Tail coal ash detection method and system based on online class incremental learning

Abstract

The application discloses a tail coal ash content detection method and system based on online class incremental learning, and the method comprises the following steps: acquiring a coal slime mixed liquid data set and performing data preprocessing to obtain a two-dimensional spectrum image data of the coal slime mixed liquid; introducing an online class incremental learning mechanism, constructing a tail coal ash content detection network model based on class incremental learning based on a backbone neural network model; introducing a herding strategy, performing tail coal ash content detection on the two-dimensional spectrum image data of the coal slime mixed liquid based on the tail coal ash content detection network model based on class incremental learning, and obtaining a tail coal ash content detection result of the coal slime mixed liquid. The application can resist the catastrophic forgetting of the model in the training process and reduce the time of the model, thereby improving the tail coal ash content detection precision of the coal slime mixed liquid. The application can be widely applied to the technical field of coal slime flotation ash content detection as a tail coal ash content detection method and system based on online class incremental learning.

Description

Technical Field

[0001] This invention relates to the field of coal slime flotation ash content detection technology, and in particular to a method and system for detecting tailings ash content based on online incremental learning. Background Technology

[0002] Coal slime flotation is an important method in coal production. The ash content information of tailings not only reflects the ash content of the clean coal but also its recovery rate. Current technologies for tailings ash content detection focus on eliminating manual coal preparation and improving detection accuracy. For example, they rely on machine vision or spectral data to build specific deep learning models to achieve detection. However, current tailings ash content detection methods firstly fail to consider the conceptual drift caused by changes in various factors in the industrial environment, such as the properties of coal slime. Secondly, existing methods can only detect coal samples with a limited ash category. When a new category of coal sample is present, the original model must be discarded, and a new model must be built by collecting new data and combining it with the existing dataset. In other words, current research methods for tailings ash content detection in coal slime flotation do not consider the problem of ash label invalidation or mismatch caused by factors such as coal slime properties, particle size distribution, and mineral impurities, resulting in new data being used to train the model with historical data. Furthermore, the arrival of new data also requires integration with valid old data and retraining the model, consuming more computational resources. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a tailings ash content detection method and system based on online incremental learning, which can resist catastrophic forgetting during model training and reduce model time, thereby improving the accuracy of tailings ash content detection in coal slime mixture.

[0004] The first technical solution adopted in this invention is: a method for detecting tailings ash content based on online incremental learning, comprising the following steps:

[0005] The coal slime mixture dataset was acquired and preprocessed to obtain two-dimensional spectral image data of the coal slime mixture.

[0006] An online incremental learning mechanism is introduced, and a tailings ash content detection network model based on incremental learning is constructed based on the backbone neural network model.

[0007] A herding strategy is introduced, and a tailings ash content detection network model based on incremental learning is used to detect tailings ash content in two-dimensional spectral image data of coal slime mixture, so as to obtain the tailings ash content detection results of coal slime mixture.

[0008] Furthermore, the step of acquiring the coal slime mixture dataset and performing data preprocessing to obtain the two-dimensional spectral image data of the coal slime mixture specifically includes:

[0009] Several groups of coal slime mixtures with different concentrations and different ash contents were obtained;

[0010] A peristaltic pump was used to drive the coal slime mixture to form a liquid flow, and a flow-through cuvette observation window was set up to collect data, resulting in a coal slime mixture image dataset.

[0011] By combining a fiber optic spectrometer and a xenon lamp light source, absorption spectral data of coal slime mixture images were acquired to obtain the spectral time-series data of the coal slime mixture images.

[0012] The spectral time-series data of the coal slime mixture image was converted into two-dimensional image data using the Gram angle field method, thus obtaining two-dimensional spectral image data of the coal slime mixture.

[0013] Furthermore, the step of introducing a herding strategy and using a class-incremental learning tailings ash detection network model to detect tailings ash content in the two-dimensional spectral image data of the coal slime mixture, and obtaining the tailings ash content detection result of the coal slime mixture, specifically includes:

[0014] The two-dimensional spectral image data of the coal slime mixture were segmented to obtain several classification datasets;

[0015] The first classification dataset is input into the backbone neural network model based on the class incremental learning tailings ash content detection network model for training, and the trained backbone neural network model is obtained.

[0016] Weight and bias data are obtained based on the trained backbone neural network model;

[0017] The incremental learning tailings ash content detection network model is updated based on the weight data and bias data to obtain the updated incremental learning tailings ash content detection network model.

[0018] A representative dataset is obtained by selecting samples from the first classification dataset using a herding strategy;

[0019] The second classification dataset and the representative dataset are combined and input into the updated class incremental learning tailings ash detection network model for training until all classification datasets are traversed to obtain the final class incremental learning tailings ash detection network model.

[0020] The tailings ash content detection network model based on the final incremental learning method is used to detect the tailings ash content in the two-dimensional spectral image data of the coal slime mixture, and the tailings ash content detection results of the coal slime mixture are obtained.

[0021] Furthermore, the step of inputting the first classification dataset into the backbone neural network model of the tailings ash content detection network model based on class incremental learning for training, to obtain the trained backbone neural network model, specifically includes:

[0022] The first classification dataset is input into the backbone neural network model of the tailings ash content detection network model based on class incremental learning. The backbone neural network model includes an input layer, an initial convolutional layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, and a fully connected layer.

[0023] The first classification dataset is obtained based on the input layer of the backbone neural network model;

[0024] Based on the initial convolutional layer of the backbone neural network model, preliminary feature extraction processing is performed on the first classification dataset to obtain the low-level features of the first classification dataset.

[0025] Based on the first residual layer, second residual layer, third residual layer and fourth residual layer of the backbone neural network model, deep feature extraction processing is performed on the low-level features of the first classification dataset to obtain the deep features of the first classification dataset.

[0026] Based on the fully connected layer of the backbone neural network model, linear transformation and mapping are performed on the deep features of the first classification dataset to obtain the detection results of the first classification dataset.

[0027] The detection accuracy of the detection results for the first classification dataset is judged;

[0028] If the detection accuracy of the detection results of the first classification dataset does not meet the preset accuracy threshold, the first classification dataset is re-inputted into the backbone neural network model for training until the detection accuracy of the detection results of the first classification dataset meets the preset accuracy threshold.

[0029] If the detection accuracy of the detection results of the first classification dataset meets the preset accuracy threshold, then the trained backbone neural network model is output.

[0030] Furthermore, the step of updating the incremental learning tailings ash content detection network model based on weight data and bias data to obtain the updated incremental learning tailings ash content detection network model specifically includes:

[0031] For the backbone neural network model, generate a new fully connected layer;

[0032] Copy the number of output categories of the fully connected layer in the original backbone neural network model to the new fully connected layer;

[0033] The new fully connected layer is assigned values ​​based on the weight data and bias data to obtain the assigned fully connected layer.

[0034] The newly assigned fully connected layer is replaced with the original fully connected layer to obtain the updated incremental learning tailings ash content detection network model.

[0035] Furthermore, the specific expression for updating the class-incremental learning tailings ash content detection network model based on weight data and bias data is as follows:

[0036] ;

[0037] In the above formula, This indicates a new fully connected layer. Indicates the feature dimension.

[0038] Furthermore, the step of selecting samples from the first classification dataset using the herding strategy to obtain a representative dataset specifically includes:

[0039] Obtain the current set of embedded category centers for the first classification dataset;

[0040] Calculate the distance from each instance in the first classification dataset to the embedding class center of the current set, and sort them in ascending order to obtain the sorted distance results;

[0041] The distance results after sorting are selected based on a preset number of representative samples to obtain a representative dataset.

[0042] Furthermore, the expression for obtaining the current set embedding category center of the first classification dataset is as follows:

[0043] ;

[0044] In the above formula, Indicates the category of the first classification dataset. This represents the total number of instances in the first category dataset. This represents the first classification dataset. Indicates the first One example, This indicates that the current set is embedded in the category center.

[0045] The second technical solution adopted in this invention is: a tailings ash content detection system based on online incremental learning, comprising:

[0046] The first module is used to acquire the coal slime mixture dataset and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture.

[0047] The second module is used to introduce an online incremental learning mechanism and construct a tailings ash content detection network model based on incremental learning, based on the backbone neural network model.

[0048] The third module introduces the herding strategy and uses a class-incremental learning tailings ash detection network model to detect tailings ash content in two-dimensional spectral image data of coal slime mixture, thus obtaining the tailings ash content detection results of coal slime mixture.

[0049] The beneficial effects of the method and system of this invention are as follows: This invention obtains a coal slime mixture dataset and performs data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture. Then, an online incremental learning mechanism is introduced. Based on the backbone neural network model, a tailings ash content detection network model based on incremental learning is constructed. Data is continuously input into the incremental learning model in the form of a data stream, and the model is continuously updated. Through the online incremental learning mechanism, incremental training can be performed on the dataset of the second coal sample while saving the weight data and classification labels of the first coal sample training. This avoids the model starting from scratch and shortens the model training time. Furthermore, a herding strategy is introduced. The tailings ash content of the coal slime mixture is detected based on the two-dimensional spectral image data of the coal slime mixture using the incremental learning tailings ash content detection network model. The herding strategy selects a small portion of representative samples from the subset of previous task data and puts them into the current training dataset. This can resist catastrophic forgetting during the model training process while ensuring that the memory cost is not too high, thereby improving the accuracy of tailings ash content detection in coal slime mixture. Attached Figure Description

[0050] Figure 1 This is a flowchart of the steps of a tailings ash content detection method based on online incremental learning according to the present invention;

[0051] Figure 2 This is a structural block diagram of a tailings ash content detection system based on online incremental learning according to the present invention;

[0052] Figure 3 This is a schematic diagram of a tailings ash content detection model based on online incremental learning provided in a specific embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of the backbone neural network model provided in a specific embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of online incremental learning training provided in a specific embodiment of the present invention.

[0055] Figure reference numerals: 1. Input layer; 2. Initial convolutional layer; 3. First residual layer; 4. Second residual layer; 5. Third residual layer; 6. Fourth residual layer; 7. Fully connected layer. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0057] First, the technical terms used in the embodiments of this invention need to be explained:

[0058] 1) Coal slime flotation: Coal slime is a byproduct produced during the coal washing process. It is a semi-solid substance formed by mixing coal powder and water. However, the quality of coal varies greatly, resulting in significant differences in the quality and usability of coal slime. There are many types of coal slime with a wide range of applications.

[0059] 2) Class-Incremental Learning: Class-incremental learning is a machine learning paradigm, also known as continuous learning, that allows a model to retain memories of previously learned classes while continuously learning new classes. This learning method simulates the human ability to retain old memories when learning new things. In class-incremental learning, the model can only access the class of the current task in each task, and cannot access the classes of previous tasks. This requires the model to effectively learn new classes without forgetting old ones. Class-incremental learning is defined as aiming to learn new features from a continuously updated data stream and resist catastrophic forgetting of old features. Assume there is... A sequence of training tasks with no overlapping categories. ,in It refers to the first The dataset for each incremental task It is a category An instance of data, It is a task The tag space, in which We only have training tasks Only when can the data set be accessed The ultimate goal of incremental learning is to build a complete classification and recognition model for all data categories. That is, the model cannot accurately identify the current task set. The model can accurately identify the categories corresponding to previous tasks. After each stage of the task, the model will evaluate the performance based on all the visible categories that have been trained.

[0060] 3) Concept Drift: Concept drift is an important concept in machine learning and data mining. It describes the phenomenon that the relationship between input data and the target variable (label) changes over time. This change means that a model learned from historical data may no longer be effective in predicting new data in the future because the relationship between the input and output data has changed.

[0061] Reference Figure 1 This invention provides a method for detecting the ash content of tailings coal based on online incremental learning, the method comprising the following steps:

[0062] S100. Obtain the coal slime mixture dataset and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture;

[0063] Specifically, several groups of coal slime mixtures with different concentrations and ash contents were obtained; the coal slime mixtures were driven by a peristaltic pump to form a liquid flow and a flow-through cuvette observation window was set to collect data, resulting in a coal slime mixture image dataset; the absorption spectrum data of the coal slime mixture image dataset was collected by combining a fiber optic spectrometer and a xenon lamp light source, resulting in spectral time-series data of the coal slime mixture images; the spectral time-series data of the coal slime mixture images were converted into two-dimensional image data using the Gram angle field method, resulting in two-dimensional spectral image data of the coal slime mixture.

[0064] In this embodiment, 110 groups of coal slime mixtures with different concentrations and ash contents were first prepared to simulate a flotation process, comprising 11 different ash types and 10 concentrations for each ash type. A peristaltic pump was then used to drive the mixture into a flow path, with a flow-through cuvette observation window installed along the flow path for data acquisition. A miniature UV-Vis fiber optic spectrometer paired with a xenon lamp was then used to continuously acquire absorption spectral data of the suspension flow. Finally, the acquired absorption spectral time-series data was preprocessed, and Gram angle field technology was applied to convert the time-series data into two-dimensional image data. The converted data was then shuffled and divided into multiple task subsets for subsequent training.

[0065] Additionally, it should be noted that when used in industrial settings, a branch line can be installed next to the flotation machine, and a fiber optic spectrometer can be used to collect real-time spectral data of the liquid flow on the branch line.

[0066] S200: Introducing an online incremental learning mechanism, and constructing a tailings ash content detection network model based on incremental learning based on the backbone neural network model;

[0067] S300. Introducing the herding strategy, the tailings ash content detection network model based on incremental learning is used to detect the tailings ash content in the two-dimensional spectral image data of the coal slime mixture, and the tailings ash content detection results of the coal slime mixture are obtained.

[0068] First, it should be noted that, as Figure 3 As shown, the total dataset is shuffled and divided into multiple task data subsets, where subset 1 = subset 1. After initializing the model (in the PyTorch architecture, the weights and biases of the model are automatically initialized), subset 1 is input into the model for initial training. If the model performance does not meet the expected standard after the first training, training needs to be stopped and readjusted. If the model performance meets or exceeds the expected standard, the corresponding model parameter weights are stored for later use, and the model enters the incremental training phase.

[0069] S310. The two-dimensional spectral image data of the coal slime mixture is divided and processed to obtain several classification datasets;

[0070] S320. Input the first classification dataset into the backbone neural network model based on the class incremental learning tailings ash content detection network model for training, and obtain the trained backbone neural network model.

[0071] Specifically, the first classification dataset is input into a backbone neural network model based on a class-incremental learning tailings ash content detection network model. This backbone neural network model includes an input layer 1, an initial convolutional layer 2, a first residual layer 3, a second residual layer 4, a third residual layer 5, a fourth residual layer 6, and a fully connected layer 7. Based on the input layer of the backbone neural network model, the first classification dataset is obtained. Based on the initial convolutional layer of the backbone neural network model, preliminary feature extraction processing is performed on the first classification dataset to obtain low-level features of the first classification dataset. Based on the first, second, third, and fourth residual layers of the backbone neural network model, deep processing is performed on the low-level features of the first classification dataset. The deep features of the first classification dataset are obtained through layer feature extraction. Based on the fully connected layer of the backbone neural network model, linear transformation and mapping are performed on the deep features of the first classification dataset to obtain the detection results of the first classification dataset. The detection accuracy of the first classification dataset is judged. If the detection accuracy of the first classification dataset does not meet the preset accuracy threshold, the first classification dataset is re-input into the backbone neural network model for training until the detection accuracy of the first classification dataset meets the preset accuracy threshold. If the detection accuracy of the first classification dataset meets the preset accuracy threshold, the trained backbone neural network model is output.

[0072] In this embodiment, as Figure 4 As shown, the incremental learning framework uses the ResNet18 network model as an example for training. The specific training process is as follows:

[0073] Input layer: Processes the input image to a standard size of 224×224 pixels.

[0074] Initial convolutional processing layer: A 7×7 convolutional kernel is used to extract a wide range of features from the input image. The feature map resolution is reduced from 224×224 to 56×56 by convolution with a stride of 2 and 3×3 max pooling. At the same time, batch normalization and ReLU activation function are combined to enhance the stability of the model and the initial feature representation ability, providing rich low-level features for subsequent layers.

[0075] The first residual layer consists of two residual blocks, each containing two 3×3 convolution operations with a kernel size of 3×3, padding of 1, a channel count of 64, and a stride of 1. It is followed by batch normalization and ReLU activation function. The input is directly added to the output through the residual connection, ensuring that while extracting shallow features, the input feature information is preserved, mitigating the gradient vanishing problem and improving the optimization stability and training efficiency of the model.

[0076] The second residual layer consists of two residual blocks, increasing the number of channels from 64 to 128. The first residual block halves the feature map size by setting the stride of the first convolution to 2, and uses 1×1 convolutions in the residual connections to adjust the number of channels, matching the input dimension to the output. The second residual block keeps the feature map size unchanged and continues with 3×3 convolutions to extract features. Each residual block includes two 3×3 convolutions, followed by batch normalization and a ReLU activation function. The residual connections preserve feature information. The main purpose of this layer is to extract deeper, intermediate-level features by downsampling and increasing the number of channels, while preserving input features to ensure training stability.

[0077] The third residual layer consists of two residual blocks, increasing the number of channels from 128 to 256. The first residual block halves the feature map size by setting the stride of the first convolution to 2, and simultaneously adjusts the number of channels within the residual connections using 1×1 convolutions to match the input dimension with the output. The second residual block maintains the same feature map size and number of channels, continuing to extract features using 3×3 convolutions. Each residual block contains two 3×3 convolutions, followed by batch normalization and a ReLU activation function after each convolution. The residual connections preserve input feature information within each block, preventing gradient vanishing. The main purpose of this layer is to extract deeper, higher-level features, further enhancing the network's ability to express complex patterns, while simultaneously improving computational efficiency by reducing the feature map resolution.

[0078] The fourth residual layer, consisting of two residual blocks, increases the number of channels from 256 to 512 and is used to extract high-level semantic features. The first residual block's first 3×3 convolution halves the feature map size with a stride of 2, while a 1×1 convolution adjusts the input and output channel counts in the residual connections to match. The second residual block maintains the same feature map size, and two more 3×3 convolutions continue to extract deeper features. Each residual block's convolution is followed by batch normalization and a ReLU activation function, directly summing the input features through the residual connections to avoid the vanishing gradient problem and retain important information. This layer, by increasing the number of channels and compressing the feature map resolution, enables the network to capture complex patterns, providing strong feature support for the final classification task.

[0079] Fully connected layer: The fully connected layer maps the 512-dimensional feature vector generated by global average pooling to the number of categories in the classification task through a linear transformation, and generates the category probability distribution through the Softmax activation function to complete the final classification output. At the same time, it reduces the number of parameters to reduce the risk of overfitting.

[0080] Finally, it should be noted that using the classic neural network ResNet18 as the backbone network has the advantage of maintaining high detection accuracy while having relatively few network parameters, making it easier to deploy on industrial field hardware platforms. Moreover, ResNet18 has a large number of pre-trained models available, which allows for rapid initiation of the training process through transfer learning in incremental learning tasks, accelerating convergence and improving initial performance.

[0081] Once the model is evaluated (i.e., the number of training epochs has been reached and the model performance meets the target), the next incremental learning training cycle can begin until all categories and tasks in the total dataset have been updated, at which point training is complete.

[0082] S330. Obtain weight data and bias data based on the trained backbone neural network model;

[0083] S340. Update the incremental learning tailings ash content detection network model according to the weight data and bias data to obtain the updated incremental learning tailings ash content detection network model.

[0084] Specifically, for the backbone neural network model, a new fully connected layer is generated; the output class number of the fully connected layer of the original backbone neural network model is copied to the new fully connected layer; the new fully connected layer is assigned values ​​according to the weight data and bias data to obtain the assigned fully connected layer; the assigned fully connected layer is replaced with the original fully connected layer to obtain the updated class incremental learning tailings ash content detection network model.

[0085] In this embodiment, the first step of the incremental training phase is to update the model, specifically the fully connected layers: retaining the weights and biases learned during previous training while allowing the model to adapt to a new number of classes. If the number of classes increases, new weights and biases will be initialized (usually with random values), while the old weights and biases remain in their original positions.

[0086] First, a new fully connected layer is generated, and its expression is:

[0087] ;

[0088] in It is the feature dimension. The number of new categories.

[0089] Assumption To find the number of output categories of the original fully connected layer, copy the original weights and biases to the beginning of the new fully connected layer. There are 1 element whose expression is:

[0090] ;

[0091] Then delete the old fully connected layer and assign the new fully connected layer to... Its expression is:

[0092] ;

[0093] The entire update process can be represented as:

[0094] ;

[0095] This process ensures that the weights and biases of the new fully connected layer are properly preserved and updated when categories are added.

[0096] After the model is updated and the herding strategy has completed the selection of representative datasets, the sub-datasets will be fed into the updated model for further training.

[0097] S350. Samples are selected from the first classification dataset using the herding strategy to obtain a representative dataset;

[0098] Specifically, the current set embedding category center of the first classification dataset is obtained; the distance from each instance in the first classification dataset to the current set embedding category center is calculated and sorted in ascending order to obtain the sorted distance results; the sorted distance results are selected according to the preset number of representative samples to obtain the representative dataset.

[0099] In this embodiment, in each incremental task, the model can generally only access the current corresponding dataset, which helps protect user privacy and reduce storage burden. However, to resist catastrophic forgetting while ensuring that memory costs are not too high, this embodiment of the invention uses a herding strategy to select a small subset of representative samples from previous task data and put them into the current training dataset: given from category dataset The herding strategy first calculates the category centers of the current set embeddings:

[0100] ;

[0101] Then calculate the distance of each instance to the category center. The samples are sorted in ascending order, and the most suitable sample is selected based on the ranking.

[0102] Additionally, it should be noted that the number of representative samples in this embodiment of the invention is limited: the number of representative samples for each category = 2000 / the number of current categories.

[0103] Therefore, the dataset structure in the current incremental training phase is as follows: dataset i = data subset i + representative dataset, where data subset i = data subset i + representative dataset ... The subsets are the divisions of the total dataset. The representative dataset is a collection of representative samples selected through the herding strategy.

[0104] S360. Combine the second classification dataset with the representative dataset and input them into the updated class incremental learning tailings ash detection network model for training until all classification datasets are traversed to obtain the final class incremental learning tailings ash detection network model.

[0105] S370. Based on the final incremental learning tailings ash detection network model, tailings ash content is detected in the two-dimensional spectral image data of coal slime mixture, and the tailings ash content detection results of coal slime mixture are obtained.

[0106] Finally, the specific implementation steps of the present invention are summarized. This invention provides a method for dynamically processing and progressively updating coal sample data online. First, data is continuously input into an incremental model in the form of a data stream, causing the model to be constantly updated. Historical data corresponds to historical labels, and new data corresponds to new labels. While learning new data, old data is less easily forgotten, effectively detecting both past and newly added coal sample ash content. The model is continuously updated based on the current coal sample quality to adapt to industrial needs.

[0107] like Figure 5As shown, this embodiment of the invention applies incremental learning technology to the detection of ash content in tailings of coal slime flotation. It can be divided into different task stages according to the ash categories to be detected. For example, if a coal sample originally has ten ash categories, it can be put into task 1 to train a ten-category model, which can detect all current ash categories. When a coal sample with new ash categories is available in a new stage, it can be used as task 2 to continue training the original ten-category model and update the categories at the same time, without having to retrain a new model by combining all the previous coal samples and the new coal samples.

[0108] Finally, it needs to be explained that the number of times the model needs to be trained in the entire incremental learning framework of this invention depends on the amount of task (coal sample changes). Specifically, suppose that this round of coal slime flotation production includes three batches of different coal samples. First, the absorption spectrum data of the first coal sample is collected, and a ResNet18 classification model is trained to identify the 10 ash components of this batch of coal samples. However, when the second batch of coal samples arrives, its configuration and other elemental contents may be different from the first batch, resulting in extremely poor classification performance of the model trained in the first batch of samples. However, starting training from scratch for the second batch of coal samples is time-consuming and labor-intensive, and the trained model can only be applied to a specific batch of coal samples. Therefore, the effectiveness of the incremental learning framework is demonstrated. While preserving the weight data and classification labels of the first coal sample training, incremental training is performed on the dataset of the second coal sample, adding 10 new classification labels suitable for the second batch of coal samples while maintaining the original 10 classifications. When the third batch of coal samples enters the production process, the configuration may change again. Incremental training is performed as described above to avoid starting training from scratch. After several days of production, if a certain type of ash content in a coal sample is found to have an extremely low probability of occurrence during model training or recognition, the model will selectively forget that label to avoid model redundancy.

[0109] Reference Figure 2 A tailings ash content detection system based on online incremental learning includes:

[0110] The first module 201 is used to acquire the coal slime mixture dataset and perform data preprocessing to obtain two-dimensional spectral image data of the coal slime mixture.

[0111] The second module 202 is used to introduce an online incremental learning mechanism and construct a tailings ash content detection network model based on incremental learning, based on the backbone neural network model.

[0112] The third module 203 is used to introduce the herding strategy, which uses a class-incremental learning tailings ash detection network model to detect tailings ash content in two-dimensional spectral image data of coal slime mixture, and obtains the tailings ash content detection results of coal slime mixture.

[0113] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0114] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An online class incremental learning based tail coal ash detection method, characterized in that, The method comprises the following steps: obtaining a coal slime mixed liquid data set and performing data preprocessing to obtain two-dimensional spectral image data of the coal slime mixed liquid; an online class incremental learning mechanism is introduced, and a class incremental learning tail coal ash content detection network model is constructed based on a backbone neural network model; the two-dimensional spectral image data of the coal slime mixed liquid is divided to obtain a plurality of classification data sets; the first classification data set is input into the backbone neural network model of the class incremental learning tail coal ash content detection network model for training to obtain a trained backbone neural network model; weight data and bias data are obtained based on the trained backbone neural network model; the class incremental learning tail coal ash content detection network model is updated based on the weight data and the bias data to obtain an updated class incremental learning tail coal ash content detection network model; representative data sets are obtained by selecting samples from the first classification data set through a herding strategy; the second classification data set and the representative data sets are combined and input into the updated class incremental learning tail coal ash content detection network model for training until all classification data sets are traversed, and a final class incremental learning tail coal ash content detection network model is obtained; the two-dimensional spectral image data of the coal slime mixed liquid is detected based on the final class incremental learning tail coal ash content detection network model to obtain a tail coal ash content detection result of the coal slime mixed liquid.

2. The tail coal ash content detection method based on online class incremental learning according to claim 1, characterized in that, The step of obtaining a coal slime mixed liquid data set and performing data preprocessing to obtain two-dimensional spectral image data of the coal slime mixed liquid specifically comprises: obtaining a plurality of groups of coal slime mixed liquids with different concentrations and different ash contents; a peristaltic pump is used to drive the coal slime mixed liquid to form a liquid flow, and a flow-through cuvette observation window is arranged to collect data, thereby obtaining a coal slime mixed liquid image data set; absorption spectral data of the coal slime mixed liquid image data set are collected by combining a fiber spectrometer and a xenon lamp light source, thereby obtaining spectral time series data of the coal slime mixed liquid image; the spectral time series data of the coal slime mixed liquid image are converted into two-dimensional image data by a Gram angle field method, thereby obtaining two-dimensional spectral image data of the coal slime mixed liquid.

3. The tail coal ash content detection method based on online class incremental learning according to claim 2, characterized in that, The step of inputting the first classification data set into the backbone neural network model of the class incremental learning tail coal ash content detection network model for training to obtain a trained backbone neural network model specifically comprises: the first classification data set is input into the backbone neural network model of the class incremental learning tail coal ash content detection network model, wherein the backbone neural network model comprises an input layer, an initial convolutional layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, and a fully connected layer; the first classification data set is obtained based on the input layer of the backbone neural network model; the first classification data set is subjected to preliminary feature extraction processing based on the initial convolutional layer of the backbone neural network model, thereby obtaining low-level features of the first classification data set; the low-level features of the first classification data set are subjected to deep feature extraction processing based on the first residual layer, the second residual layer, the third residual layer, and the fourth residual layer of the backbone neural network model, thereby obtaining deep-level features of the first classification data set; The deep features of the first classification data set are linearly transformed and mapped based on the full connection layer of the backbone neural network model to obtain a detection result of the first classification data set; The detection accuracy of the detection result of the first classification data set is judged; If the detection accuracy of the detection result of the first classification data set does not meet the preset accuracy threshold, the first classification data set is re-input into the backbone neural network model for training until the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold; If the detection accuracy of the detection result of the first classification data set meets the preset accuracy threshold, the trained backbone neural network model is output.

4. The tail coal ash content detection method based on online class incremental learning according to claim 3, characterized in that, The step of updating the class-incremental learning tail coal ash content detection network model according to the weight data and the bias data to obtain an updated class-incremental learning tail coal ash content detection network model specifically includes: A new full connection layer is generated for the backbone neural network model; The output class number of the original full connection layer of the backbone neural network model is copied to the new full connection layer; The new full connection layer is valued according to the weight data and the bias data to obtain a valued full connection layer; The valued full connection layer is replaced with the original full connection layer to obtain the updated class-incremental learning tail coal ash content detection network model.

5. The tail coal ash content detection method based on online class incremental learning according to claim 4, characterized in that, The expression of updating the class-incremental learning tail coal ash content detection network model according to the weight data and the bias data is specifically as follows: ； In the above formula, denotes a new fully connected layer, denotes the feature dimension.

6. The tail coal ash content detection method based on online class incremental learning according to claim 5, characterized in that, The step of selecting samples from the first classification data set to obtain a representative data set by using the herding strategy specifically includes: A current set embedding class center of the first classification data set is obtained; The distance of each instance in the first classification data set to the current set embedding class center is calculated and sorted in ascending order to obtain a sorted distance result; The representative data set is obtained by selecting the sorted distance result according to a preset number of representative samples.

7. The tail coal ash content detection method based on online class incremental learning according to claim 6, characterized in that, The expression of obtaining the current set embedding class center of the first classification data set is specifically as follows: ； In the above formulae, denotes a class of the first classification dataset, denotes a total number of instances of the first classification dataset, denotes the first classification dataset, denotes the first classification dataset, denotes the first classification dataset, denotes a current set embedding class center.

8. An online class-incremental learning based tail coal ash detection system, characterized in that, The following modules are included: A first module is configured to obtain a coal slime mixed liquid data set and perform data preprocessing to obtain a two-dimensional spectrum image data of the coal slime mixed liquid; A second module is configured to introduce an online class-incremental learning mechanism, construct a class-incremental learning tail coal ash content detection network model based on a backbone neural network model; A third module is configured to divide the two-dimensional spectrum image data of the coal slime mixed liquid to obtain a plurality of classification data sets; The first classification data set is input into the backbone neural network model of the class-incremental learning tail coal ash content detection network model for training to obtain a trained backbone neural network model; Weight data and bias data are obtained based on the trained backbone neural network model; The class-incremental learning tail coal ash content detection network model is updated according to the weight data and the bias data to obtain an updated class-incremental learning tail coal ash content detection network model; Samples are selected from the first classification data set by using the herding strategy to obtain a representative data set; The second classification data set is combined with the representative data set and input into the updated class-incremental learning tail coal ash content detection network model for training until all classification data sets are traversed, and a final class-incremental learning tail coal ash content detection network model is obtained; Based on the final class-incremental learning tail coal ash content detection network model, the two-dimensional spectral image data of the slime mixed liquid are detected to obtain the tail coal ash content detection result of the slime mixed liquid.

Images