial production lines.

[0004] In order to realize online detection, the industry has also made some explorations. For example, some schemes indirectly calculate the purity by measuring physical parameters such as resistivity and density, but such methods are not sensitive to impurity types, have limited accuracy, and are easily disturbed by production environment (such as temperature and humidity). In recent years, online monitoring systems combining spectral analysis technology and artificial intelligence models have made progress in some fields (such as liquid composition analysis). Such systems can quickly obtain material information using a spectrometer and analyze it through a deep learning model to achieve fast composition prediction and feedback control. However, when such technology is directly applied to the online detection of solid powder or particulate materials such as graphite, it faces unique technical challenges that existing technology has not effectively solved. Specifically, the physical state of solid materials, such as water content, particle size, and compaction density, is extremely uneven on the production line. This unevenness can cause serious scattering and noise interference to the spectral signal, resulting in poor repeatability and low accuracy of the measurement results, so that the advantages of existing intelligent analysis systems cannot be fully utilized.

[0005] Therefore, the existing technology still lacks a graphite purity detection scheme that can effectively overcome the unevenness of the physical state of solid materials while achieving high-precision and high-efficiency online detection and automatic closed-loop control. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide an industrial production line-oriented graphite composition online detection system and method to solve the technical problems in the prior art that the graphite purity detection method cannot consider real-time, high precision and automation, especially the lack of an online detection scheme that effectively overcomes the influence of the unevenness of the physical state of solid samples on measurement accuracy.

[0007] According to the present invention, an online graphite composition detection system for industrial production lines includes: a sampling module for real-time acquisition of graphite material samples on an industrial production line; a preprocessing module connected to the sampling module for physical state adjustment of the acquired graphite material samples; a spectral analysis module equipped with a spectral detection device for spectral scanning of the preprocessed samples to obtain spectral data related to graphite purity; a data processing module for receiving and processing the spectral data, wherein the deep learning model used by the data processing module includes: a feature extraction unit for extracting local features reflecting local spectral changes from the spectral data; and an attention unit connected after the feature extraction unit for weighting the local features to focus on key features related to graphite purity and calculating the graphite purity value based on the weighted features; and a feedback control module connected to the data processing module and the production control system of the production line for feeding back the calculated graphite purity value to the production control system.

[0008] Optionally, the pretreatment module includes: a drying unit for drying the graphite material sample to reduce the water content of the sample; and a homogenization unit for homogenizing the graphite material sample to form a physically homogeneous test sample.

[0009] Optionally, the spectral detection device is a near-infrared spectrometer or a Raman spectrometer.

[0010] Furthermore, when the spectral detection device is a near-infrared spectrometer, its detection wavelength range is 800-2500nm.

[0011] Optionally, before inputting the spectral data into the deep learning model, the data processing module is further configured to perform at least one of the following preprocessing on the spectral data: normalization processing; wavelet transform denoising processing; and polynomial baseline correction processing.

[0012] In a preferred embodiment of the present invention, the feature extraction unit includes at least one one-dimensional convolutional layer and a pooling layer.

[0013] In another preferred embodiment of the present invention, the feature extraction unit includes a bidirectional gated loop unit.

[0014] Optionally, the attention unit is a multi-head attention mechanism module.

[0015] Optionally, the data processing module also has an online learning function, which is used to fine-tune and update the deep learning model using newly collected and confirmed sample data on the production line through a transfer learning strategy.

[0016] According to the graphite component online detection method for industrial production line provided by the application, the graphite material sample is collected in real time on the industrial production line, the collected graphite material sample is subjected to physical state adjustment pretreatment, the pretreated sample is subjected to spectrum scanning to obtain spectrum data related to graphite purity, the spectrum data is processed by using a deep learning model, the processing includes extracting local features reflecting local changes of the spectrum from the spectrum data, performing weighted processing on the local features to focus on key features related to graphite purity, and calculating a graphite purity value based on the weighted processed features, and the calculated graphite purity value is fed back to the production control system of the production line.

[0017] Compared with the prior art, the application has the following beneficial effects: 1、The pre-treatment module is arranged to ensure the consistency of the physical state of the sample entering the spectrum analysis, thereby eliminating the main interference of the non-uniformity of the solid material on the detection result from the source. In combination with the deep learning model comprising the feature extraction unit and the attention unit, the spectrum features related to the purity can be accurately identified from the complex background noise, the second-level online detection of the graphite purity is realized, and the accuracy and efficiency of the detection are significantly improved.

[0018] 2、The feedback control module links the high-precision detection result with the production control system, can automatically adjust the production parameters in real time, effectively avoids the batch of defective products caused by quality problems, and significantly improves the automation level and product quality control capability of the production line.

[0019] 3、The physical state of the solid sample is adjusted, so that the whole system has strong stability and reliability in the real industrial environment. At the same time, through the online learning function, the system can be self-optimized, continuously adapt to the changes of the production line, and long-term maintain high precision, further enhancing the robustness of the system. BRIEF DESCRIPTION OF DRAWINGS

[0020] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings: Figure 1 A whole architecture schematic diagram of the graphite component online detection system for industrial production line provided by the embodiment of the application is provided. Figure 2 A flowchart of the graphite component online detection method for industrial production line provided by the embodiment of the application is provided. Figure 3 A structure schematic diagram of the deep learning model in the data processing module provided by the embodiment of the application is provided. Figure 4A timing diagram of signaling interaction of each module of the system provided by the embodiment of the present application is shown.

[0021] In the drawings, reference signs are explained as follows: 10-conveyor belt; 20-sampling module; 30-preprocessing module; 40-spectral analysis module; 50-data processing module; 60-feedback control module; 70-production control system; 51-data preprocessing unit; 52-feature extraction unit; 521-convolution layer; 522-pooling layer; 53-attention unit; 54-full connection layer; 55-output layer. DETAILED DESCRIPTION

[0022] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These are within the scope of the present application.

[0023] Embodiment 1 The present embodiment provides a graphite component online detection system for industrial production line based on near-infrared spectral analysis and deep learning and a corresponding method. The system can realize real-time and high-precision detection of the purity of graphite materials on the industrial production line, and form an automatic closed-loop control with the production control system.

[0024] As shown in Figure 1 , it is a whole architecture schematic diagram of the graphite component online detection system for industrial production line in one embodiment of the present application. The system is deployed on the graphite production line, for example, it can be deployed at a certain key node of the conveyor belt 10 conveying graphite powder or particles. The system mainly includes a sampling module 20, a preprocessing module 30, a spectral analysis module 40, a data processing module 50 and a feedback control module 60. These modules work cooperatively and interact with the original production control system 70 of the production line (such as programmable logic controller) to detect the results.

[0025] Specifically, the sampling module 20 is arranged above or beside the conveyor belt 10, and its function is to automatically collect a certain amount of sample from the flowing graphite material at a preset time interval or according to external instructions. As an optional implementation manner, the sampling module 20 can be a non-contact sampling device, for example, a precision mechanical arm guided by a laser ranging sensor. The mechanical arm can accurately position to the material flow on the conveyor belt 10 and quickly grab about 10 grams of graphite sample. The collected sample is then quickly sent to the preprocessing module 30 through an air transmission pipeline, which can effectively avoid cross contamination and environmental impact of the sample during transmission.

[0026] The pretreatment module 30 is connected to the outlet of the sampling module 20, and its core function is to eliminate the interference of the physical state inhomogeneity of the sample on the subsequent spectral measurement. It can be understood that in the production process of graphite, the physical properties of the material such as water content, particle size, and compaction density will change constantly, and these changes will cause the baseline of the spectrum to drift and the scattering effect, which is one of the main reasons for low online detection accuracy. Therefore, the pretreatment module 30 of the embodiment integrates two key parts of a drying unit and a homogenization unit. The drying unit can be a microwave drying cavity, which can quickly and uniformly heat the inside of the sample when the sample enters, and reduce the water content to a very low stable level, such as below 0.1%, within a short time (for example, 1-2 minutes). The homogenization unit can be a container with a high-speed rotating blade or stirring paddle, i.e. a high-speed homogenizer, which can rotate at a speed of not less than 2000 revolutions per minute. After the dried sample enters the homogenization unit, it is broken and mixed through high-speed shearing and collision, so as to form a sample to be measured with relatively uniform particle size distribution and compaction density. Through such physical state adjustment and processing, it can be ensured that the sample sent into the spectral analysis module 40 each time has highly consistent physical properties, laying a solid foundation for obtaining high-quality and highly repeatable spectral data.

[0027] The spectral analysis module 40 is connected to the outlet of the pretreatment module 30, and is responsible for performing spectral scanning on the treated sample. In an embodiment of the application, the core component of the spectral analysis module 40 is a spectral detection device, which can be a near-infrared spectrometer. The detection wavelength range of the spectrometer is set to 800 nanometers to 2500 nanometers, and the spectral resolution is 1 nanometer. It should be noted that this wavelength range covers the frequency doubling and frequency mixing absorption regions of various chemical bonds such as carbon-hydrogen bonds and oxygen-hydrogen bonds, and the spectral characteristics of these regions are highly related to the carbon content in graphite and the content of trace impurities (such as organic matter and water residue). The homogenized sample is automatically sent into the reflective sample cell of the spectrometer. In order to further improve the representativeness of the measurement, the spectrometer will perform multi-point scanning on the sample in the sample cell at least at three different positions, and average the spectral data obtained by multiple scanning to obtain the final original absorption spectrum data. The entire spectral acquisition process, including sample feeding, scanning, and data output, is controlled to be completed within 5 seconds, fully meeting the real-time requirements of online detection.

[0028] The data processing module 50 is the core computing unit of the system, which is responsible for receiving and processing the spectral data transmitted by the spectral analysis module 40, and calculating the purity value of the graphite accordingly. In terms of hardware configuration, the data processing module 50 can be a high-performance industrial computer, which is internally configured with a graphics processing unit to accelerate the complex operation of the subsequent deep learning model. Please refer to Figure 3which shows the structure of the deep learning model adopted inside the data processing module 50. The complete workflow of the data processing module 50 can be described as follows: Firstly, before inputting the spectral data into the deep learning model, a series of mathematical processing need to be performed by the data preprocessing unit 51, which is crucial for improving the performance of the model. In this embodiment, the steps performed by the data preprocessing unit 51 include but are not limited to: 1. Smoothing the original spectral data, for example, a Savitzky-Golay filter with a window size of 11 and a polynomial order of 2 can be used to filter out random noise in the spectral signal; 2. Normalization, for example, minimum-maximum normalization, which scales the absorbance value of each point in the spectral vector to a fixed interval (such as [0, 1]), to eliminate the overall signal intensity fluctuations caused by factors such as optical path changes; 3. Wavelet transform denoising, for example, the Daubechies wavelet basis function can be used to decompose and reconstruct the signal for 3 layers, which can effectively remove noise while maximizing the preservation of the true shape information of the spectral peaks; 4. Polynomial baseline correction, for example, a 3rd order polynomial can be used to fit the baseline of the spectrum and subtract it from the original spectrum to correct the baseline drift caused by physical factors such as particle scattering.

[0029] After the above preprocessing, the clean and standardized spectral data is input into a specially designed deep learning model for analysis. The model is mainly composed of a feature extraction unit 52 and an attention unit 53.

[0030] Among them, the function of the feature extraction unit 52 is to automatically learn and extract deep local features that can reflect the local changes of the spectrum from the serialized spectral data. In this embodiment, the feature extraction unit 52 is composed of multiple one-dimensional convolution layers 521 and pooling layers 522 stacked alternately. Specifically, the spectral data (which can be regarded as a one-dimensional vector) is first passed through the first one-dimensional convolution layer (for example, using 32 convolution kernels with a size of 5), which is equivalent to sliding scanning the entire spectrum with a set of learnable filters to capture basic patterns in the spectrum, such as the rising edge, falling edge and peak shape of the absorption peak. The result of convolution is then passed through a max pooling layer, which reduces the data dimension by taking the maximum value in the local region, while retaining the most significant features and giving the model invariance to small shifts in the spectral peak position. Subsequently, the data continues to pass through the second one-dimensional convolution layer (for example, using 64 convolution kernels with a size of 5) and another max pooling layer to learn and combine the simple features extracted by the previous layer to form more complex and abstract combined features. Through such a hierarchical structure, the feature extraction unit 52 can effectively convert the subtle changes in the original spectral image that are difficult for the human eye to distinguish into a set of high-dimensional feature vectors.

[0031] The attention unit 53 is connected immediately after the feature extraction unit 52, which is to weight the extracted local feature sequence so that the model can focus on the key features most relevant to the purity of graphite. In the present embodiment, the attention unit 53 adopts a multi-head attention mechanism module, for example, an attention module containing 4 “heads”. Each “head” independently learns a set of query, key, and value weight matrices, and calculates the correlation scores between different positions in the feature sequence, and then weights and sums the features according to these scores. That is, it is equivalent to letting the model examine the feature sequence from multiple different subspaces and angles, and automatically determine which wavebands (i.e. which features) contribute most to the final purity prediction. For example, the model may automatically learn that the spectral features related to carbon-hydrogen bonds near 1650 nanometers should have much higher weights than the features in other noise regions. The multi-head mechanism concatenates and linearly transforms the weighted results of each “head” to obtain the final weighted feature representation.

[0032] Finally, the feature vector after the attention unit 53 weighting processing is flattened and sent to the regression part composed of several fully connected layers 54. For example, two fully connected layers can be set, the first layer has 128 neurons, and the second layer has 1 neuron, i.e. the output layer 55. The output layer 55 finally outputs a single continuous value, which is the model's predicted graphite purity value (for example, a value between 0 and 100).

[0033] The feedback control module 60 is connected with the data processing module 50 and the production control system 70, and is used to receive the purity value calculated by the data processing module 50. In the present embodiment, the feedback control module 60 can be an interface card supporting standard industrial communication protocols (such as Modbus protocol). The module is internally set with one or more purity thresholds, for example, the qualified line is 99.5%. When the calculated purity value is lower than the threshold, the feedback control module 60 will immediately generate an alarm signal. The alarm signal can drive the on-site audible and visual alarm to prompt the operator to pay attention on the one hand; on the other hand, more importantly, it will send a specific control instruction to the production control system 70 through the aforementioned communication protocol. Correspondingly, the production control system 70 can automatically execute the preset adjustment operation after receiving the instruction, for example, adjusting the addition amount of a certain chemical reagent in the upstream purification process, or changing the temperature parameters of a certain reaction furnace, so as to realize real-time and automatic adjustment of the production process, and nip the quality problem in the bud. The detection result is also displayed synchronously on the industrial touch screen on site for the operator to check.

[0034] Please refer to Figure 2Fig. 1 shows a complete flow of an industrial production line-oriented graphite composition online detection method provided by an embodiment of the present application. The method corresponds to the above-mentioned system and specifically includes the following steps: Step S101, real-time collection of graphite samples, i.e., the system automatically collects materials on the conveying belt 10 through the sampling module 20. Step S102, sample pretreatment, wherein the collected samples are sent to the pretreatment module 30 for drying and homogenization treatment to obtain physically uniform samples to be tested. Step S103, spectral scanning, i.e., the pretreated samples are scanned in the spectral analysis module 40 to obtain their near-infrared spectral data. Step S104, spectral data pretreatment, wherein the data pretreatment unit 51 in the data processing module 50 performs operations such as smoothing, normalization, denoising, and baseline correction on the original spectral data. Step S105, purity calculation through the model, i.e., the pretreated spectral data is input into the deep learning model, sequentially passes through the feature extraction unit 52 and the attention unit 53, and finally the purity value is calculated by the output layer 55. Step S106, purity judgment, i.e., the feedback control module 60 compares the calculated purity value with the preset threshold value. Step S107, feedback control, i.e., according to the judgment result, the feedback control module 60 sends the purity value, the alarm signal, or the control instruction to the production control system 70 and can trigger the corresponding automatic control action.

[0035] For further reference Figure 4 Fig. 2 shows a timing diagram of signaling interaction of modules of the system, which clearly shows the dynamic process of the entire closed-loop control. Specifically, after the sampling module (S) completes sampling, it sends a “sampling complete” signal to the data processing module (D). In response, the data processing module (D) immediately instructs the spectrometer (G) to start scanning. After the spectrometer (G) completes scanning, it returns the spectral data to the data processing module (D). The data processing module (D) then activates the internal data processing and model calculation process. After the calculation is completed, the data processing module (D) sends the results (such as the purity value or the alarm judgment) to the feedback control module (F). The feedback control module (F) then converts this information into control instructions or alarm signals that the production control system (P) can understand, thereby completing a complete online detection and feedback control cycle.

[0036] Through the system and method provided by the present embodiment, it is verified through a large number of experiments (for example, through 10-fold cross-validation) that the average prediction accuracy of the deep learning model can reach 98.5%, and the root mean square error between the prediction result and the chemical analysis reference value is less than 0.5%. The entire detection cycle from sampling to outputting the result can be completed within 3 minutes, realizing the online detection of graphite purity at the second level, and successfully integrating with the production line control system, realizing real-time response and automatic closed-loop control of quality problems.

[0037] Embodiment 2 This embodiment aims to illustrate that the feature extraction unit 52 in the data processing module 50 in the technical scheme of the present application is not limited to a specific implementation manner. As a variant, this embodiment provides a feature extraction unit adopting a recurrent neural network variant structure.

[0038] In this embodiment, the overall architecture of the graphite component online detection system for industrial production lines, including the hardware composition and functions of the sampling module 20, the pretreatment module 30, the spectrum analysis module 40, and the feedback control module 60, are basically the same as those described in Embodiment 1. The core difference of this embodiment lies in the internal structure of the deep learning model adopted in the data processing module 50.

[0039] Specifically, the feature extraction unit 52 no longer uses a one-dimensional convolutional network, but is replaced by a recurrent neural network structure that can better capture the dependence relationship before and after the sequence data. As a preferred implementation manner, this embodiment adopts a two-layer stacked bidirectional gated recurrent unit as the feature extraction unit.

[0040] In the working process, the steps of sampling, sample pretreatment, spectrum acquisition, and spectrum data pretreatment (S101 to S104) are completely consistent with Embodiment 1. The spectrum data vector processed by the data pretreatment unit 51 is regarded as a time sequence, in which each wavelength point corresponds to a time step in the sequence. The sequence is input into the bidirectional gated recurrent unit network.

[0041] It can be understood that the gated recurrent unit is a kind of recurrent neural network, which solves the gradient vanishing problem in traditional recurrent neural networks through internal “update gate” and “reset gate”, and can effectively learn the long-term dependence relationship in the sequence. When the spectrum sequence is input into the gated recurrent unit, the unit updates its internal hidden state at each wavelength point by combining the input information of the current wavelength point and the hidden state transmitted from the previous wavelength point. This process is carried out along the spectrum from low wavelength to high wavelength (or vice versa), so that the hidden state at each wavelength point contains information summaries of all previous wavelength points.

[0042] The “bidirectional” structure means that data is processed in two directions at the same time: one gated recurrent unit layer processes from the starting wavelength (such as 800 nm) to the ending wavelength (such as 2500 nm), and the other gated recurrent unit layer processes from the ending wavelength to the starting wavelength in reverse. At each wavelength point, the hidden states of the two directions are spliced together to form the final feature representation of the point. The advantage of this is that for any absorption peak in the spectrum, the model can not only consider the information of the baseline and adjacent peaks before it, but also consider the spectrum information after it, thereby obtaining a more comprehensive understanding of the feature.

[0043] The hidden state sequence output by the bi-directional gated recurrent unit network (whose dimension is the same as the input spectrum sequence length, but the number of channels becomes the number of hidden units of the gated recurrent unit) also constitutes a set of high-dimensional local features. This feature sequence is then sent to the same attention unit 53 (e.g., a multi-head attention mechanism module) as in Embodiment 1 for weighted processing to focus on the band features that are most important for purity prediction. The subsequent steps, including purity regression calculation by the fully connected layer 54 and closed-loop control by the feedback control module 60, are also exactly the same as in Embodiment 1.

[0044] This embodiment demonstrates that using a recurrent neural network and its variants (such as gated recurrent units or long short-term memory networks) as a feature extraction unit can also effectively extract sequence features related to graphite purity from spectral data and achieve high-precision prediction. This fully supports the generalization of the "feature extraction unit" functional module in the technical solution of the present application, indicating that any neural network structure that can extract local or context-related features from sequence data can fall within the protection concept of the present application. The overall performance of the system is comparable to that of Embodiment 1, and can also achieve high accuracy and high efficiency of online detection.

[0045] Embodiment 3 This embodiment aims to illustrate that the spectral detection device configured by the spectral analysis module 40 in the technical solution of the present application is not limited to a specific type, and provides a variant scheme using Raman spectroscopy as a detection means.

[0046] In this embodiment, the overall structure of the system, including the sampling module 20, the preprocessing module 30, the data processing module 50, and the feedback control module 60, has a basic composition and connection relationship similar to that of Embodiment 1. The core difference lies in the spectral analysis module 40.

[0047] Specifically, the spectral detection device in the spectral analysis module 40 is replaced by a near-infrared spectrometer with a Raman spectrometer. For example, a confocal Raman spectrometer equipped with a 785 nm excitation wavelength laser can be used. It should be noted that Raman spectroscopy measures the inelastic scattering light that occurs after the interaction of laser light with the molecular vibration or rotational energy level of the sample to obtain the chemical structure and lattice structure information of the substance. For carbon materials, Raman spectroscopy is extremely sensitive to the integrity of the crystal structure, defect density, and graphitization degree, and these characteristics are closely related to the purity of graphite.

[0048] Correspondingly, the working process of the system is also adjusted adaptively. The sampling step S101 and the sample pretreatment step S102 are the same as in Embodiment 1, that is, the collected sample is dried and homogenized to ensure the consistency of the physical state of the sample for submission. This is also important for obtaining high-quality Raman spectra, as it can reduce the interference of fluorescence background and uneven heating effects.

[0049] In the spectral scanning step S103, the pretreated sample is automatically placed on the sample stage of the Raman spectrometer. A 785-nanometer laser beam is focused on the sample surface, exciting Raman scattering signals. The detector of the spectrometer collects these scattered lights and generates a Raman spectrum. The abscissa of the Raman spectrum is usually expressed in Raman shift (unit: cm -1 ).

[0050] The deep learning model in the data processing module 50 can have the same basic architecture as in Embodiment 1, that is, it still uses a model composed of a feature extraction unit 52 (e.g., a one-dimensional convolutional network) and an attention unit 53 (e.g., a multi-head attention mechanism). However, this model needs to be retrained or transferred using a large amount of Raman spectrum data related to graphite purity, so that it can understand and analyze the features of the Raman spectrum.

[0051] In the model processing process, the focus of its analysis is different from that of near-infrared spectroscopy. For the Raman spectrum of graphite, there is usually a D peak (representing disorder induction) at about 1350 cm -1 , the intensity of which is proportional to the defects and disorder degree in the lattice; there is a G peak (representing graphite structure) at about 1580 cm -1 , corresponding to the in-plane vibration of sp 2 carbon atoms in the ideal graphite lattice. The feature extraction unit 52 of the model will learn how to accurately capture the peak shape features of these two key peaks, such as the peak position, peak height, peak width (full width at half maximum), and intensity ratio of the D peak and the G peak. The subtle changes in these parameters all contain important information about the purity and crystalline quality of graphite. The attention unit 53 will learn to assign higher weights to the spectral region where the D peak and the G peak are located, while ignoring the background noise in other regions. Finally, the fully connected layer 54 will map these deep features extracted from the Raman spectrum and weighted to the final graphite purity value.

[0052] The subsequent purity judgment step S106 and feedback control step S107 are exactly the same as in Embodiment 1.

[0053] The embodiment shows that the overall technical framework of "pretreatment + spectral analysis + deep learning (feature extraction + attention) + feedback control" proposed by the application has good universality. It is not only suitable for near-infrared spectroscopy, but also can be successfully applied to Raman spectroscopy and other spectral analysis technologies that can reflect material composition or structure information. This strongly supports the protection scope of the "spectral detection device" in the application scheme, making the system not only able to detect purity, but also to monitor the crystalline quality of graphite simultaneously, providing more rich quality control dimensions for the production process.

[0054] Embodiment 4 The embodiment further elaborates the adaptive optimization and predictive analysis functions of the system based on embodiment 1, aiming to further improve the robustness, intelligence level and practical value of the system in long-term industrial operation. The hardware structure of the system is exactly the same as that of embodiment 1, and the special feature lies in the software function expansion of the data processing module 50 and the feedback control module 60.

[0055] In this embodiment, in addition to the deep learning model for purity calculation, an online learning function module is integrated into the data processing module 50. This function aims to solve the "working condition drift" problem that may occur in the industrial field, such as raw material batch replacement, production equipment aging, seasonal changes in environmental temperature and humidity, etc. These factors may cause slow and continuous changes in the spectral characteristics of graphite samples, thereby reducing the prediction accuracy of the initially trained model over time.

[0056] The specific workflow of the online learning function is as follows: 1. Data accumulation: the system will store a portion of the online collected sample spectrum data and its corresponding model predicted purity value periodically (e.g., every hour) while performing routine online detection. At the same time, the operator will periodically (e.g., every day) extract a small amount of samples from the production line to the laboratory for determination of their true purity using high-precision chemical analysis and other reference methods. These spectrum data pairs (spectrum data + true purity value) with "true labels" are stored in a special database for training. 2. Trigger and update: the system has a built-in online learning trigger. When the number of new samples accumulated in the training database reaches a preset threshold (e.g., 50 new effective samples are accumulated), or at fixed time intervals (e.g., every 24 hours), the trigger will automatically start the fine-tuning update program of the model. 3. Transfer learning strategy: in order to achieve rapid and efficient updating, the system adopts the strategy of transfer learning, rather than completely retraining the model. Specifically, the system will load the current deep learning model in use, and "freeze" the weights of most of the bottom convolutional layers 521 in the feature extraction unit 52. It can be understood that these bottom network learns relatively general spectral basic features (such as peaks and valleys), which are relatively stable under different working conditions. Then, the system only uses the newly accumulated data to retrain the higher layer network of the model, especially the weights of the attention unit 53 and the fully connected layer 54. This process has small calculation amount and short time consumption (usually within a few minutes), but can make the model quickly learn the subtle change rules of the spectrum under the new working condition, so as to adapt to the new data distribution, restore and maintain its high prediction accuracy. The updated model will automatically replace the old model and be put into subsequent online detection.

[0057] In addition, the feedback control module 60 in the present embodiment also adds the function of historical data trend analysis to realize predictive maintenance and quality early warning. Specifically, the feedback control module 60 not only processes the current single-point detection result, but also continuously records and analyzes the sequence of purity detection results in the past period of time (for example, the past 24 hours), and calculates the moving average and standard deviation of the sequence in real time. If the system monitors that the moving average of the purity value presents a continuous downward trend, even if each purity value itself is still within the qualified range (for example, slowly decreases from 99.8% to 99.6%), the system will generate an “early warning of quality trend” signal in advance. The early warning signal will be sent to the control room through the on-site human-machine interface or to the management personnel and operators, prompting them to pay attention to the potential fluctuations of the upstream raw material batch or related process parameters. This predictive function based on trend analysis can help enterprises discover and solve problems earlier, thereby avoiding the production of batches of substandard products to the greatest extent, and significantly reducing production risks and quality costs, and upgrading the quality control from “post-alarm” (i.e., alarm after detecting unqualified products) to a higher level of “prevention”.

[0058] Through the online learning and predictive analysis functions added by the present embodiment, the graphite composition online detection system for industrial production lines provided by the present application is not only a static detection tool, but also an intelligent quality control system that can evolve itself and actively warn, greatly enhancing its long-term stability and application value in complex and variable real industrial environments.

[0059] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module, unit thereof in the form of pure computer readable program code, the same functions can also be realized by logically programming the method steps in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers. Therefore, the system provided by the present application and each device, module, unit thereof can be considered as a hardware component, and the devices, modules, units included therein for realizing various functions can also be considered as structures within the hardware component; the devices, modules, units for realizing various functions can also be considered as both software modules realizing methods and structures within hardware components.

[0060] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other without conflict.

Claims

1. An online graphite composition detection system for industrial production lines, characterized in that, include: The system includes a sampling module, a preprocessing module, a spectral analysis module, a data processing module, and a feedback control module. The sampling module is used to collect graphite material samples in real time on the industrial production line; The preprocessing module is connected to the sampling module and is used to perform physical state adjustment processing on the collected graphite material samples; The spectral analysis module is equipped with a spectral detection device for performing spectral scanning on the pretreated sample to obtain spectral data related to graphite purity. The data processing module is used to receive and process the spectral data; The feedback control module is connected to the data processing module and the production control system of the production line, and is used to feed back the calculated graphite purity value to the production control system.

2. The online graphite composition detection system for industrial production lines according to claim 1, characterized in that, The pretreatment module includes a drying unit and a homogenization unit; The drying unit is used to dry the graphite material sample to reduce the moisture content of the sample. The homogenization unit is used to homogenize the graphite material sample to form a test sample with uniform physical state.

3. The online graphite composition detection system for industrial production lines according to claim 1, characterized in that, The deep learning model used in the data processing module includes a feature extraction unit and an attention unit; The feature extraction unit is used to extract local features reflecting local spectral changes from the spectral data; The attention unit is connected after the feature extraction unit and is used to perform weighted processing on local features to focus on key features related to graphite purity, and to calculate the graphite purity value based on the weighted features.

4. The online graphite composition detection system for industrial production lines according to claim 1, characterized in that, The spectral detection equipment is a near-infrared spectrometer or a Raman spectrometer; When the spectral detection device is a near-infrared spectrometer, its detection wavelength range is 800-2500nm.

5. The online graphite composition detection system for industrial production lines according to claim 1, characterized in that, Before inputting the spectral data into the deep learning model, the data processing module is further configured to perform at least one of the following preprocessing steps on the spectral data: Normalization; wavelet transform denoising; polynomial baseline correction.

6. The online graphite composition detection system for industrial production lines according to claim 3, characterized in that, The feature extraction unit includes at least one one-dimensional convolutional layer and a pooling layer.

7. The online graphite composition detection system for industrial production lines according to claim 3, characterized in that, The feature extraction unit includes a bidirectional gated loop unit.

8. The online graphite composition detection system for industrial production lines according to claim 3, characterized in that, The attention unit is a multi-head attention mechanism module.

9. The online graphite composition detection system for industrial production lines according to claim 1, characterized in that, The data processing module also has an online learning function, which is used to fine-tune and update the deep learning model using newly collected and confirmed sample data on the production line through a transfer learning strategy.

10. A method for online detection of graphite composition in industrial production lines, characterized in that, Includes the following steps: Step S1: Collect graphite material samples in real time on the industrial production line; Step S2: Perform physical state adjustment pretreatment on the collected graphite material samples; Step S3: Perform a spectral scan on the pretreated sample to obtain spectral data related to graphite purity; Step S4: Process the spectral data using a deep learning model. The processing includes: extracting local features reflecting local spectral changes from the spectral data; weighting the local features to focus on key features related to graphite purity; and calculating the graphite purity value based on the weighted features. Step S4: Feed back the calculated graphite purity value to the production control system of the production line.

Images