A method and device for classifying and identifying a CO concentration source, an electronic device, and a medium

Abstract

The application provides a CO concentration source classification and identification method and device, electronic equipment and medium, relates to the coal mine safety technical field, and includes: collecting original CO concentration data from different generation sources, obtaining target CO concentration data by cropping, and constructing a labeled CO concentration sample database; preprocessing the target CO concentration data to obtain standardized CO concentration sequences; extracting features from the standardized CO concentration sequences to obtain curve features representing the concentration change rules of different sources, and assembling CO concentration sequence feature vectors; performing feature selection and dimensionality reduction processing on the CO concentration sequence feature vectors to obtain optimized low-dimensional feature vectors; inputting the optimized low-dimensional feature vectors into a classification and identification model to determine the classification and identification result of the CO concentration source. Through the application, the accuracy of distinguishing and classifying the CO concentration changes of different sources is improved.

Description

Technical Field

[0001] This application relates to the field of coal mine safety technology, and more specifically, to a method, apparatus, electronic device, and medium for classifying and identifying CO concentration sources. Background Technology

[0002] Currently, real-time and accurate monitoring of CO gas in underground coal mines is a key technological means to prevent major disasters such as spontaneous combustion of coal. Traditional monitoring methods mainly rely on single-point detection and alarm mechanisms based on CO concentration thresholds. However, underground CO sources are complex and diverse, including not only spontaneous combustion of coal but also various interference sources such as mining operations, blasting construction, and exhaust gases from transportation equipment. The CO concentration change trends from these different sources often exhibit similar patterns, making monitoring methods that rely solely on concentration values ​​inherently limited.

[0003] Existing technologies face two prominent bottlenecks in practical applications. First, at the data analysis level, traditional methods mainly rely on statistical characteristics such as the mean and peak values ​​of CO concentration. Second, at the identification model level, most methods employ single, shallow classification algorithms, which lack the ability to classify nonlinear, high-dimensional feature data, leading to a high misclassification rate for CO sources related to spontaneous combustion of coal.

[0004] Therefore, the current technology system cannot accurately identify CO concentration, making it difficult to meet the accuracy and reliability requirements of coal mine safety monitoring. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, apparatus, electronic device and medium for classifying and identifying CO concentration sources, so as to overcome at least one of the above-mentioned defects.

[0006] In a first aspect, embodiments of this application provide a method for classifying and identifying CO concentration sources. The method includes: collecting raw CO concentration data from different sources, cropping target CO concentration data, and constructing a labeled CO concentration sample database; preprocessing the target CO concentration data to obtain a standardized CO concentration sequence; extracting features from the standardized CO concentration sequence to obtain curve features characterizing the concentration variation patterns of different sources, and constructing a CO concentration sequence feature vector; performing feature selection and dimensionality reduction processing on the CO concentration sequence feature vector to obtain an optimized low-dimensional feature vector; and inputting the optimized low-dimensional feature vector into a classification and identification model to determine the classification and identification result of the CO concentration source.

[0007] In one optional embodiment of this application, a labeled CO concentration sample database is constructed in the following manner: the original CO concentration data is cropped to retain stable bands that contain the complete process of concentration change. The stable band refers to the data segment from the start of concentration change to the point of stabilization, and the stable band includes the target CO concentration data. A CO source label is set for each stable band to construct a CO concentration sample database caused by different reasons. The CO source label is used to identify the specific source of the target CO concentration data in the stable band.

[0008] In one optional embodiment of this application, a standardized CO concentration sequence is obtained by: processing outliers in the target CO concentration data of the stable band using a preset criterion, the preset criterion being used to identify and remove data points that deviate from the average value; filling in missing values ​​in the target CO concentration data of the stable band using an interpolation method, the interpolation method being used to estimate missing values ​​based on time series trends; removing periodic interference components in the target CO concentration data of the stable band using an empirical mode decomposition reconstruction method to obtain deperiodic CO concentration data for periodic fluctuations; and performing normalization processing on the processed deperiodic CO concentration data to convert it to a uniform numerical scale to obtain a standardized CO concentration sequence.

[0009] In one optional embodiment of this application, the CO concentration sequence feature vector is constructed in the following manner: the standardized CO concentration sequence is input into a deep neural network module, and a curve feature representing the standardized CO concentration sequence is output, wherein the deep neural network module is used to represent the relationship between the standardized CO concentration sequence and the curve feature, and the curve feature is a feature extracted from the standardized CO concentration sequence that can distinguish the morphology and pattern of different generating sources; the curve features are combined to construct the CO concentration sequence feature vector.

[0010] In one optional embodiment of this application, the optimal low-dimensional feature vector is obtained by: performing feature selection using a recursive feature elimination method on the CO concentration sequence feature vector to remove unimportant features; and performing dimensionality reduction using principal component analysis on the features after feature selection to obtain the optimal low-dimensional feature vector.

[0011] In one optional embodiment of this application, the classification and identification result of CO concentration source is determined by the following method: the optimized low-dimensional feature vector is input into the classification and identification model to obtain the corresponding source classification probability distribution, wherein the classification and identification model is used to characterize the mapping relationship between the optimized low-dimensional feature vector and the CO source category; the category is determined based on the source classification probability distribution to obtain the classification and identification result of CO concentration source.

[0012] In one optional embodiment of this application, the classification and recognition model is trained in the following manner: a training sample set is obtained, the training sample set including multiple training samples, each training sample including a sample optimized low-dimensional feature vector and a corresponding sample CO source label; the sample optimized low-dimensional feature vector is used as the input of the classification and recognition model to be trained, and the corresponding sample CO source label is used as the output of the classification and recognition model to be trained, so as to train the classification and recognition model to be trained.

[0013] Secondly, embodiments of this application also provide a CO concentration source classification and identification device, the device comprising: a CO concentration sample database construction module, used to collect raw CO concentration data from different sources, crop target CO concentration data, and construct a labeled CO concentration sample database; a CO concentration sequence standardization module, used to preprocess the target CO concentration data to obtain a standardized CO concentration sequence; a CO concentration sequence feature vector component module, used to extract features from the standardized CO concentration sequence, obtain curve features characterizing the concentration change patterns of different sources, and construct a CO concentration sequence feature vector; an optimized low-dimensional feature vector construction module, used to perform feature selection and dimensionality reduction processing on the CO concentration sequence feature vector to obtain an optimized low-dimensional feature vector; and a classification and identification result determination module, used to input the optimized low-dimensional feature vector into a classification and identification model to determine the classification and identification result of the CO concentration source.

[0014] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of the method described above are performed.

[0015] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the method described above.

[0016] The CO concentration source classification and identification method, apparatus, electronic device, and medium provided in this application establish a data foundation by constructing a labeled sample database, performing standardized preprocessing on the target CO concentration data, extracting curve features representing source characteristics from the concentration sequence and constructing a feature vector, obtaining optimized feature representations through feature selection and dimensionality reduction, and finally using a classification and identification model to accurately classify CO concentration sources. This achieves deep feature mining and optimized representation of CO concentration change curves, thereby enabling more accurate differentiation of CO concentration change patterns from different sources. It solves the problems of misjudgment and missed judgment of CO sources caused by insufficient feature extraction and limited model recognition capabilities in traditional methods, improving the accuracy and reliability of CO concentration source identification in underground coal mines.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the CO concentration source classification and identification method provided in the embodiments of this application; Figure 2 A flowchart illustrating the construction of a labeled CO concentration sample database provided in this application embodiment; Figure 3 A flowchart illustrating the standardized CO concentration sequence provided in this application embodiment; Figure 4 This is a flowchart illustrating the construction of a CO concentration sequence feature vector provided in an embodiment of this application; Figure 5 A schematic diagram illustrating feature importance analysis provided in an embodiment of this application; Figure 6 This is a flowchart illustrating the process of obtaining the optimized low-dimensional feature vector provided in an embodiment of this application; Figure 7 A flowchart illustrating the classification and identification results for determining the source of CO concentration provided in an embodiment of this application; Figure 8 A schematic diagram illustrating the combination of PCA and SVM for CO source classification, provided for embodiments of this application; Figure 9 This is a schematic diagram illustrating the CO concentration source classification results provided in the embodiments of this application; Figure 10 A schematic diagram illustrating the SHAP feature contribution summary provided in the embodiments of this application; Figure 11 This is a schematic diagram of the CO concentration source classification and identification device provided in the embodiments of this application; Figure 12 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0021] First, the applicable scenarios for this application will be introduced. This application can be applied to the field of coal mine safety technology.

[0022] Research has found that spontaneous combustion of coal seams is one of the most common and dangerous hazards during coal mining and production, with its early characteristics often manifested as the abnormal generation and accumulation of CO. As the main indicator gas in the spontaneous combustion process, CO concentration changes directly affect the safety of underground workers and are also a crucial basis for identifying and controlling coal mine fire hazards. However, the sources of CO underground are highly complex and diverse. Besides the oxidation reaction during spontaneous combustion, they can also originate from raw coal escape, disturbances during mining operations, blasting operations, roof collapses in goaf areas, and exhaust gases from transportation equipment, among other factors. These different sources often intertwine and overlap in time and space, making it difficult to accurately reflect the true development state of spontaneous combustion by simply relying on gas concentration monitoring.

[0023] Current traditional monitoring methods mostly focus on single-point detection and threshold-based early warning of CO concentration, lacking in-depth analysis of its formation mechanism and source differences. When an abnormal increase in CO is observed underground, it is impossible to quickly determine whether it originates from spontaneous combustion of coal or from other non-combustion conditions, thus limiting the timeliness and specificity of prevention and control measures. In this situation, early warnings are highly likely to fail or be misjudged, thereby increasing the risk of spontaneous combustion of coal evolving into major disasters such as fires or even gas explosions.

[0024] Currently, the field of coal spontaneous combustion monitoring and early warning faces four major core technical challenges in identifying CO concentration sources: First, there are limitations in data acquisition and preprocessing. The complex underground environment, such as high humidity, dust, and electromagnetic interference, can easily lead to large fluctuations and noise in CO sensor data. Furthermore, the sampling timing and parameter dimensions are not uniform in different scenarios, such as goaf and blasting areas. There is a lack of standardized preprocessing procedures, making it difficult to form high-quality datasets to support subsequent analysis. Second, sample labeling is difficult. The sources of CO in the mine are hidden, such as the oxidation of deep residual coal. When manually confirming the source, labeling errors are easy to occur. In extreme working conditions, such as the CO pulse during roof collapse, samples are scarce, which leads to class imbalance in the dataset and affects the model training effect. Third, there are shortcomings in the feature engineering process. Existing studies mostly use only the single feature of CO concentration, without fully integrating multi-dimensional features that are strongly correlated with CO generation, such as temperature, oxygen concentration, and equipment operating conditions. This makes it difficult to distinguish the differentiated features of CO from different sources and reduces the classification and identification accuracy. Fourth, research on the identification and classification of CO concentration sources from spontaneous coal combustion is still in its developmental stage. Classification and identification methods are rarely used and lack accuracy. Traditional machine learning models, such as single SVM and decision trees, have limited ability to fit complex nonlinear data and cannot effectively distinguish similar characteristic scenarios such as residual coal oxidation and raw coal escape, resulting in a high misjudgment rate of CO sources related to spontaneous coal combustion and affecting the reliability of early warning.

[0025] Based on this, embodiments of this application provide a method, apparatus, electronic device, and medium for classifying and identifying CO concentration sources, to solve the problem of low accuracy in CO source identification caused by insufficient feature extraction and limited model recognition capabilities in the prior art. Specifically, by employing a deep neural network model to automatically extract curve features from the CO concentration sequence and combining them with a classification and recognition model for classification and identification, accurate differentiation and classification of CO concentration changes from different sources are achieved, thereby improving the accuracy and reliability of CO concentration source identification in underground coal mines.

[0026] Please see Figure 1 , Figure 1 This is a flowchart illustrating the CO concentration source classification and identification method provided in an embodiment of this application. Figure 1 As shown in the embodiments of this application, the CO concentration source classification and identification method includes: S101. Collect raw CO concentration data from different sources, crop the target CO concentration data, and construct a labeled CO concentration sample database.

[0027] Raw CO concentration data from multiple sources were collected, stable band data were obtained by cropping, CO source labels were established, and a sample database of CO concentrations caused by different factors was constructed.

[0028] CO concentration data comes from five sources: O represents spontaneous combustion oxidation of coal, A represents blasting operations, B represents exhaust gas from transportation equipment, C represents roof collapse, and D represents raw coal leakage. Raw coal leakage is relatively low and has a minor impact, so it is usually not considered. Spontaneous combustion oxidation of coal easily leads to spontaneous combustion disasters and is the primary target for identification; the label can be set to O, A, B, and C after removing raw coal leakage.

[0029] The CO source classification and labeling system constructed in this application covers four main categories: spontaneous combustion of coal (O2), blasting operations (A), exhaust gas from transportation equipment (B), and roof collapse (C). Firstly, CO concentration data can be collected through three methods: numerical simulation, laboratory experiments, and field experiments.

[0030] In a preferred embodiment, numerical simulation generates CO concentration data by establishing a coal mine ventilation-gas diffusion model and setting different boundary conditions, such as the temperature of the residual coal oxidation zone and the instantaneous CO release during blasting operations; laboratory experiments simulate the processes of coal spontaneous combustion and equipment exhaust gas in a controlled environment such as a large coal spontaneous combustion test bench; and field experiments deploy sensors at key locations such as the coal mining face, return airway, and goaf in the coal mine for actual monitoring.

[0031] The collected parameters include not only core data such as CO concentration, timestamp, and monitoring point location (CO concentration is in ppm), but also multi-dimensional parameters strongly correlated with CO generation, including environmental parameters such as temperature and oxygen concentration, as well as detailed operating information such as the start and stop of mining machinery, blasting operation time, and the operating status of transportation equipment.

[0032] Secondly, the original continuous monitoring data is extracted to retain the complete "stable band" of CO concentration from the beginning of the change to the return to a stable state. This band should be able to fully characterize the concentration change characteristics of a specific source. Then, each stable band is cross-validated based on multi-source information such as on-site records and video monitoring, and the specific source category is manually labeled, including: coal spontaneous combustion oxidation, blasting operations, exhaust gas from transportation equipment, and roof collapse, etc.

[0033] Finally, all these labeled stable band data from different sources were organized into a unified, structured CO concentration sample database, with each sample containing concentration sequence data and corresponding source labels.

[0034] This application establishes a high-quality, multi-source labeled database, providing a solid data foundation for subsequent waveform feature-based classification and recognition, and effectively solving the data acquisition limitations and sample labeling problems mentioned in the background technology.

[0035] Specifically, please refer to Figure 2 , Figure 2The flowchart for constructing a labeled CO concentration sample database provided in the embodiments of this application is as follows: Figure 2 As shown, a labeled CO concentration sample database is constructed using the following method: S201. The original CO concentration data is cropped to retain stable bands that contain the complete concentration change process.

[0036] The stable band refers to the data segment from when the concentration begins to change until it returns to stability. The stable band includes the target CO concentration data.

[0037] Here, the algorithm automatically detects when the concentration values ​​of multiple consecutive data points in the concentration sequence, such as 3 to 5 points, exceed the baseline stable value by a certain range, for example, exceeding the baseline average value by 3 times the standard deviation. This moment is determined as the starting point of the concentration change.

[0038] After the concentration reaches its peak and begins to decline, the end point of the change event is determined when the concentration value falls back to the baseline fluctuation range, for example, within the range of plus or minus 10 percent of the baseline average value, and remains there for a predetermined period of time, such as 60 seconds.

[0039] Starting from the identified starting point and ending point, this complete CO concentration data segment is extracted to form a stable band.

[0040] In one embodiment, a momentary spike is generated during the blasting operation. The stable band begins from the period of stable concentration before the blast and ends when the concentration returns to the normal background value after the spike.

[0041] In another embodiment, the spontaneous combustion oxidation of coal exhibits a slow, continuous upward process. The stable band begins when the concentration starts to deviate from the baseline and continues until a new stable state is formed at a higher concentration level or the concentration begins to decline.

[0042] S202. Establish CO source labels for each stable band and construct a sample database of CO concentrations caused by different reasons.

[0043] CO source labels are used to identify the specific source of a stable band.

[0044] In numerical simulation, the labels are directly determined by the boundary conditions and operating parameters set during the simulation. For example, if the scenario is set as spontaneous combustion of coal, then all the generated data will be labeled as spontaneous combustion oxidation of coal.

[0045] Regarding laboratory experimental data, the labeling is determined by the experimental design and control conditions. For example, if diesel engine exhaust is simulated on a test bench, the corresponding data will be labeled as transportation equipment exhaust.

[0046] Regarding field experimental data, the labels need to be determined through multi-source information fusion and cross-validation: reviewing mine operation logs and scheduling records to clarify the specific operational activities that occurred at the data collection time point; retrieving industrial video surveillance to visually confirm the underground operation situation; interviewing on-site operators to verify production activities within a specific time period; and accurately matching the above information with the timestamps of CO concentration changes and the location of monitoring points to assign an accurate CO source label to each stable band.

[0047] All labeled stable bands are associated with their corresponding source labels and organized into a structured database, typically stored in tabular or specific data file formats. Each record contains the sample ID, CO concentration sequence (i.e., stable band data), source label, and other possible metadata, such as acquisition time, location, and environmental parameters.

[0048] This step addresses the core issues of difficult sample labeling and class imbalance in the background technology. By validating data from multiple sources, the accuracy of the labels is improved, avoiding the subjective bias of manual labeling. The final database is a high-quality dataset with reliable labels, covering diverse sources, and suitable for training supervised learning models.

[0049] S102. Preprocess the target CO concentration data to obtain a standardized CO concentration sequence.

[0050] Preprocessing of the target CO band data using 3σ goal The method of handling outliers, Holt interpolation to handle missing values, empirical mode decomposition and reconstruction to remove periodicity, and normalization processing are used to obtain a CO concentration sequence applicable to the detection and classification of CO concentration changes caused by various reasons.

[0051] Specifically, please refer to Figure 3 , Figure 3 A flowchart illustrating the standardized CO concentration sequence provided in this application embodiment is shown below. Figure 3 As shown, the standardized CO concentration sequence was obtained in the following manner: S301. Use preset criteria to process outliers in the target CO concentration data of the stable band.

[0052] Preset criteria are used to identify and remove data points that deviate from the average data.

[0053] For example, the preset criterion can be 3σ. goal Guidelines.

[0054] First, calculate the arithmetic mean μ of the CO concentration data segment. goal and standard deviation σ goal Then, iterate through all data points and find those whose values ​​are not in the interval [μ].goal -3σ goal ,μ goal +3σ goal Points within the range are identified as outliers. For identified outliers, especially consecutive outlier segments, they are not simply deleted, but replaced by linear interpolation of the preceding and following valid data points to maintain the continuity of the data sequence.

[0055] Among them, the decision interval [μ] goal -3σ goal ,μ goal +3σ goal According to the characteristics of normal distribution, data points outside this interval have an extremely low probability and are considered anomalies.

[0056] It should be noted that the arithmetic mean μ goal and standard deviation σ goal It is calculated from the target CO concentration data obtained after cropping, i.e. the original data within the stable band, and is used to identify and remove abnormal data points.

[0057] S302. Use interpolation to fill in the missing values ​​in the target CO concentration data of the stable band.

[0058] Interpolation is used to estimate missing values ​​based on time series trends.

[0059] For example, the interpolation method is Holt interpolation, also known as Holt-Winters two-parameter exponential smoothing. This method considers not only the current level L of the data, but also... t Also consider its changing trend T t It uses two smoothing equations: Horizontal equation: L t =α×x t +(1-α)×(L t-1 +T t-1 ); Trend equation: T t =β×(L t -L t-1 )+(1-β)×T t-1 .

[0060] Where, x t Let L be the observed CO concentration at time t. t Let T be the estimated level value at time t. t Let L be the estimated trend value at time t. t-1 Let T be the estimated level value at time t-1. t-1The estimated trend value at time t-1 is determined by optimizing the horizontal smoothing parameter α and the trend smoothing parameter β on the complete historical data through grid search, with the optimization objective being to minimize the prediction error.

[0061] Based on the valid data segments before and after the missing values, the horizontal and trend components are calculated iteratively to predictively fill in the missing segments in accordance with the time series trend.

[0062] In this step, compared with traditional linear interpolation, Holt interpolation can better capture and continue the inherent trend of CO concentration changes, avoid the situation where the filler value contradicts the actual change law of the environment, and ensure the physical meaning and continuity of the data sequence.

[0063] S303. To address the periodic fluctuation interference in the target CO concentration data in the stable band, the empirical mode decomposition reconstruction method is used to remove the periodic interference components in the target CO concentration data in the stable band, resulting in deperiodic CO concentration data.

[0064] This step aims to filter out periodic fluctuations in the data caused by regular daily operations such as shift work, scheduled ventilation, and centralized transportation, which are unrelated to the target CO source. By purifying the signal, subsequent analysis can focus more on the concentration changes caused by specific events such as spontaneous combustion of coal and explosions.

[0065] This step involves a three-step process of decomposition, identification, and reconstruction.

[0066] a. Decomposition Process: The preprocessed CO concentration sequence X(t) is input into the Empirical Mode Decomposition (EMD) algorithm. The EMD algorithm adaptively decomposes the complex signal X(t) into a set of simple intrinsic mode functions (IMFs) arranged from high frequency to low frequency: c1(t), c2(t), ..., c n (t) and a residual term r representing the overall trend n (t).

[0067] For example, this can be understood as separating an audio signal (original signal) containing various sounds into individual frequencies of different pitches. For instance, the first intrinsic mode function (IMF1) is like a soprano, the second IMF2 is like a baritone, and so on. Simultaneously, a background sound is also separated, such as the residual term representing the overall trend of the signal. Each IMF component corresponds to an oscillation at a different frequency, and the residual corresponds to the long-term trend of the signal.

[0068] b. Identification and Removal Process: For each IMF component obtained from decomposition, such as c2(t), perform a Fast Fourier Transform (FFT) to analyze its frequency components and find its dominant frequency, i.e., the most important fluctuation period; compare the dominant frequency of the IMF with the periodic frequency range preset according to the actual working system of the mine; determine the IMF components whose dominant frequencies fall within the preset periodic range as periodic interference components.

[0069] c. Reconstruction process: Remove all IMF components identified as periodic interferences from the original decomposition results.

[0070] The remaining, non-periodic IMF components are combined with the residual term r. n Adding (t) together, we reconstruct a new CO concentration sequence X. new (t).

[0071] The formula is expressed as: X new (t)=Σ(non-periodic IMF)+r n (t) A deperiodic CO concentration sequence was obtained, in which regular fluctuations introduced by routine fixed operations were significantly suppressed. Traditional methods may use fixed filters, which are difficult to adapt to complex non-stationary signals in underground mines. The adaptability of the EMD method allows it to more accurately remove interference from the data itself, without relying on manually preset fixed parameters, thus making it more robust and effective in processing data from complex environments such as coal mines.

[0072] S304. Perform normalization processing on the processed deperiodic CO concentration data to convert it to a uniform numerical scale and obtain a standardized CO concentration sequence.

[0073] For example, normalization usually refers to min-max normalization, which normalizes each independent stable band separately. A stable band refers to the data segment clipped from S201. The calculation formula is: X normalized =(XX min )÷(X max -X min ).

[0074] Where X is the deperiodized CO concentration data, X min and X max These are the minimum and maximum values ​​within the stable band, respectively, X. min and X max It can be calculated in real time from the currently processed stable band, and the processed data is scaled to the range of [0,1].

[0075] Normalization eliminates the absolute dimensional differences between CO concentration events from different sources and with varying intensities. For example, it standardizes the comparison between a blasting pulse of 1000 ppm and a slow rise in coal spontaneous combustion oxidation of 50 ppm, bringing them to the same numerical scale. This ensures that subsequent machine learning models are not dominated by high-volume data and can fairly learn morphological features across all categories, a crucial step in guaranteeing model convergence and classification performance.

[0076] S103. Extract features from the standardized CO concentration sequence to obtain curve features that characterize the concentration variation patterns of different sources, and construct a CO concentration sequence feature vector.

[0077] DNN was used to extract features from CO concentration sequences in different coal mines, and corresponding quantitative characterization models were established. Curve features of CO concentration changes caused by different sources were proposed. Based on the CO concentration sequence features from different sources, a feature vector of CO concentration sequences in coal mines was constructed.

[0078] For example, time-domain statistical feature extraction can start from the basic statistical properties of the standardized CO concentration sequence and extract the following key indicators: Basic statistics include the maximum value C. max Minimum value C min Average value μ feat , median C med Standard deviation σ feat .

[0079] Morphological characteristics include peak time T peak Ascent time T rise decay time T decay .

[0080] Among them, peak time is the time taken from the start of the sequence to reaching the peak; rise time is the time required to rise from the baseline to the peak; decay time is the time required to decay from the peak to a stable value.

[0081] Distribution characteristics include skewness S kew and kurtosis K urt .

[0082] It should be noted that the average value μ feat and standard deviation σ feat These are characteristic parameters calculated based on the standardized CO concentration sequence, and are different from μ calculated based on the target CO concentration data in S301. goal and σ goal The objects they calculate and their functions are different.

[0083] Skewness is used to measure the asymmetry of concentration distribution; kurtosis is used to reflect the sharpness of concentration distribution.

[0084] It should be noted that the minimum value X here... min and maximum value X max It is only used for normalization calculations and is not a statistical feature extracted later.

[0085] Features of parametric model fitting: Based on the physical mechanisms of different CO sources, a specific mathematical model is used for fitting, and the model parameters are extracted as features: The exponential decay model is applicable to scenarios such as roof collapse, blasting, and other instantaneous releases followed by natural decay processes.

[0086] The logical / exponential growth model is applicable to scenarios involving the cumulative effects of slow oxidation processes such as spontaneous combustion of coal.

[0087] The Gaussian / impulse model is suitable for short-lived events such as vehicles passing by or equipment starting and stopping.

[0088] Specifically, please refer to Figure 4 , Figure 4 The flowchart for constructing the CO concentration sequence feature vector provided in the embodiments of this application is as follows: Figure 4 As shown, the CO concentration sequence feature vector is constructed in the following way: S401. Input the standardized CO concentration sequence into the deep neural network module and output the curve features representing the standardized CO concentration sequence.

[0089] The deep neural network module is used to characterize the relationship between the standardized CO concentration sequence and the curve features, which are features extracted from the standardized CO concentration sequence that can distinguish the morphology and patterns of different sources.

[0090] Deep neural networks (DNNs) can employ either one-dimensional convolutional neural networks (1D-CNNs) or multilayer perceptrons (MLPs). Taking 1D-CNN as an example, its input layer receives a fixed-length, standardized CO concentration sequence, such as 300 data points corresponding to 5 minutes of monitoring data. The network contains multiple convolutional layers, pooling layers, and fully connected layers.

[0091] The network's weights and biases are core parameters. These parameters are not set manually, but are automatically learned by supervising the network through a labeled CO concentration sample database built using S101, with the training aiming at classification accuracy.

[0092] The goal of training is to enable the DNN to map CO concentration sequences from different sources to different output categories. During this process, the intermediate layers of the network will automatically learn the key features that can distinguish these categories.

[0093] After training, a new standardized CO concentration sequence is input into the DNN. Instead of using the network's final output (classification result), the output of the last fully connected layer, also known as the bottleneck layer, is used as the curve features. This output is a high-dimensional vector, such as 128 or 256-dimensional, where each value is an abstract representation of the input sequence by the DNN. These features comprehensively characterize the deep-seated patterns of the concentration sequence, including its morphology, trend, and fluctuation patterns, such as the slow upward slope, the sharpness of the pulses, and the intervals of the multi-peak distribution.

[0094] This step addresses the shortcomings in the feature engineering stage of the background technology. It utilizes DNNs to automatically learn and extract deep curve features, replacing the shallow statistical features that rely on human experience, thereby improving the discriminative power and information richness of the features.

[0095] S402. Combine the curve features to form a CO concentration sequence feature vector.

[0096] The combination of these steps here means directly using this high-dimensional vector as the output of this step: the CO concentration sequence feature vector. This CO concentration sequence feature vector can be represented as V=[f1,f2,f3,...,f n ], where n is the feature dimension, i.e., the number of features, f i It is the numerical value of each feature calculated by the DNN.

[0097] For example, the feature vector can be represented as: V=[C max , μ feat ,T rise ,T decay ,S kew ,K urt That is: V = [maximum value, average value, rise time, decay time, skewness, kurtosis].

[0098] This application visualizes and quantifies the abstract curve features learned by the DNN into a structured data representation, namely a feature vector. This vector provides input for subsequent feature optimization S104 and classification recognition S105.

[0099] In summary, step S103 uses a pre-trained DNN model to transform a CO concentration curve into a high-dimensional feature vector containing its essential pattern, laying a solid foundation for achieving high-precision source classification.

[0100] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the feature importance analysis provided in an embodiment of this application.

[0101] like Figure 5As shown in the figure, this plot illustrates the global importance scores of multiple time-domain statistical, morphological, and distribution features extracted from the standardized CO concentration sequence in the classification model. The horizontal axis represents the median importance score of each feature; a higher score indicates a greater contribution of the feature to distinguishing different CO sources. The vertical axis lists the names of the evaluated features, including maximum, minimum, mean, standard deviation, kurtosis, skewness, rise time, peak time, and decay time. Figure 5 It can be seen that the maximum value and rise time have the highest importance scores, indicating that these two features are the most discriminative in distinguishing different sources such as spontaneous combustion oxidation of coal, blasting operations, exhaust gas from transportation equipment, and roof collapse. Statistical features such as kurtosis, skewness, standard deviation, and mean also have high importance; while the peak time, minimum value, and decay time have relatively low importance scores. This result verifies that the feature extraction method described in this application can effectively capture the key patterns of CO concentration changes from different sources, providing an important basis for subsequent feature selection and dimensionality reduction processing.

[0102] S104. Perform feature selection and dimensionality reduction on the CO concentration sequence feature vector to obtain the optimized low-dimensional feature vector.

[0103] Feature selection and dimensionality reduction are performed on the feature vector of CO concentration sequence in underground coal mines to establish an optimized low-dimensional feature vector.

[0104] Specifically, please refer to Figure 6 , Figure 6 The flowchart for obtaining the optimized low-dimensional feature vector provided in the embodiments of this application is as follows: Figure 6 As shown, the optimized low-dimensional feature vector is obtained in the following way: S501. A recursive feature elimination method is used to select features from the CO concentration sequence feature vector to remove unimportant features.

[0105] Recursive Feature Elimination (RFE) is a wrapper-style feature selection method that evaluates the importance of features by recursively building a model.

[0106] First, after initializing all features, an iterative loop is entered. Then, a base model is trained using the current feature set, and the weight coefficients corresponding to each feature in the model are obtained. The absolute value of each weight coefficient represents the importance of that feature to the classification decision. Features with the smallest absolute weight coefficients are removed, for example, 10% are removed each time. The above process is repeated based on the new feature subset.

[0107] Finally, the iteration stops when the number of features decreases to a preset target value. The optimal target number of features is determined through cross-validation, weighing model performance such as accuracy against the number of features.

[0108] S502. After feature selection, principal component analysis is used to reduce the dimensionality of the features to obtain the optimal low-dimensional feature vector.

[0109] Principal Components Analysis (PCA) is an unsupervised dimensionality reduction method. Its core idea is to perform a coordinate transformation to map the original features to a new set of uncorrelated orthogonal features, namely principal components, and then sort them in descending order of variance.

[0110] First, the feature data after RFE screening is assumed to be k1-dimensional and is centered (mean-removed). Then, the covariance matrix of the data is calculated.

[0111] Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; the eigenvalues ​​are sorted in descending order, and their corresponding eigenvectors are the principal component directions.

[0112] Finally, the principal components corresponding to the top k2 largest eigenvalues ​​are selected, and the original data is projected onto the new space formed by these k2 principal components.

[0113] The number of principal components, k2, is a core output parameter of PCA. It is usually determined by setting the cumulative variance contribution rate, such as 95%, which means that the first k2 principal components selected can explain 95% of the variance of the original data.

[0114] Eigenvalues ​​and eigenvectors are calculated using the PCA algorithm, and the magnitude of the eigenvalue represents the amount of information carried by the corresponding principal component.

[0115] S105. Input the optimized low-dimensional feature vector into the classification and recognition model to determine the classification and recognition results of CO concentration sources.

[0116] Specifically, please refer to Figure 7 , Figure 7 A flowchart illustrating the classification and identification results for determining the source of CO concentration provided in this application embodiment is shown below. Figure 7 As shown, the classification and identification results of CO concentration sources were determined using the following methods: S601. Input the optimized low-dimensional feature vector into the classification and recognition model to obtain the corresponding source classification probability distribution.

[0117] Among them, the classification and recognition model is used to characterize the mapping relationship between the optimized low-dimensional feature vector and the CO source category.

[0118] Preferably, the classification and recognition model is a hybrid model of DNN-SVM.

[0119] This is a cascaded structure: DNN part: that is, the deep neural network trained in S103, whose parameters, weights and biases have been fixed, and here it acts as a fixed, high-performance feature extractor.

[0120] The SVM (Support Vector Machine) part: This is a support vector machine classifier, and it is the core of this step. Its role is to learn the complex mapping relationship from optimizing low-dimensional feature vectors to the source classes of CO.

[0121] Furthermore, the classification and recognition model can be trained in the following ways: Obtain a training sample set, which includes multiple training samples. Each training sample includes the optimized low-dimensional feature vector and the corresponding sample CO source label.

[0122] The optimized low-dimensional feature vector of the sample is used as the input to the classification and recognition model to be trained, and the corresponding sample CO source label is used as the output of the classification and recognition model to be trained.

[0123] During the SVM training phase, the DNN parameters remain unchanged, and only the parameters of the SVM part are trained. For nonlinear SVMs, such as those using the RBF kernel function, key parameters include the penalty factor C and the kernel function parameters. The penalty factor C controls the model's tolerance for misclassified samples; the kernel function parameters, such as the γ of the RBF kernel, control the width of the kernel function, affecting the distribution of samples in the feature space. These parameters are determined through grid search combined with cross-validation, with the optimization objective being to maximize classification accuracy and other metrics.

[0124] The penalty factor, often represented by the letter C, is also known as the regularization parameter or penalty coefficient. It is used to control the complexity and tolerance of the model and defines the severity of the penalty imposed by the model on misclassified samples.

[0125] Kernel functions are used to solve nonlinear classification problems. Essentially, they are mapping functions that can map linearly inseparable data in the original low-dimensional space to a higher-dimensional feature space, making the data linearly separable in this new space.

[0126] Kernel functions do not actually compute the complex mappings in high-dimensional space. Instead, they obtain the inner product result in high-dimensional space through a function that can be computed in the original space, namely the kernel function. This is called the kernel trick, which can save computational overhead.

[0127] The optimized low-dimensional feature vector of the sample to be identified is input into the trained DNN-SVM hybrid model.

[0128] The model typically calculates the probability that a sample belongs to each preset CO source category using probability output methods such as Platt scaling.

[0129] The output is a probability distribution vector. In an optional embodiment of this application, the probability P = [P O =0.02, P A =0.80, P B =0.05, P C =0.03], where O represents spontaneous combustion and oxidation of coal, A represents blasting operations, B represents exhaust gas from transportation equipment, and C represents roof collapse. This distribution characterizes the model's confidence in the source of the sample.

[0130] For further details, please refer to Figure 8 , Figure 8 The diagram provided in this application illustrates how to combine PCA and SVM to classify CO sources, demonstrating how to identify CO source categories through dimensionality reduction and classifiers.

[0131] The elements in the diagram and their technical meanings in this application are as follows: Original feature space: A two-dimensional plane established with horizontal coordinate X1 and vertical coordinate X2, representing two initial feature dimensions extracted from the CO concentration sequence, such as peak concentration and rate of increase. This space actually corresponds to a two-dimensional projection view of the high-dimensional CO concentration sequence feature vector extracted and initially constructed by the deep neural network in this application.

[0132] The space after PCA transformation is a new coordinate system established with X1' (the first principal component axis) and X2' (the second principal component axis). The X1' axis points to the optimal projection direction, i.e., the direction with the largest data variance, representing the most significant difference in CO concentration variation patterns from different sources. The X2' axis is the orthogonal direction with the second largest variance. In this application, principal component analysis (PCA) is used to reduce the dimensionality of high-dimensional feature vectors, aiming to filter out noise, retain the most discriminative information, and form optimized low-dimensional feature vectors for classification.

[0133] Category Distribution and Projection: The scatter clouds of ω1 and ω2 in the figure represent two different CO sources. For example, ω1 corresponds to spontaneous combustion oxidation of coal, and ω2 corresponds to blasting operations. The projections of ω1 and ω2 show the distribution range of these sample points along the X1' axis, i.e., the principal component direction, intuitively reflecting the separability of different categories of samples after PCA compression.

[0134] SVM decision-making process: Maximum Margin Direction and Decision Hyperplane: The goal of an SVM classifier is to find a decision hyperplane, represented as a straight line in a two-dimensional graph, to maximize the margin between the two classes. The maximum margin direction is the direction perpendicular to this hyperplane, and it determines the optimal segmentation boundary of the classifier.

[0135] Decision Regions: The regions marked "classified as ω1" and "classified as ω2" in the diagram are the classification decision regions defined by the SVM based on the decision hyperplane. Sample points falling within the corresponding regions will be classified by the classifier as the corresponding CO source category.

[0136] Angle θ: Represents the angle between the original feature axis X1 and the PCA principal component axis X1', or the angle between the decision boundary and a certain coordinate axis. It quantifies the degree of PCA rotation transformation of the feature space, or the spatial orientation of the decision boundary, and is the geometric expression of linear transformation.

[0137] Figure 8 The correspondence between the workflow shown and the method of this application is as follows: Feature dimensionality reduction: Corresponding to step S104 of this application, the high-dimensional CO concentration sequence feature vector is projected onto a low-dimensional principal component space such as X1' and X2' through PCA to obtain the optimized low-dimensional feature vector.

[0138] Classification and Recognition: Corresponding to step S105 of this application, the dimensionality-reduced feature vector is input into the SVM classifier. The SVM learns the distribution of samples such as ω1 and ω2 in the low-dimensional space and determines a decision hyperplane with the maximum margin, thereby achieving accurate classification of CO concentration sources.

[0139] S602. Based on the probability distribution of source classification, classify the categories to obtain the classification and identification results of CO concentration sources.

[0140] In an optional embodiment of this application, the determination rule can be the maximum probability rule, that is, the category with the largest value in the probability distribution vector P is selected as the final classification result.

[0141] For example, in the embodiment of S601 described above, P A =0.80, which is much higher than other probabilities, therefore it is determined that the CO concentration sequence originated from blasting operations.

[0142] In another optional embodiment of this application, a probability threshold, such as 0.6, can be set. The classification result is adopted only when the maximum probability value exceeds this threshold; otherwise, the sample is marked as having an unknown origin or requires manual judgment. This helps control the risk of misjudgment and improves the reliability of the system.

[0143] Based on the optimization of low-dimensional feature vectors, a classification and identification system for CO concentration sequences in underground coal mines based on DNN-SVM is established, including CO concentration sequence classification and source identification.

[0144] The model is trained and cross-validated by dividing it into known and unknown samples. The known samples are used for training and testing, while the unknown samples are used for validation. This establishes a mapping relationship between CO concentration sequences from different sources and classification labels.

[0145] Furthermore, the results output by the classification model are evaluated using metrics such as accuracy, recall, and F1-score to optimize model parameters and ensure the reliability of the classification system.

[0146] The reserved test dataset is input into the trained classification model to obtain the prediction results.

[0147] For further details, please refer to Figure 9 , Figure 9 This is a schematic diagram of the CO concentration source classification results provided in the embodiments of this application.

[0148] This figure, presented as a scatter plot, illustrates the model's true classification performance on the reserved test set. The horizontal axis represents the sample number, and the vertical axis represents four predefined event categories, from top to bottom: blasting operations, exhaust gas from transportation equipment, roof collapse, and spontaneous combustion oxidation of coal.

[0149] In the diagram, blue circles represent the true labels of the samples, and red stars represent the predicted labels of the model. From Figure 9 It can be seen that the predicted results for the three categories of coal spontaneous combustion oxidation, roof collapse, and blasting operations are highly consistent with the true labels. The blue circles and red stars in most samples overlap or are close in position, indicating that the model has a high accuracy rate in identifying CO concentration changes from these sources. In the transportation equipment exhaust category, a small number of samples show predicted labels that deviate from the true labels. This may be because the CO concentration change characteristics of this source are similar to those of other sources, leading to some false positives from the model. Overall, the model's predictions in all four categories have a high degree of consistency with the true labels, and the classification results are reliable.

[0150] Then, the predicted results are compared with the true labels, and a series of evaluation indicators are calculated, as shown in Table 1 below: Table 1

[0151] Accuracy: The proportion of correctly classified samples out of the total sample. It is applicable when the number of samples in each class is balanced.

[0152] Recall: The proportion of samples correctly identified by the model for a specific source, such as coal mining machine cutting coal, out of the total number of samples from that source. It focuses on missed cases.

[0153] Precision: The proportion of samples correctly predicted by the model from a given source. Pay particular attention to false positives.

[0154] F1-score: The harmonic mean of precision and recall, which is a metric for comprehensively evaluating model performance.

[0155] Based on the above evaluation results, if the model performance does not meet expectations, parameter tuning is performed. Automated methods such as grid search or random search are typically used to find the optimal combination of hyperparameters, such as the learning rate, maximum tree depth, and regularization coefficient, within a predefined parameter combination space to achieve the highest cross-validation score, e.g., F1-score.

[0156] Then, SHAP was used to perform a visualization and interpretation analysis of the characteristics of CO concentration sequences from different sources, and the contribution rate of each feature to the classification results was analyzed.

[0157] Using a dedicated SHAP library, an interpreter can be created for the best-trained classification model, such as TreeExplainer for tree models.

[0158] An interpreter is a mathematical mapping tool that decomposes and quantifies the prediction results of a complex black-box model into the contribution values ​​of each input feature, i.e., SHAP values, by building a local surrogate model or computing feature perturbations.

[0159] The interpreter is run on a test set or a representative sample set to compute a SHAP value for each feature of each sample. This value quantifies the contribution of that feature to the final prediction result for that sample, whether it positively boosts or negatively drags it down.

[0160] For example, the visualization analysis method is shown below: In one embodiment, the summary plot shows the global ranking of the importance of all features to the model output and can display the relationship between feature values ​​and SHAP values, i.e., influence.

[0161] Please see here. Figure 10 , Figure 10 This is a schematic diagram of the SHAP feature contribution summary provided in the embodiments of this application. The diagram shows the global ranking of the influence of each feature on the model output and intuitively reveals the non-linear relationship between the feature value and the classification result.

[0162] Figure 10The vertical axis is arranged vertically from high to low according to the average global contribution of features to the model's prediction results, i.e., the Mean Absolute Power (SHAP) value. Each row in the graph represents a feature, sorted from top to bottom according to the average influence of the features on the model's predictions; each point in the graph represents a sample. The horizontal axis represents the SHAP value. A positive value indicates that the feature has a positive effect on the model's classification of the sample, while a negative value indicates a negative effect. The color of the point represents the original value of the feature in the corresponding sample; the redder the color, the larger the feature value, and the bluer the color, the smaller the feature value. The SHAP value on the horizontal axis reflects whether the feature has a positive or negative effect on the model's final classification decision. A positive effect is indicated by a SHAP value greater than zero, and a negative effect is indicated by a SHAP value less than zero.

[0163] from Figure 10 It can be seen that features such as maximum value and rise time have high average absolute SHAP values, indicating that they are the main basis for model decision-making. Taking rise time as an example, when the scatter points are red and concentrated on the positive half of the horizontal axis, it indicates that the higher the value of this feature, the stronger its positive driving effect on the current classification result. This quantitative relationship reveals the nonlinear decision-making logic inside the model, that is, specific sources such as coal spontaneous combustion oxidation are usually accompanied by an abnormal upward trend of CO concentration index.

[0164] For features such as mean and standard deviation, higher or lower feature values ​​show a clear positive-negative separation trend on their SHAP value axis. The distribution width of the scatter points along the horizontal axis reflects the fluctuation range of feature influence. The SHAP value distribution of each feature in the figure shows a clear clustering effect, and the high feature values ​​(red dots) and low feature values ​​(blue dots) have good distinguishability on the positive and negative half axes. This proves that the features extracted in this application have strong discriminative power, and the model demonstrates high decision consistency and generalization ability when processing CO monitoring data from different sources. This visualization result not only makes the classification decision process of the black-box model transparent, but also provides a physically meaningful reference for analyzing the generation mechanism and evolution law of different CO sources.

[0165] This application constructs and implements an automatic classification and identification method for CO concentration sequences from different sources in underground coal mines based on multi-dimensional feature vectors, forming an identification system that can automatically classify CO concentration fluctuations in underground coal mines into different sources, thereby enabling real-time monitoring and early warning support.

[0166] This application saves the trained optimal model as a file and deploys it to a production server; develops a data access interface to receive real-time or near-real-time CO concentration sequence data from a coal mine monitoring system; and develops processing logic to automatically perform feature extraction, model prediction, and result output on newly input data.

[0167] The final output of this application is no longer just a single concentration exceedance alarm, but an intelligent early warning with source tags. For example: Alarm: The current CO concentration is fluctuating abnormally, identified as a source of coal spontaneous combustion oxidation. Furthermore, it can operate 24 hours a day, and once a dangerous source such as coal spontaneous combustion oxidation is identified, it will immediately issue an early warning to safety personnel through monitoring screens, SMS, App push, etc.

[0168] This application utilizes a DNN-SVM hybrid model, combining the powerful feature representation learning capabilities of DNN with the classification performance of SVM in small sample, high-dimensional feature spaces, to improve the classification accuracy of complex nonlinear CO concentration change patterns. In particular, it effectively distinguishes easily confused scenarios such as coal spontaneous combustion oxidation and raw coal escape. Furthermore, by outputting probability distributions and setting judgment rules, it provides transparency and controllability in decision-making, reducing the risk of misjudgment and omission of key hidden dangers such as coal spontaneous combustion.

[0169] The CO concentration source classification and identification method, device, electronic device and medium provided in this application solve the problems of low CO source identification accuracy and high false positive rate caused by insufficient feature extraction and limited model recognition capability in traditional monitoring methods by constructing a sample database that integrates multi-source data, automatically extracting deep features of CO concentration change curves using deep neural networks, and using a DNN-SVM hybrid model for accurate classification. This achieves automated and high-precision identification and classification of CO concentration changes from different sources in coal mines, effectively improving the accuracy and reliability of early warning of coal spontaneous combustion.

[0170] This application provides a feasible method for studying the characteristics of CO concentration variations from different sources in underground coal mines and for intelligently identifying the causes of CO. This method collects CO gas concentrations from different sources through numerical simulations, laboratory experiments, and field experiments, establishing a database of typical CO concentration variations that integrates various parameters related to the underground coal mine environment and operating conditions. By extracting features from the trends of CO concentration variations from different sources, and combining statistical analysis and machine learning methods to classify and identify CO concentration variation waveforms, a system for identifying the sources of CO concentration in underground coal mines based on the characteristics of CO concentration variation curves is constructed.

[0171] This application addresses several key issues: Firstly, it solves the problem that in coal mine CO concentration monitoring, the complex formation environments of CO from different sources and their similar concentration trends make it difficult and inefficient to manually distinguish the causes of CO, thus affecting the timeliness and accuracy of early warnings for potential hazards such as spontaneous combustion of coal. Secondly, through multi-scenario data collection combining numerical simulation, laboratory experiments, and field experiments, a CO concentration database integrating environmental and operational conditions is constructed. Furthermore, by employing a targeted feature extraction, statistical analysis, and machine learning-based identification method, it overcomes the limitations of traditional CO monitoring methods, which only obtain concentration values ​​but cannot correlate with causes, and suffer from low identification accuracy due to single data sources and analytical models. This provides a theoretical basis and application foundation for the automated and precise identification of CO concentration sources in coal mines.

[0172] In the field of coal spontaneous combustion monitoring and early warning, CO, as a hallmark early-stage gas released during the coal spontaneous combustion process, exhibits a strong coupling relationship between its concentration changes and the stage of coal spontaneous combustion, making it a core indicator for assessing the risk level of coal spontaneous combustion. However, in the current coal spontaneous combustion monitoring and early warning technology system, CO monitoring is limited to concentration quantification and cannot achieve source tracing and risk attribution, leading to delays and misjudgments in early warning of coal spontaneous combustion.

[0173] The implementation of this application improves the accuracy of CO source identification, effectively solves the problem of misjudgment and omission of spontaneous combustion of coal caused by source confusion in traditional methods, provides early warning for spontaneous combustion hazards such as oxidation of residual coal found in mines, reduces the risk of major disasters such as gas explosions caused by spontaneous combustion of coal, and has practical value for ensuring safe production in coal mines and improving the level of safety monitoring in the industry.

[0174] Based on the same inventive concept, this application also provides a CO concentration source classification and identification device corresponding to the CO concentration source classification and identification method. Since the principle of the device in this application is similar to the CO concentration source classification and identification method described above in this application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0175] Please see Figure 11 , Figure 11 This is a schematic diagram of the CO concentration source classification and identification device provided in an embodiment of this application. Figure 11 As shown, the CO concentration source classification and identification device 800 includes: CO concentration sample database construction module 801 is used to collect raw CO concentration data from different sources, crop the target CO concentration data, and construct a labeled CO concentration sample database. CO concentration sequence standardization module 802 is used to preprocess the target CO concentration data to obtain a standardized CO concentration sequence; CO concentration sequence feature vector component module 803 is used to extract features from the standardized CO concentration sequence, obtain curve features that characterize the concentration variation patterns of different sources, and construct a CO concentration sequence feature vector. The optimal low-dimensional feature vector construction module 804 is used to perform feature selection and dimensionality reduction on the CO concentration sequence feature vector to obtain the optimal low-dimensional feature vector. The classification and identification result determination module 805 is used to input the optimized low-dimensional feature vector into the classification and identification model to determine the classification and identification result of the CO concentration source.

[0176] Please see Figure 12 , Figure 12 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. Figure 12 As shown, the electronic device 900 includes a processor 810, a memory 820, and a bus 830.

[0177] The memory 820 stores machine-readable instructions executable by the processor 810. When the electronic device 900 is running, the processor 810 and the memory 820 communicate via the bus 830. When the machine-readable instructions are executed by the processor 810, they can perform the operations described above. Figure 1 The steps of the CO concentration source classification and identification method in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0178] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the CO concentration source classification and identification method in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0179] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0180] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0181] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0182] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0183] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0184] Finally, it should be noted that the above embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for classifying and identifying CO concentration sources, characterized in that, include: Raw CO concentration data from different sources were collected, and the target CO concentration data were cropped to construct a labeled CO concentration sample database. The target CO concentration data is preprocessed to obtain a standardized CO concentration sequence; Feature extraction is performed on the standardized CO concentration sequence to obtain curve features characterizing the concentration variation patterns of different sources, and a CO concentration sequence feature vector is constructed. The CO concentration sequence feature vector is subjected to feature selection and dimensionality reduction processing to obtain the optimized low-dimensional feature vector; The optimized low-dimensional feature vector is input into the classification and recognition model to determine the classification and recognition result of the CO concentration source.

2. The method according to claim 1, characterized in that, A labeled CO concentration sample database was constructed using the following method: The original CO concentration data is cropped to retain stable bands that contain the complete process of concentration change. The stable bands refer to the data segments from the start of concentration change to the point of recovery to stability, and the stable bands include the target CO concentration data. A CO source label is established for each stable band, and a sample database of CO concentrations caused by different reasons is constructed. The CO source label is used to identify the specific source of the target CO concentration data in the stable band.

3. The method according to claim 2, characterized in that, The standardized CO concentration sequence was obtained in the following way: Outliers in the target CO concentration data of the stable band are processed using preset criteria, which are used to identify and remove data points that deviate from the average data value. The missing values ​​in the target CO concentration data of the stable band are filled using an interpolation method, which is used to estimate the missing values ​​based on the time series trend. To address the periodic fluctuation interference present in the target CO concentration data of the stable band, the empirical mode decomposition reconstruction method is used to remove the periodic interference components in the target CO concentration data of the stable band, resulting in deperiodic CO concentration data. Normalization was performed on the processed deperiodic CO concentration data to convert it to a uniform numerical scale, resulting in a standardized CO concentration sequence.

4. The method according to claim 1, characterized in that, The CO concentration sequence feature vector is constructed using the following method: The standardized CO concentration sequence is input into a deep neural network module, which outputs curve features characterizing the standardized CO concentration sequence. The deep neural network module is used to characterize the relationship between the standardized CO concentration sequence and the curve features, which are features extracted from the standardized CO concentration sequence that can distinguish the morphology and pattern of different sources. The curve features are combined to form a CO concentration sequence feature vector.

5. The method according to claim 4, characterized in that, The optimized low-dimensional feature vector is obtained through the following method: The CO concentration sequence feature vector is subjected to a recursive feature elimination method for feature selection to remove unimportant features; Principal component analysis is used to reduce the dimensionality of the features after feature selection to obtain the optimized low-dimensional feature vector.

6. The method according to claim 5, characterized in that, The classification and identification results of CO concentration sources were determined using the following methods: The optimized low-dimensional feature vector is input into the classification and recognition model to obtain the corresponding source classification probability distribution, wherein the classification and recognition model is used to characterize the mapping relationship between the optimized low-dimensional feature vector and the CO source category; Based on the probability distribution of the source classification, the category is determined to obtain the classification and identification results of the CO concentration source.

7. The method according to claim 6, characterized in that, The classification and recognition model is trained in the following manner: Obtain a training sample set, which includes multiple training samples, each of which includes a sample optimized low-dimensional feature vector and a corresponding sample CO source label; The optimized low-dimensional feature vector of the sample is used as the input to the classification and recognition model to be trained, and the corresponding sample CO source label is used as the output of the classification and recognition model to be trained.

8. A CO concentration source classification and identification device, characterized in that, include: The CO concentration sample database construction module is used to collect raw CO concentration data from different sources, crop the target CO concentration data, and construct a labeled CO concentration sample database. The CO concentration sequence standardization module is used to preprocess the target CO concentration data to obtain a standardized CO concentration sequence. The CO concentration sequence feature vector component module is used to extract features from the standardized CO concentration sequence, obtain curve features that characterize the concentration variation patterns of different sources, and construct a CO concentration sequence feature vector. The optimal low-dimensional feature vector construction module is used to perform feature selection and dimensionality reduction on the CO concentration sequence feature vector to obtain the optimal low-dimensional feature vector. The classification and identification result determination module is used to input the optimized low-dimensional feature vector into the classification and identification model to determine the classification and identification result of the CO concentration source.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is in operation, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method as described in any one of claims 1 to 7.

Images