A semi-automatic method and system for monitoring the plastic layer index of coal.

Abstract

This invention relates to the field of coal caking index monitoring, and more particularly to a semi-automatic coal caking index monitoring method and system. The method includes the following steps: acquiring a dataset; acquiring a pre-constructed index monitoring network; inputting samples from the dataset into the index monitoring network in batches to complete iterative training of the index monitoring network, thereby achieving semi-automatic coal caking index monitoring; wherein any iterative training process includes: constructing a small-sample feature extraction loss function, and using the small-sample feature extraction loss function to supervise the training of the index monitoring network. This effectively improves the accuracy of caking index monitoring for small-sample coal types.

Description

Technical Field

[0001] This invention relates to the field of coal caking index monitoring, and more particularly to a semi-automatic coal caking index monitoring method and system. Background Technology

[0002] The plastic layer index is a key indicator for evaluating the quality of coking coal. It directly reflects the formation and solidification characteristics of plastic bodies during coal pyrolysis and has significant guiding value in production practices in industries such as coal and metallurgy. Currently, the determination of the coal plastic layer index mainly relies on a semi-automatic plastic layer index measuring instrument. By recording the temperature-displacement curve of the coal sample during heating, core parameters such as the maximum thickness and final shrinkage of the plastic layer are obtained through manual interpretation or algorithm analysis.

[0003] With the development of intelligent detection technology, neural network-based data analysis methods have been gradually applied to the automatic analysis of the plastic layer index, significantly improving detection efficiency and standardization. However, existing neural networks exhibit significant limitations when processing small sample coal data. Due to the limited number of samples for small coal types (such as scarce coal types or coal samples from special geological areas), traditional neural network architectures struggle to effectively capture the feature information contained in the data, easily leading to underfitting or overfitting of the model. Especially in the plastic layer index detection scenario, the temperature-displacement curves of small sample coal types are complex and variable, and subtle differences in curve morphology may correspond to different plastic body formation patterns. Traditional networks cannot accurately extract this unique information, ultimately resulting in significant errors in the plastic layer index detection results for small sample coal types, severely restricting the refined evaluation and rational utilization of coal resources. Therefore, it is urgent to develop a precise detection method for the plastic layer index of small sample coal types to overcome the technical bottleneck of traditional neural networks in processing small sample data. Summary of the Invention

[0004] To address the issue of improving the accuracy of coal mortar index monitoring for small sample coal types, this invention provides a semi-automatic coal mortar index monitoring method and system.

[0005] In a first aspect, the present invention provides a semi-automatic method for monitoring the coal caking layer index, employing the following technical solution:

[0006] A semi-automatic method for monitoring the plastic layer index of coal includes the following steps:

[0007] Obtain a dataset containing several samples, which can be divided into large samples and small samples. Each sample is a labeled index monitoring curve of the semi-automatic coal colloid layer.

[0008] Obtain a pre-constructed index monitoring network, which is a network with a small sample feature extraction structure added between the shared layer and the fully connected layer of the Conv1D network;

[0009] The samples in the dataset are input into the index monitoring network in batches to complete the iterative training of the index monitoring network, so as to realize the semi-automatic coal mortar index monitoring.

[0010] Each iteration of the training process includes: constructing a few-sample feature extraction loss function. The small sample input in this batch is designated as the target small sample. Samples belonging to the same type of coal as the target small sample are designated as reference samples. The common features of the output data of all reference samples in the shared layer are extracted and designated as common features. The distinguishing features between the target small sample and the large sample input in this batch in the output data of the shared layer are extracted and designated as unique features. Indicates common features, This represents the output data of the target small sample in the small sample extraction structure. Indicates unique characteristics, The correlation of the data is represented; the training of the index monitoring network is jointly supervised by the loss function extracted from small sample features and the loss function built into the Conv1D network.

[0011] This invention addresses the issue that some coal types have limited sample sizes, hindering traditional networks from effectively extracting features from small sample coal types. Therefore, it adds a small-sample feature extraction structure to the traditional network specifically for extracting unique information from small sample coal types, thereby improving the accuracy of monitoring the plastic layer index for small sample coal types. Furthermore, by setting a small-sample feature extraction loss function for the small-sample structure, it enables the structure to extract small-sample features more quickly and effectively. Moreover, by introducing unique features into the small-sample feature extraction loss function, the structure can extract more unique information from small samples that distinguishes them from larger samples, thus compensating for the inability of traditional networks to extract unique information from small samples. Finally, by introducing shared features into the small-sample feature extraction loss function, the structure can extract shared information among small samples, preventing the extraction of discrepancies and noise, thereby improving the accuracy of feature extraction.

[0012] Preferably, acquiring the pre-built index monitoring network includes:

[0013] The first predetermined number of network layers in the Conv1D network are designated as shared layers. The network layers between the shared layers and the fully connected layers in the Conv1D network are denoted as large sample feature extraction structures. A threshold structure and a small sample feature extraction structure are added after the shared layers. The small sample feature extraction structure is located after the threshold structure and parallel to the large sample feature extraction structure. The threshold structure is used to divert the data stream of small sample information to the small sample extraction structure and the data stream of large sample information to the large sample feature extraction structure.

[0014] This invention takes into account that the front layers of the network only have the ability to extract common shallow features. Therefore, a shared layer is reserved in the network to extract features from both large and small samples. This allows large samples to participate in the training at the shared layer, effectively reducing the training scale of network parameters unique to small samples and preventing the problem that the number of small samples is insufficient to complete the training of large-scale unique network parameters.

[0015] Preferably, the network layers in the small sample feature extraction structure are the same as those in the large sample feature extraction structure.

[0016] Preferably, the method for achieving semi-automatic coal mortar index monitoring includes:

[0017] The index monitoring curve of the newly collected coal semi-automatic plastic layer is input into the trained index monitoring network to obtain the semi-automatic plastic layer index of the newly collected coal.

[0018] Preferably, the common features extracted from the output data of all reference samples in the shared layer are denoted as common features, including:

[0019] The output data of each reference sample in the shared layer is obtained. The mean of all output data of each reference sample in the shared layer is used as the comprehensive output data of each reference sample. Principal component analysis algorithm is used to reduce the dimensionality of the comprehensive output data of all reference samples, and the resulting dimensionality-reduced data is used as the common feature.

[0020] This invention uses principal component analysis (PCA) to reduce the dimensionality of reference samples and remove the differences between them, thus accurately extracting the common information between the reference samples.

[0021] Preferably, the method for obtaining the unique feature includes:

[0022] The process involves: acquiring the output data of the large samples in the shared layer and using the average of the output data of the large samples in the shared layer as the comprehensive output data of the large samples; acquiring the output data of the target small samples in the shared layer and using the average of the output data of the target small samples in the shared layer as the comprehensive output data of the target small samples; calculating the absolute value of the difference between the corresponding positions of the comprehensive output data of the large samples and the comprehensive output data of the target small samples to obtain the difference data; setting the positions in the difference data with values ​​greater than a preset difference threshold to 1 and the positions with values ​​not greater than the preset difference threshold to 0 to obtain sparse data; performing an OR operation on the corresponding positions of the sparse data of all large samples to obtain comprehensive sparse data; and performing a multiplication operation on the comprehensive output data of the target small samples and the comprehensive sparse data to obtain the unique features.

[0023] This invention excludes information that is only slightly different from the large sample, and extracts the unique information of the target small sample more accurately.

[0024] Preferably, the step of using a small-sample feature extraction loss function and the loss function inherent in the Conv1D network to jointly supervise the training of the exponential monitoring network includes:

[0025] The first loss value is calculated using the small sample feature extraction loss function, and the second loss value is calculated using the loss function built into the Conv1D network. The parameters in the exponential monitoring network are then updated using the first and second loss values ​​via back gradient.

[0026] Preferably, the index monitoring curve is a temperature-displacement curve measured using a semi-automatic gel layer index measuring instrument.

[0027] Preferably, the samples are divided into large and small samples by comparing the sample size of each type of coal with a preset value.

[0028] This invention distinguishes between large and small samples by comparing preset values ​​of sample size. This method is relatively simple and has high efficiency.

[0029] Secondly, this invention provides a semi-automatic coal mortar index monitoring system, which adopts the following technical solution:

[0030] A semi-automatic coal caking layer index monitoring system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned semi-automatic coal caking layer index monitoring method is implemented.

[0031] By adopting the above technical solution, a computer program for the semi-automatic coal colloid index monitoring method is generated and stored in a memory for loading and execution by a processor. Terminal equipment is then manufactured based on the memory and processor for convenient use.

[0032] The present invention has the following technical effects:

[0033] This invention takes into account that the number of samples extracted for some coal types is small, which makes it impossible for traditional networks to obtain feature extraction capabilities for small sample coal types. Therefore, by adding a small sample feature extraction structure to the traditional network, it is specifically used to extract the unique information of small sample coal types, thereby improving the accuracy of monitoring the plastic layer index of small sample coal types.

[0034] Furthermore, by setting a small sample feature extraction loss function for the small sample structure, the small sample structure can acquire the ability to extract small sample features faster and better.

[0035] Furthermore, specific features are introduced into the small sample feature extraction loss function to enable the small sample feature extraction structure to extract more specific information of small samples that are different from large samples, thereby making up for the inability of traditional networks to extract specific information of small samples.

[0036] Furthermore, common features are introduced into the few-sample feature extraction loss function to enable the few-sample feature extraction structure to extract common information between few samples, thereby preventing the few-sample feature extraction structure from extracting some differentiated noise information, thus improving the accuracy of feature extraction by the few-sample feature extraction structure. Attached Figure Description

[0037] Figure 1 This is a flowchart of a semi-automatic coal caking layer index monitoring method according to an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the network structure of the index monitoring network provided in an embodiment of the present invention. Detailed Implementation

[0039] This invention discloses a semi-automatic method for monitoring the plastic layer index of coal, referring to... Figure 1 This includes steps S1-S3:

[0040] S1: Obtain the dataset, which contains several samples. The samples can be divided into large samples and small samples. Each sample is an index monitoring curve of the semi-automatic coal colloid layer with labels.

[0041] Specifically, a semi-automatic plastic layer index meter is used to measure the temperature-displacement curves of various types of coal, which are recorded as index monitoring curves. Based on experience, the index monitoring curves are analyzed to obtain the plastic layer index, which is then used as a label for the corresponding index monitoring curve. The plastic layer index can be the maximum thickness and final shrinkage of the plastic layer, or other data; this embodiment does not limit the specific data.

[0042] The labeled index monitoring curve is used as a sample, and the dataset consists of all samples.

[0043] Sample data for each type of coal is obtained from the dataset. Samples of coal types with a sample size smaller than a preset value are denoted as small samples; samples of coal types with a sample size not less than a preset value are denoted as large samples.

[0044] S2: Obtain a pre-constructed index monitoring network, which is a network with a small sample feature extraction structure added between the shared layer and the fully connected layer of the Conv1D network.

[0045] It should be noted that in order to perform gelatin index detection, a network for gelatin index detection needs to be constructed first.

[0046] It should be further explained that traditional neural networks use a single network structure to extract features from all samples. However, some coal types have a smaller sample size, resulting in less data being used in network training. This makes the network more suitable for feature extraction from large samples, leading to poor performance on small samples. To improve the network's ability to extract features from small samples, we consider adding a separate network structure to the traditional network specifically for extracting feature information from small samples.

[0047] Preferably, as an example, a pre-built index monitoring network is obtained, including:

[0048] The first predetermined number of network layers in the Conv1D network are designated as shared layers. The network layers between the shared layers and the fully connected layers in the Conv1D network are denoted as large sample feature extraction structures. A threshold structure and a small sample feature extraction structure are added after the shared layers. The small sample feature extraction structure is located after the threshold structure and parallel to the large sample feature extraction structure. The threshold structure is used to divert the data stream of small sample information to the small sample extraction structure and the data stream of large sample information to the large sample feature extraction structure.

[0049] The network layers in the small sample feature extraction structure are the same as those in the large sample feature extraction structure, except that the number of network layers in the small sample feature extraction structure can be less than the number of network layers in the large sample feature extraction structure. Figure 2 This is a schematic diagram of the network structure of the index monitoring network. Figure 2 The dashed rectangle represents a shared layer, the diamond represents a threshold structure, the upper solid rectangle represents a large-sample feature extraction structure, the lower solid rectangle represents a small-sample feature extraction structure, and the cylinder represents a fully connected layer.

[0050] Understandably, due to the limited amount of data in small sample datasets, large-scale feature extraction structures trained with a small number of samples are not accurate enough. Normally, the features extracted by the earlier layers of the network are mostly shared shallow information. Therefore, the earlier feature extraction structures can be trained using all samples. This reduces the size of the feature extraction structures trained on small samples alone, thereby improving the feature extraction capability.

[0051] To facilitate understanding, the data flow of the index monitoring network is explained below: Samples are input into the index monitoring network and first processed by the shared layer. After processing by the shared layer, the data flows into the threshold structure. The threshold structure determines whether a sample is large or small by identifying the number of coal samples belonging to that category. If the sample is determined to be large, the data flows into the large sample feature extraction structure; if it is determined to be small, the data flows into the small sample extraction structure. After processing by the corresponding large or small sample feature extraction structure, the data flows into the fully connected layer, and after further processing by the fully connected layer, the output result is obtained.

[0052] S3: Input the samples in the dataset into the index monitoring network in batches to complete the iterative training of the index monitoring network, so as to realize the semi-automatic coal mortar index monitoring.

[0053] S30: Input the samples in the dataset into the index monitoring network in batches to complete the iterative training of the index monitoring network.

[0054] It should be noted that the limited number of small samples means that traditional neural networks lack the ability to extract features specific to these small samples, resulting in poor monitoring of the plastic layer index for small coal types. To enable the index monitoring network to learn the unique features of small samples, a loss function needs to be set for the small sample feature extraction structure. This loss function will then supervise the extraction structure to ensure it can extract the unique information of the small samples.

[0055] Each iteration of the training process includes: constructing a few-shot feature extraction loss function; and using the few-shot feature extraction loss function and the loss function built into the Conv1D network to jointly supervise the training of the exponential monitoring network.

[0056] S300: Construct a loss function for small sample feature extraction.

[0057] Preferably, as an example, a few-sample feature extraction loss function is constructed, including:

[0058]

[0059] Specifically, the small sample input in this batch is denoted as the target small sample; samples belonging to the same type of coal as the target small sample are denoted as reference samples; the common features of the output data of all reference samples in the shared layer are extracted and denoted as common features; and the distinguishing features between the target small sample and the large sample input in this batch in the output data of the shared layer are extracted and denoted as unique features. Indicates common features, This represents the output data of the target small sample in the small sample extraction structure. Indicates unique characteristics, S represents the correlation of the data; S represents the loss function for small sample feature extraction.

[0060] It is understandable that the target small sample is the output data in the small sample extraction structure. This reflects the target small sample feature information extracted by the small sample extraction structure, including unique features. It reflects the unique characteristics that distinguish small samples from large samples. This value reflects the correlation between the feature information and specific information of the target small sample extracted by the small sample extraction structure. A larger value indicates that the small sample extraction structure can extract the specific information of the target small sample better. Common Features It reflects the common characteristic information of samples of the same type as the target small sample. The correlation between the feature information and common information of the target small sample extracted by the small sample extraction structure is important. Since the information that distinguishes the target small sample from the large sample may also contain noise information, the extraction of such information may interfere with the detection results. Therefore, by analyzing the correlation between the feature information and common information of the target small sample extracted by the small sample extraction structure, the small sample extraction structure can extract more of the unique common information of the small sample rather than some noise information.

[0061] The above embodiments involve common features, unique features, the correlation between common features and the output data of the target small sample in the small sample extraction structure, and the correlation between unique features and the output data of the target small sample in the small sample extraction structure. The following will explain the methods for determining the correlation between common features, unique features, the correlation between common features and the output data of the target small sample in the small sample extraction structure, and the correlation between unique features and the output data of the target small sample in the small sample extraction structure.

[0062] First, we will introduce the method for obtaining common features.

[0063] Preferably, as an example, common feature acquisition methods include:

[0064] The output data of each reference sample in the shared layer is obtained. The mean of all output data of each reference sample in the shared layer is used as the comprehensive output data of each reference sample. Principal component analysis algorithm is used to reduce the dimensionality of the comprehensive output data of all reference samples, and the resulting dimensionality-reduced data is used as the common feature.

[0065] It is understandable that after the principal component analysis algorithm is used to reduce the dimensionality of the comprehensive output data, the common correlation information in the comprehensive output data will be retained and some special information will be removed, thereby extracting the common feature information.

[0066] Then, the methods for obtaining unique features will be introduced.

[0067] Preferably, as an example, methods for obtaining unique features include:

[0068] The process involves: acquiring the output data of the large samples in the shared layer and using the average of the output data of the large samples in the shared layer as the comprehensive output data of the large samples; acquiring the output data of the target small samples in the shared layer and using the average of the output data of the target small samples in the shared layer as the comprehensive output data of the target small samples; calculating the absolute value of the difference between the corresponding positions of the comprehensive output data of the large samples and the comprehensive output data of the target small samples to obtain the difference data; setting the positions in the difference data with values ​​greater than a preset difference threshold to 1 and the positions with values ​​not greater than the preset difference threshold to 0 to obtain sparse data; performing an OR operation on the corresponding positions of the sparse data of all large samples to obtain comprehensive sparse data; and performing a multiplication operation on the comprehensive output data of the target small samples and the comprehensive sparse data to obtain the unique features.

[0069] It is understandable that the data at different positions in the comprehensive output data reflect different data characteristics. By filtering out the data with the largest differences between the comprehensive output data of the target small sample and the comprehensive output data of the large sample, the unique features of the target small sample compared with the large sample are identified.

[0070] The following section introduces a method for calculating the correlation between common features and the output data of the target small sample in the small sample extraction structure.

[0071] Preferably, as an example, a method for calculating the correlation between shared features and the output data of the target small sample in the small sample extraction structure includes:

[0072] The mutual information between the common features and the output data of the target small sample in the small sample extraction structure is calculated as the correlation between the common features and the output data of the target small sample in the small sample extraction structure.

[0073] It is understandable that mutual information can reflect the correlation between different data sequences. It can reflect both linear and nonlinear correlations, and therefore it can better reflect the correlation between common features and the output data of the target small sample in the small sample extraction structure.

[0074] Finally, we introduce a method for understanding the correlation between unique features and the output data of target small samples in the small sample extraction structure.

[0075] Preferably, as an example, a method for calculating the correlation between specific features and the output data of the target small sample in the small sample extraction structure includes:

[0076] The mutual information between the specific features and the output data of the target small sample in the small sample extraction structure is calculated as the correlation between the specific features and the output data of the target small sample in the small sample extraction structure.

[0077] S301: The training of the exponential monitoring network is jointly supervised by the loss function extracted from small sample features and the loss function built into the Conv1D network.

[0078] Preferably, as an example, the training of the exponential monitoring network is jointly supervised using both the few-sample feature extraction loss function and the loss function built into the Conv1D network, including:

[0079] After the small samples in this batch are input into the index monitoring network, the loss value is calculated using the small sample feature extraction loss function and recorded as the first loss value;

[0080] Finally, the output data is obtained in the exponential monitoring network. Based on the output data and the label, the loss value is calculated using the loss function built into the Conv1D network in the exponential monitoring network and recorded as the second loss value.

[0081] Using the first and second loss values, the network parameters in the exponential monitoring network are updated in reverse using the gradient descent method.

[0082] It is understandable that, since only small samples extract structures, only small sample information is used to update the information of small sample extracted structures, thereby preventing large sample information from overwhelming small sample information and causing the network to lose the ability to extract small sample features.

[0083] It should be noted that updating network parameters in reverse using gradient descent based on the loss value is an existing technique and will not be elaborated upon here.

[0084] S31: To achieve semi-automatic monitoring of coal mortar index.

[0085] Preferably, as an example, to achieve semi-automatic monitoring of the coal seam index, the following is included:

[0086] The index monitoring curve of the newly collected coal semi-automatic plastic layer is input into the trained index monitoring network to obtain the semi-automatic plastic layer index of the newly collected coal.

[0087] This invention also discloses a semi-automatic coal caking layer index monitoring system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a semi-automatic coal caking layer index monitoring method according to the present invention.

[0088] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0089] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (DRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (DRAM), high-bandwidth memory, hybrid memory cube, etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

Claims

1. A semi-automatic method for monitoring the coal mortar index, characterized in that, Including the following steps: Obtain a dataset containing several samples, which can be divided into large samples and small samples. Each sample is a labeled index monitoring curve of the semi-automatic coal colloid layer. Obtain a pre-constructed index monitoring network, which is a network with a small sample feature extraction structure added between the shared layer and the fully connected layer of the Conv1D network; The samples in the dataset are input into the index monitoring network in batches to complete the iterative training of the index monitoring network, so as to realize the semi-automatic coal mortar index monitoring. Each iteration of the training process includes: constructing a few-sample feature extraction loss function. The small sample input in this batch is designated as the target small sample. Samples belonging to the same type of coal as the target small sample are designated as reference samples. The common features of the output data of all reference samples in the shared layer are extracted and designated as common features. The distinguishing features between the target small sample and the large sample input in this batch in the output data of the shared layer are extracted and designated as unique features. Indicates common features, This represents the output data of the target small sample in the small sample extraction structure. Indicates unique characteristics, The correlation of the data is represented; the training of the index monitoring network is jointly supervised by the loss function extracted from small sample features and the loss function built into the Conv1D network.

2. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The acquisition of the pre-built index monitoring network includes: The first predetermined number of network layers in the Conv1D network are designated as shared layers. The network layers between the shared layers and the fully connected layers in the Conv1D network are denoted as large sample feature extraction structures. A threshold structure and a small sample feature extraction structure are added after the shared layers. The small sample feature extraction structure is located after the threshold structure and parallel to the large sample feature extraction structure. The threshold structure is used to divert the data stream of small sample information to the small sample extraction structure and the data stream of large sample information to the large sample feature extraction structure.

3. The semi-automatic coal mortar index monitoring method according to claim 2, characterized in that, The network layers in the small sample feature extraction structure are the same as those in the large sample feature extraction structure.

4. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The method for achieving semi-automatic coal mortar index monitoring includes: The index monitoring curve of the newly collected coal semi-automatic plastic layer is input into the trained index monitoring network to obtain the semi-automatic plastic layer index of the newly collected coal.

5. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The common features extracted from the output data of all reference samples in the shared layer are denoted as common features, including: The output data of each reference sample in the shared layer is obtained. The mean of all output data of each reference sample in the shared layer is used as the comprehensive output data of each reference sample. Principal component analysis algorithm is used to reduce the dimensionality of the comprehensive output data of all reference samples, and the resulting dimensionality-reduced data is used as the common feature.

6. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The method for obtaining the unique features includes: The process involves: acquiring the output data of the large samples in the shared layer and using the average of the output data of the large samples in the shared layer as the comprehensive output data of the large samples; acquiring the output data of the target small samples in the shared layer and using the average of the output data of the target small samples in the shared layer as the comprehensive output data of the target small samples; calculating the absolute value of the difference between the corresponding positions of the comprehensive output data of the large samples and the comprehensive output data of the target small samples to obtain the difference data; setting the positions in the difference data with values ​​greater than a preset difference threshold to 1 and the positions with values ​​not greater than the preset difference threshold to 0 to obtain sparse data; performing an OR operation on the corresponding positions of the sparse data of all large samples to obtain comprehensive sparse data; and performing a multiplication operation on the comprehensive output data of the target small samples and the comprehensive sparse data to obtain the unique features.

7. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The method of using a small-sample feature extraction loss function and the loss function built into the Conv1D network to jointly supervise the training of the exponential monitoring network includes: The first loss value is calculated using the small sample feature extraction loss function, and the second loss value is calculated using the loss function built into the Conv1D network. The parameters in the exponential monitoring network are then updated using the first and second loss values ​​via back gradient.

8. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The index monitoring curve is a temperature-displacement curve measured using a semi-automatic gel layer index meter.

9. The semi-automatic coal mortar index monitoring method according to claim 1, characterized in that, The samples are divided into large and small samples by comparing the sample size of each type of coal with preset values.

10. A semi-automatic coal mortar index monitoring system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a semi-automatic coal caking layer index monitoring method according to any one of claims 1-9.

Images