Working face mine pressure prediction method based on domain generalization small sample learning and model construction method thereof

Abstract

The invention provides a working face mine pressure prediction method based on domain generalization small sample learning and a model construction method thereof, and belongs to the technical field of signal processing. Constructing a small sample data set, and designing an efficient feature extraction network to perform feature extraction on the preprocessed mine pressure data; based on a feature extraction network, a generative adversarial network model for generators and discriminators of different domains is designed. Meanwhile, in domain generalization problem processing of source domain data and target domain data, an attention weighted mine pressure prediction network with a long and short-term memory network coding module, a guiding attention weighting module and a soft attention weighting module is constructed by utilizing the instructive coding function of a generator. The problem that the feature dependency relationship in the training data cannot be captured on the premise of small samples is further solved. According to the method, the prediction robustness can be improved, a new solution is provided for the complexity of mine pressure display of the coal mine working face, and the continuity and reliability of coal mine safety production are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to a working face mine pressure prediction method based on domain generalization small sample learning and a model construction method thereof. Background Art

[0002] With the continuous increase in the depth and intensity of coal mining, the environment faced by the working face of the mine is becoming more and more complex, and abnormal mine pressure phenomena occur frequently, such as coal wall spalling and support "crushing", which directly affect the safe production and equipment operation of coal mines. Abnormal mine pressure not only brings serious economic losses, but also threatens the life safety of workers. Roof accidents account for 40% to 50% of coal mine casualties. Therefore, accurate prediction of mine pressure on the working face has become an important means to ensure safe production in coal mines.

[0003] At present, the mainstream methods of mine pressure prediction can be divided into three categories: non-deterministic methods based on statistics and fuzzy mathematics, deterministic methods based on physical mechanics theory, and intelligent technical methods based on large amounts of monitoring data.

[0004] Non-deterministic methods based on statistics and fuzzy mathematics: This type of method uses statistical analysis and fuzzy mathematical modeling to predict changes in mine pressure using statistical laws or fuzzy relationships in historical data. For example, models such as linear regression and time series analysis are used to extract trends and periodic characteristics of data. The advantage of this type of method is that it is simple and easy to use, but its ability to capture complex nonlinear relationships is weak, especially when faced with mine pressure with obvious dynamic and nonlinear characteristics, it often shows large errors.

[0005] Deterministic methods based on physical mechanics theory: This type of method establishes a physical and mechanical model to predict mine pressure based on the geological conditions and roof structure of the mine, such as stress field inversion, vibration wave detection and other technologies. These methods can reveal the physical mechanism of mine pressure changes, and the theoretical explanation is relatively intuitive. However, due to the complex and changeable environment of the mine, the physical and mechanical model often needs to rely on a large number of assumptions and is difficult to cope with random factors under actual mining conditions. In addition, the physical and mechanical model usually has high requirements for the accuracy of parameters, and the parameters are difficult to obtain accurately in the actual environment, which limits its scope of application.

[0006] Smart technology-based approach: In recent years, with the development of big data and artificial intelligence technology, intelligent technologies based on machine learning and deep learning have been widely used in mine pressure prediction. For example, models such as support vector machine (SVM), artificial neural network (ANN), long short-term memory network (LSTM) can automatically learn the complex nonlinear relationship in mine pressure data, and have significantly improved the prediction accuracy compared with traditional methods.

[0007] However, such methods usually rely on a large amount of labeled data, and require rich historical data support when training models. In actual coal mining, it is often difficult to obtain enough high-quality labeled data, especially in new mining areas or under special mining conditions. In addition, these intelligent models usually lack domain generalization capabilities, that is, the prediction effect of a model trained in one mining area will be significantly reduced when it is applied to mining areas with different geological conditions or mining methods. Summary of the invention

[0008] In view of the above problems, the present invention proposes a working face mine pressure prediction method based on domain generalization small sample learning and a model construction method thereof. The core innovation of this method is that it can make full use of existing data under small sample conditions to effectively predict mine pressure, and it has cross-domain generalization capabilities and can adapt to data changes in different mining environments. Through the fusion of multi-bracket data and in-depth analysis of time series characteristics, the model can overcome the limitations of traditional methods when data is insufficient or domain differences are large while maintaining high prediction accuracy.

[0009] A first aspect of the present invention provides a method for constructing a working face mine pressure prediction model based on domain generalization small sample learning, comprising the following steps: S1, multiple mine pressure sensors are arranged at different coal mine working faces to collect mine pressure data from the source domain and the target domain, and the data are preprocessed in a consistent manner to construct a small sample data set containing source domain and target domain data; S2, constructing a working face mine pressure prediction model, the model comprising a feature extraction network and an attention weighted mine pressure prediction network; The feature extraction network includes a block unit, a feature encoding unit, a position encoding unit, a transformer network modeling unit, and a feature flattening and output unit; the block unit is used to perform block processing on the original mine pressure data, and divide the data into multiple continuous segments with overlaps to retain the correlation between adjacent segments; the feature encoding unit uses a multi-layer perceptron to extract key features of each segment and generate a feature vector; the position encoding unit encodes the position of each segment; the transformer network modeling unit is used to perform relational modeling on the features, effectively capture global features and local details, and enhance the accuracy and correlation of feature expression; the feature flattening and output unit is used to flatten the output result of the transformer network modeling unit, convert it into a feature vector of a specific length, and pass it to the subsequent fully connected layer; The attention-weighted mine pressure prediction network uses a long short-term memory network and a dual attention mechanism to perform efficient feature fusion and prediction. S3, based on the feature extraction network, constructs a domain generalization training strategy based on the generative adversarial network. Based on the small sample data set in S1, the feature extraction network is trained through adversarial learning between the generator and the discriminator to obtain the final working face mine pressure prediction model.

[0010] Preferably, the preprocessing in S1 includes denoising and standardization processing; The denoising process includes two processes: data filtering and data denoising. The data filtering process uses the arithmetic mean filtering method to reduce random errors. The input N working face mine pressure time series data are represented as source domain data. and target domain data , thereby obtaining the sampling value, minimizing the sum of the squares of the deviations between it and each mine pressure sampling value;

[0011]

[0012] Where N is the sample size; is the sampling value of the mine pressure at time n; is the arithmetic mean of the sampling values; the filtered mine pressure source domain data S1 and target domain data T1 are obtained; Data denoising is to use wavelet threshold denoising method to compress the signal and extract useful mine pressure data from the noise signal; assuming that the noisy signal is , then the formula is expressed as:

[0013] in f(k) is a useful signal; e(k) is the noise coefficient function; ϵ is the standard deviation of the noise coefficient; extract the useful mine pressure source domain data S2 and target domain data T2 from the noise signal; The standardization process is to normalize the acquired mine pressure source domain data S2 and target domain data T2.

[0014] Preferably, the position encoding unit creates a position index for each working surface position in the source domain and the target domain based on the block processing of the blocking unit, and then calculates the position encoding according to the feature vector length of each position of the feature encoding unit, and the encoding is implemented by sine and cosine functions, wherein each dimension corresponds to a combination of a sine and a cosine value; for each position in the source domain and the target domain, the position encoding is specifically calculated using the following formula:

[0015]

[0016] in is the position index, j is the current dimension index, dim is the dimension index; then these calculated sine and cosine values ​​are interleaved and filled into the position encoding vector; finally, the position encoding vector is added to the corresponding working face mine pressure feature vector.

[0017] Preferably, the transformer network modeling unit designs a multi-head self-attention layer and a feedforward layer; the input data comes from a feature encoding unit and a position encoding unit, and first passes through a multi-head self-attention layer, in which 8 heads are set, and the dimension of each head is 8; the multi-head self-attention captures the relationship between different features by dividing the input data into multiple heads and calculating the attention weight of each head respectively, and finally splices the outputs of each head together to enhance the model's ability to represent information; next, the output passes through a feedforward layer, which is composed of two fully connected layers, the output dimension of the first layer is 64, and the output dimension of the second layer is 20; the activation function uses ReLU, and finally, it is processed by layer normalization.

[0018] Preferably, the domain generalization training strategy based on the generative adversarial network in S3 constructs the generator G derived from the aforementioned feature extraction network, and combines the feature mapping module to achieve effective data conversion; the feature mapping module includes a feature enhancement layer and feature map layers The feature enhancement layer consists of two fully connected layers. The output dimension of the first layer is 128, and the output dimension of the second layer is 64. The ReLU activation function is used; the feature mapping layer uses a linear transformation to transform the feature vector of the source domain obtained by the feature extraction network And the feature vector of the target domain The features are mapped to the target space to maintain the similarity between the source domain and the target domain data. The expression formula is as follows:

[0019] The generator G shares the parameters of the source domain and the target domain. In the generator, the working face mine pressure data of the source domain and the target domain are input separately and processed by the shared feature mapping module, so that the generated features of the two are as close as possible. The goal of the generator is to minimize the distance between the generated features of the source domain and the target domain. The optimization objective function formula is as follows:

[0020] in, and Represent the generated features of the source domain and the target domain respectively; The constructed discriminator D is used to effectively distinguish the working face mine pressure data of the source domain and the target domain. The discriminator first receives the feature data output by the generator and transforms its shape into 8x16 by flattening. Next, the feature data is processed by two fully connected layers. The output dimension of the first layer is 64 and the output dimension of the second layer is 32. Both use the ReLU activation function to enhance the nonlinear representation ability of the feature. Subsequently, the processed feature data is passed to the output layer, which contains a single neuron and uses Sigmoid The activation function performs binary classification, and the output value represents the similarity between the source domain and the target domain features, which is mathematically expressed as the following formula:

[0021] in, D(X) is the output of the discriminator, i.e. the input of the generator G, σ for Sigmoid function, and are the weight and bias of the output layer respectively; the goal of the discriminator is to improve the discrimination ability of the model by maximizing the difference between the source domain features and the target domain features. Its optimization objective function is as follows:

[0022] Among them Table Expected Function During the training process, the generator and discriminator are optimized alternately to achieve an effective fusion of source domain and target domain features.

[0023] Preferably, the attention-weighted mine pressure prediction network includes a long short-term memory network encoding module, a guided co-attention weighting module and a soft attention weighting module; first, the long short-term memory network encoding module is responsible for time-dependent modeling of sequence data, capturing the long-term and short-term characteristics of mine pressure data to enhance the model's understanding of time-varying laws; then, the guided co-attention weighting module introduces guiding weights in the feature selection process, focusing attention on feature intervals with more predictive significance, thereby improving feature extraction efficiency; finally, the soft attention weighting module dynamically weights the extracted features by assigning different weight values.

[0024] Preferably, the long short-term memory network encoding module is set to a single layer, and the number of hidden units is 128 to balance the model complexity and computational efficiency; for the input sequence , the output of the LSTM layer obtains N hidden states ,Finally, the output dimension after the LSTM layer is 128, forming a new feature representation.

[0025] Preferably, the guided co-attention weighted module first receives the feature representation from the long short-term memory network and the shared features of the generator output , by calculating the attention weight, the shared features are guided to the weighted process of the feature value. The calculation formula of the attention weight is as follows:

[0026] in, and is the weight matrix obtained through learning. Through these weights, the features are weighted to obtain the weighted feature representation Such as the formula:

[0027] By guiding the attention weighting, the information of the source domain and the target domain is effectively fused.

[0028] Preferably, the soft attention weighting module takes as input the feature output processed by the guided co-attention weighting module. , these features will be passed to the decoding network for further processing; the module first calculates the attention weight of the features of each time step to other time steps, and the weight is calculated by the following formula:

[0029] in, is the learned weight matrix, Represents the time step i Attention weights for all input features; the features obtained by weighted summation are expressed as formula:

[0030] in, is the output feature after processing by the guided co-attention weighted module i Dimension C provides a more stable and accurate input for the decoding network R, ensuring that the model can still effectively learn the shared knowledge between the source domain and the target domain under small sample conditions.

[0031] The second aspect of the present invention provides a method for predicting working face mine pressure based on domain generalization small sample learning, which deploys the working face mine pressure prediction model constructed by the construction method described in the first aspect into the monitoring system of the mining area, and includes the following process: Obtain mine pressure data in real time through sensors; The mine pressure data is preprocessed by denoising and standardization; The preprocessed mine pressure data is input into the deployed working face mine pressure prediction model to output real-time mine pressure dynamic prediction.

[0032] The third aspect of the present invention provides a working face mineral pressure prediction device based on domain generalization small sample learning, the device includes at least one processor and at least one memory, the processor and the memory are coupled; the memory stores a computer execution program of the working face mineral pressure prediction model constructed by the construction method described in the first aspect; when the processor executes the computer execution program stored in the memory, the processor executes a working face mineral pressure prediction method based on domain generalization small sample learning.

[0033] The fourth aspect of the present invention provides a computer-readable storage medium, which stores a computer program or instruction of a working face mine pressure prediction model constructed by the construction method described in the first aspect. When the program or instruction is executed by a processor, the processor executes a working face mine pressure prediction method based on domain generalization small sample learning.

[0034] Compared with the prior art, the present invention has the following beneficial effects: 1. Improve the intelligence and automation of mine pressure prediction: The present invention significantly improves the intelligence and automation of mine pressure prediction in the working face by constructing a mine pressure prediction model based on a generative adversarial network. The designed feature extraction module can not only effectively extract the complex nonlinear features in the mine pressure data under small sample conditions, but also solves the problem of poor model adaptability in different mining environments by introducing cross-domain generalization capabilities. Compared with traditional methods that require a large amount of labeled data, the present invention can achieve high-precision predictions using only a small amount of data, greatly reducing the cost and difficulty of data collection, and significantly improving the intelligence and automation of the mine pressure monitoring system.

[0035] 2. Improve the robustness and cross-domain generalization ability of the model: The generative adversarial network model constructed by the present invention effectively solves the problem of differences in the distribution of mine pressure data between different domains through interactive training of the generator and the discriminator, and realizes efficient generalization of source domain and target domain data. Combined with the long short-term memory network and the attention weighted mechanism, the present invention can not only capture the long-range dependencies in the time series characteristics, but also optimize the feature extraction process when the data samples are limited, and improve the adaptability of the model in a changing mining environment. Compared with traditional physical and mechanical models and statistical methods, the present invention shows stronger robustness and generalization ability when facing the randomness and nonlinear characteristics of complex mine environments.

[0036] 3. Improve prediction accuracy and small sample learning ability: In traditional methods, mine pressure prediction relies on a large amount of monitoring data, especially when the data is insufficient, the prediction accuracy will drop significantly. In contrast, the present invention, by designing an attention weighted model and an efficient feature extraction module based on a small number of samples, can still accurately capture the key features of mine pressure changes under small sample conditions, avoiding the risk of the model falling into local optimality. At the same time, by utilizing the guidance and soft attention weighting mechanism, multi-scaffold and time series data are effectively integrated, so that the prediction accuracy of the present invention is significantly improved, which solves the limitations of traditional methods when data is insufficient or feature dependencies are not fully captured.

[0037] 4. Improve the real-time monitoring and decision-making support capabilities for safe production: The present invention has been successfully applied to the prediction of mine pressure on actual working faces through the optimized cross-domain small sample learning model. It can monitor the changes in mine pressure in real time and provide accurate prediction results, helping managers to quickly identify potential risks and formulate response strategies. Compared with the existing technology that requires a large amount of manpower intervention, the present invention realizes intelligent and automated operation of the entire process, effectively improves the level of coal mine safety management, and ensures the continuity of production and the efficiency of operations. At the same time, by reducing potential accidents, safety risks are significantly reduced, production efficiency is improved, and the sustainable development of mine safety production is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall logic block diagram of the working face mine pressure prediction method of the present invention.

[0039] Figure 2 This is a schematic diagram of the original feature block of the present invention.

[0040] Figure 3 Schematic diagram of the overall structure of the feature extraction network of the present invention.

[0041] Figure 4 A schematic diagram of generating an adversarial model for the present invention.

[0042] Figure 5 This is a schematic diagram of the attention-weighted mine pressure prediction network structure of the present invention.

[0043] Figure 6 Schematic diagram of the structure of the soft attention weighted module of the present invention.

[0044] Figure 7 This is a simplified diagram of the structure of the working face mine pressure prediction equipment. DETAILED DESCRIPTION

[0045] The invention will be further described below in conjunction with specific embodiments.

[0046] The present invention collects a small amount of working face mine pressure data, and after data preprocessing, designs an efficient feature extraction network based on mine pressure data block unit, feature encoding unit, position encoding unit, transformer network modeling unit and feature flattening and output unit to extract features from the preprocessed mine pressure data. Subsequently, based on the feature extraction network, a generative adversarial network model of a generator module encoding unit and a discriminator module distinguishing unit for different domains is constructed. At the same time, in the model's processing of the domain generalization problem of source domain data and target domain data, the generator's guided encoding function is used to construct an attention weighted network with a long short-term memory network encoding module, a guided attention weighting module and a soft attention weighting module, so as to further solve the problem that the feature dependency in the training data cannot be captured under the premise of a small sample, causing the algorithm to fall into a local optimum. Finally, the optimized model is successfully applied to the mine pressure prediction of the actual working face, which can monitor and predict the trend of mine pressure changes in real time, and help managers identify potential risks as early as possible and make corresponding response decisions. This real-time prediction mechanism not only improves the safety management level of coal mine production, but also effectively ensures the continuity of the production process, reduces potential accidents, and ensures the efficient operation of mining operations. Figure 1 shown.

[0047] Composition of the initial data set: Multiple mine pressure sensors are arranged at different coal mine working faces to collect mine pressure data from the source domain and the target domain. Consistent data preprocessing operations are performed on the two types of data, including denoising and standardization, and noise and outliers are removed to ensure data consistency and comparability. The source domain data is used for the initial training of the model, while the target domain data is used as a verification and test set in the cross-domain generalization task. Finally, a small sample data set containing source and target domain data is constructed to provide high-quality input data for subsequent generative adversarial networks and domain generalization models.

[0048] Feature extraction network construction: The feature extraction network is based on block units, feature encoding units, position encoding units, transformer network modeling units, and feature flattening and output units, which can capture the global trend and local detail changes of the data at the same time. First, the original mine pressure data is processed in blocks, and the data is divided into multiple continuous segments with overlaps to retain the correlation between adjacent segments. Then, each segment is feature encoded and its key feature vector is extracted. Subsequently, an advanced transformer network is used to model the relationship between these features, further enhancing the accuracy and correlation of feature expression, laying the foundation for subsequent weighted fusion and generator construction in generative adversarial networks. In complex working face environments, this feature extraction module can effectively extract global and detail features. Even when the amount of mine pressure data is small or the data changes dynamically, it can still maintain high-precision prediction results and enhance the robustness and stability of mine pressure prediction.

[0049] Construct a domain generalization generative adversarial network training strategy: Based on the feature extraction network, a domain generalization training strategy based on the generative adversarial network is constructed. First, the mine pressure data of the source domain and the target domain that have passed the feature extraction network are input into the feature map, and encoded with the network generator of the feature extraction network to extract the shared features and unique features of the two. Subsequently, these features are input into the discriminator module, and the discriminator is used to distinguish the difference in feature distribution between the source domain and the target domain, and this difference is gradually narrowed through adversarial training to achieve effective generalization of cross-domain features. At the same time, the long short-term memory network is combined with the guided attention mechanism to further enhance the model's feature capture ability under small sample conditions, effectively solve the problem of reduced prediction accuracy of traditional models between different domain data, and ultimately improve the cross-domain generalization ability of mine pressure prediction.

[0050] Constructing an attention-weighted mine pressure prediction network: In order to improve the prediction accuracy of mine pressure under small sample conditions, an attention-weighted mine pressure prediction network including a long short-term memory network encoding module, a guided co-attention weighting module and a soft attention weighting module was constructed. First, the long short-term memory network encoding module is responsible for modeling the time dependency of sequence data and capturing the long-term and short-term characteristics of mine pressure data to enhance the model's understanding of the temporal variation law. Then, the guided co-attention weighting module introduces guiding weights in the feature selection process to focus attention on feature intervals with more predictive significance and improve the efficiency of feature extraction. Finally, the soft attention weighting module dynamically weights the extracted features by assigning different weight values ​​to ensure that the model can still achieve efficient feature fusion and prediction when facing sparse data or incomplete features. The entire model effectively solves the problem of capturing feature dependencies in small sample data by jointly using long short-term memory networks and dual attention mechanisms, ensuring the high accuracy and robustness of mine pressure prediction, and is suitable for complex and changeable working face environments.

[0051] Model deployment and use: In the actual working face environment, the mine pressure prediction model designed by the present invention obtains mine pressure data in real time through sensors installed on the working face monitoring equipment, and inputs these data into the optimized mine pressure prediction model. The model can quickly analyze the current mine pressure change trend and accurately identify potential risk pressure fluctuations.

[0052] 1. Dataset Construction and Preprocessing 1. Collection of original mine pressure data: As a data-driven technology, the present invention needs to obtain mine pressure data of different mining faces during actual mining for subsequent model training. Considering that mine pressure data is usually difficult to collect and continuous collection is challenging, sensors and monitoring equipment are deployed on the working face and its surrounding areas to collect mine pressure data in a shorter period of time. These data include mine pressure readings in different areas, which are divided into source domain mine pressure data S and target domain mine pressure data T.

[0053] 2. Data filtering and processing: Under the Internet of Things architecture of coal mines, due to the fact that the underground working face sensors are subject to various interferences, there are large noise and random errors in the measured mine pressure data, which brings challenges to the prediction of the support mine pressure. The present invention constructs a data filtering and processing module, uses the arithmetic mean filtering method to reduce random errors, and combines the wavelet denoising technology to compress the signal to achieve effective elimination of noise in the signal. This method is mainly used for smoothing the sampled values ​​of periodic pulsating parameters such as pressure and flow, and is the simplest and most commonly used digital filtering method. Specifically, the arithmetic mean filtering method is to sample the signal multiple times continuously at a certain moment, and perform arithmetic averaging on the sample data to obtain the signal value at that moment. According to the arithmetic mean filtering algorithm, the input N working face mine pressure time series data is expressed as and , thereby obtaining the sampled value and , so that the sum of the squares of the deviations between it and each mine pressure sampling value is minimized, the formula is as follows:

[0054]

[0055] Where N is the sample size; is the sampling value of the mine pressure at time n; is the arithmetic mean of the sampled values. As can be seen from the formula, arithmetic mean filtering is essentially to add and average N sampled values ​​as the value at time n. Through this process, the filtered source domain data S1 and target domain data T1 can be obtained.

[0056] 3. Data denoising: After filtering the time series data of the support pressure of the working face, the present invention further uses the wavelet threshold denoising method to compress the signal and extract the useful support pressure data from the noise signal. Wavelet denoising is a technology that eliminates noise through shortwaves. It can extract the detailed features of the signal and integrate it with low-pass filtering, thereby effectively removing noise interference. Suppose the noisy signal is , it is expressed as the following formula:

[0057] in f(k) is a useful signal; e(k) is the noise coefficient function; ϵ is the standard deviation of the noise coefficient. According to the above formula, useful support mine pressure data S2 and T2 can be extracted from the noise signal.

[0058] 4. Data standardization: In order to further improve the generalization effect after model training and better perform domain alignment in the process of domain generalization, the present invention performs standardization operations on the working face mine pressure data S2 and T2 after filtering and denoising. As shown in the following formula:

[0059] in, It indicates the pressure reading collected at time point n in the working face mine pressure data after standardized operation. Respectively represent the average value of all mine pressure readings and the calculated standard deviation. Through the above calculation, the mine pressure data of the source domain working face that can be used as model input can be obtained. And the target area working face mine pressure data .

[0060] 2. Feature Extraction Network Construction The feature extraction network is based on the block unit, feature encoding unit, position encoding unit, transformer network modeling unit and feature flattening and output unit, which can capture the global trend and local detail changes of the data at the same time. First, the original data of the source domain and the target domain are processed in blocks, and the data is divided into 8 overlapping continuous segments, which are composed of two hidden layers to retain the correlation between adjacent segments. Then, the multi-layer perceptron in the feature encoding unit is used to extract the key features of each segment and generate a feature vector. Subsequently, the position of each segment is encoded, and the TSM network is used to model the relationship between these features. The self-attention layer and the feedforward layer of the multi-head self-attention mechanism are mainly used to further enhance the accuracy and correlation of the feature expression, laying the foundation for the subsequent weighted fusion and generator construction in the generative adversarial network. In the complex working face environment, the feature extraction model can effectively extract global and detail features. Even when the amount of mine pressure data is small or the dynamic changes are large, it can still maintain high-precision prediction results and enhance the robustness and stability of mine pressure prediction.

[0061] 1. Block unit design The feature extraction network constructed by the present invention first extracts the source domain working face mine pressure data And the target area working face mine pressure data The source domain and the target domain are divided into blocks to support the self-attention learning of the TSM unit. Figure 2 The split operation shown.

[0062] For the source domain working face mine pressure data And the target area working face mine pressure data , divided into 8 segments according to the total time segment length, with an overlapping segment ratio of 10%, thus obtaining the source domain block segment set , and the target domain fragment set .

[0063] 2. Feature encoding unit design In the present invention, a feature encoding unit is constructed as a feature extraction model for the data fragments after segmentation. This unit structure consists of a 64-layer fully connected layer, a 32-layer fully connected layer and a 16-layer fully connected layer.

[0064] 3. Position encoding unit design Based on the designed block unit, a position index is created for each working surface position in the source domain and the target domain, ranging from 0 to 7. Then, the position encoding is calculated based on the length of the feature vector for each position based on the feature encoding unit, that is, 16. This encoding is implemented by sine and cosine functions, where each dimension corresponds to a combination of a sine and a cosine value. Specifically, for each position in the source domain and the target domain, the position encoding is calculated using the following formula:

[0065]

[0066] Where pos is the position index, j is the current dimension index, and dim is the dimension index. These calculated sine and cosine values ​​are then interleaved into the position encoding vector. Finally, the position encoding vector is added to the corresponding working face mine pressure feature vector to ensure that the model can take into account the temporal position information of the mine pressure data when processing the data. In this way, the model can also better adapt to and understand the changes in the mine pressure time period of different working faces.

[0067] 4. Transformer Network Modeling (TSM) Unit The core unit in the feature extraction model constructed by the present invention is the transformer network modeling (TSM) unit. In TSM, a multi-head self-attention layer and a feedforward layer are designed. The input data comes from the feature encoding unit and the position encoding unit, and first passes through a multi-head self-attention layer, in which 8 heads are set, and the dimension of each head is 8. Multi-head self-attention captures the relationship between different features by dividing the input data into multiple heads and calculating the attention weight of each head respectively. Finally, the outputs of each head are spliced ​​together to enhance the model's ability to represent information. Next, the output passes through a feedforward layer, which consists of two fully connected layers, the output dimension of the first layer is 64, and the output dimension of the second layer is 20. The activation function uses ReLU to improve the nonlinear expression ability of the model. Finally, after layer normalization, the stability of data distribution and the convergence of the model are ensured. Through this structure, the TSM unit can effectively capture global features and local details, and improve the accuracy and robustness of feature extraction.

[0068] 5. Feature Flattening and Output Unit The feature flattening and output unit constructed in the present invention is mainly used to process the output from the TSM unit. First, the output result of the TSM unit is flattened and converted into a length of The flattening operation ensures that the feature dimension is adapted and can be effectively passed to the subsequent fully connected layer, which ends with a 128-node fully connected layer. The feature vector of the source domain is obtained. And the feature vector of the target domain The output unit further integrates the global features and local detail features extracted by the TSM unit to ensure that the final output can fully reflect the mining pressure data characteristics of the source domain and the target domain, and improve the prediction accuracy and stability of the model. Multiple units are connected in sequence to form this feature extraction model, and its overall structure is as follows: Figure 3 shown.

[0069] 3. Design of training strategy for generative adversarial network model Generator Module Encoding Module: The generator G constructed by the present invention is derived from the aforementioned feature extraction model and combined with the feature mapping module to achieve effective data conversion. The feature mapping module consists of two main parts: feature enhancement layer and feature map layers The feature enhancement layer consists of two fully connected layers. The output dimension of the first layer is 128, and the output dimension of the second layer is 64. The ReLU activation function is used to enhance the expressiveness of the features. The feature mapping layer uses a linear transformation to transform the feature vector of the source domain obtained by the S2 feature extraction model. And the feature vector of the target domain The features are mapped to the target space to maintain the similarity between the source domain and the target domain data. It is expressed as the following formula:

[0070] The generator G shares the parameters of the source domain and the target domain. In the generator, the working face mine pressure data of the source domain and the target domain are input separately and processed by the shared feature mapping module to make the generated features of the two as close as possible. This design not only enhances the comparability of the generated features, but also provides a greater recognition challenge for the subsequent discriminator. The goal of the generator is to minimize the distance between the generated features of the source domain and the target domain, and optimize the objective function as follows:

[0071] in, and Represent the generated features of the source domain and the target domain respectively. This process ensures the robustness and adaptability of the model when processing data from different domains.

[0072] Discriminator module: The discriminator module constructed by the present invention is used to effectively distinguish the working face mine pressure data in the source domain and the target domain. The discriminator module D first receives the feature data output by the generator and transforms its shape into 8x16 by flattening operation for subsequent processing. Next, the feature data is processed by two fully connected layers. The output dimension of the first layer is 64 and the output dimension of the second layer is 32. Both use the ReLU activation function to enhance the nonlinear representation ability of the feature. Subsequently, the processed feature data is passed to the output layer, which contains a single neuron and uses Sigmoid The activation function performs binary classification. The output value represents the similarity between the source domain and the target domain features, which is mathematically expressed as the following formula:

[0073] Among them, D(X) is the output of the discriminator, that is, the input of the generator G, σ for Sigmoid function, and are the weights and biases of the output layer respectively. The goal of the discriminator is to improve the model’s discrimination ability by maximizing the difference between the source domain features and the target domain features. Its optimization objective function is as follows:

[0074] Among them In this process, the discriminator module ensures that the source domain and target domain data can be effectively distinguished to improve the robustness of the model. Through such a design, the discriminator module can enhance the model's ability to judge feature similarity when facing complex mine pressure data, and improve the accuracy and reliability of mine pressure prediction at the working face.

[0075] Domain generalization adversarial training strategy: The domain generalization adversarial training strategy of the present invention aims to improve the adaptability of the model between different domains through adversarial learning between the generator and the discriminator. During the training process, the generator and the discriminator are optimized alternately to achieve an effective fusion of the source domain and target domain features. The goal of the generator is to generate target domain features that are similar to the source domain features so that they can "fool" the discriminator, while the goal of the discriminator is to accurately distinguish between source domain features and target domain features. Specifically, the loss function of the generator is defined as the following formula:

[0076] Through such adversarial training strategies, the generator and the discriminator are optimized in a continuous game, so that the generated target domain features are close to the source domain features in distribution, thereby enhancing the generalization ability of the model and improving the prediction accuracy under different mining pressure data conditions. Figure 4 shown.

[0077] 4. Attention-weighted mine pressure prediction network design The attention-weighted mine pressure prediction network proposed in the present invention is mainly composed of a long short-term memory network (LSTM) encoding module, a guided attention weighting module and a soft attention weighting module. Different from traditional methods, this model uses a domain generalization small sample learning strategy to fully extract the temporal features and shared features of the mine pressure data under the data conditions of the source domain and the target domain. In the feature extraction network, the multi-head self-attention mechanism and the feedforward layer are combined to effectively capture the global and local detail features. Subsequently, the LSTM encoding module further extracts temporal features to improve the model's adaptability to dynamic changes in mine pressure data. Through the guided co-attention mechanism, the shared knowledge of the source domain and the target domain is fully learned, which enhances the generalization ability of the model. Under small sample conditions, the soft attention weighting mechanism effectively solves the problem of information loss in feature learning, so that the model can still maintain high prediction accuracy and robustness when facing limited data. Finally, through the domain generalization training and testing strategy, the model can make accurate predictions on unlabeled target domain data, reflecting excellent domain adaptation and small sample learning capabilities. Its overall structure is as follows Figure 5 .

[0078] Long short-term memory network (LSTM) encoding module: The long short-term memory network encoding module constructed by the present invention is designed to capture the temporal characteristics of the working face mine pressure data to improve the adaptability of the prediction model to dynamic changes in data. The input of this module is the processed mine pressure data, which is subjected to sequence modeling through a long short-term memory (LSTM) layer. The LSTM layer can effectively capture long-term dependencies and adjust the preservation and forgetting of information through a gating mechanism, thereby extracting the temporal characteristics in the data. In terms of specific design, the LSTM is set to a single layer with 128 hidden units to balance the model complexity and computational efficiency. For the input sequence , N hidden states can be obtained through the output of the LSTM layer ,Finally, the output dimension after the LSTM layer is 128, forming a new feature representation.,Through this design, the long short-term memory network encoding module can effectively,extract the time features in the mine pressure data, enhance the model's ability to,memorize historical information, and provide a feature basis for the subsequent,attention weighting module.

[0079] Guided co-attention weighted module: The guided attention weighted module in this invention aims to strengthen the feature sharing between the source domain and the target domain through the co-attention mechanism. In this module, the domain shared feature information generated by the generator in the previous chapter is used to guide the weighting of the N feature values ​​generated by the long short-term memory network to achieve effective learning of the working face mine pressure data.

[0080] Specifically, the module first receives the feature representation from the long short-term memory network and the shared features of the generator output By calculating the attention weight, the shared features are guided into the weighted process of the feature value. The calculation of the attention weight can be expressed as the following formula:

[0081] in, and is the learned weight matrix. Through these weights, the features are weighted to obtain the weighted feature representation Such as the following formula:

[0082] Through this guided attention weighting, the module can effectively fuse the information of the source domain and the target domain, promote the learning of shared knowledge, and thus improve the accuracy and robustness of the model in the prediction of mine pressure data. The introduction of this mechanism not only enhances the model's ability to focus on features, but also ensures the effective fusion of features from different domains, providing a solid foundation for subsequent feature fusion and prediction.

[0083] Soft-Attention Weighted Module: The soft-attention weighted module in this invention aims to effectively solve the problem of dynamic information loss in small sample learning through the soft-attention weighted mechanism. In the case of small samples, the amount of mine pressure data is limited, which makes the model susceptible to noise and uncertainty during feature learning and prediction. In this module, the input is the feature output processed by the guided co-attention weighted module. , these features will be passed to the decoding network for further processing. The module first calculates the attention weights of the features of each time step to other time steps. These weights are calculated by the following formula:

[0084] in, is the learned weight matrix, Represents the time step i The attention weights for all input features. This weighting mechanism can help the model better capture key features when samples are scarce, thereby reducing interference with unimportant features and improving the robustness of the model. The features obtained by weighted summation are expressed as the following formula:

[0085] C provides a more stable and accurate input for the decoding network R, ensuring that the model can still effectively learn the shared knowledge between the source domain and the target domain under the condition of small samples. This structural design enables the model to better utilize limited data to accurately predict mine pressure when facing the challenge of small sample learning, improving its performance and adaptability. Its overall structure is as follows: Figure 6 .

[0086] Domain generalization training and testing strategy: The training strategy of the present invention aims to perform preliminary training of the model through labeled data sets in the source domain to ensure that the model can learn basic mine pressure features. In the generative adversarial module, unlabeled data in the target domain is used for feature adversarial training to achieve alignment of source and target domain features. This method enhances the generalization ability of the model through adversarial training, enabling it to better adapt to feature differences between different domains, especially under small sample conditions.

[0087] In the testing phase, the model directly predicts on the data in the target domain. Through this strategy, the method can effectively cope with the challenge of small sample learning and ensure accurate prediction of working face pressure in the absence of rich data. This process fully demonstrates the advantages of small sample learning methods based on domain generalization in complex environments and improves the reliability and practicality of the model in practical applications.

[0088] 5. Model deployment and application Based on the designed working face mine pressure prediction model and combined with the described domain generalization small sample learning training strategy, the optimized mine pressure prediction model is finally obtained. This optimized model can efficiently predict the changes in working face mine pressure under different mining environments. The optimized model is deployed in the monitoring system of the mining area to help the mining area realize intelligent mine pressure monitoring and early warning. The specific implementation steps include: Deployment of mine pressure data collection equipment: Install high-precision mine pressure sensors on the working faces of the mining area to collect mine pressure data of the source and target working faces in real time to ensure the accuracy and completeness of the data and provide high-quality data support for subsequent model training and prediction.

[0089] Deployment of the mine pressure prediction model: The finally optimized working face mine pressure prediction model is deployed to the monitoring system of the mining area, and the pre-processed mine pressure data is input for real-time prediction to obtain the dynamic changes of the mine pressure of the current mining area working face.

[0090] Model verification and prediction effect evaluation: After the model is deployed, actual mine pressure data is collected regularly to verify and evaluate the model prediction effect to ensure that the model's prediction results are consistent with the actual situation, and the model is adjusted and optimized when necessary.

[0091] Intelligent early warning and decision-making: Based on the mine pressure change trend predicted by the model and the actual situation of the mining area, the system will automatically issue early warning prompts to help mine managers adjust mining strategies in a timely manner, prevent potential mine pressure imbalances, and ensure the safety and efficiency of mining operations.

[0092] like Figure 7As shown, the present invention also provides a working face mine pressure prediction device based on domain generalization small sample learning, the device includes at least one processor and at least one memory, and also includes a communication interface and an internal bus; the memory stores a computer execution program; the memory stores a computer execution program of the working face mine pressure prediction model constructed by the construction method as described above; when the processor executes the computer execution program stored in the memory, the processor can execute a working face mine pressure prediction method based on domain generalization small sample learning. The internal bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus. The memory may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a U disk, a mobile hard disk, a read-only memory, a disk or an optical disk.

[0093] The device may be provided as a terminal, a server or other forms of device. In an exemplary embodiment, the electronic device may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0094] The present invention also provides a computer-readable storage medium, which stores a computer execution program of the working face mine pressure prediction model constructed by the construction method described above. When the computer execution program is executed by a processor, the processor can execute a working face mine pressure prediction method based on domain generalization small sample learning.

[0095] Specifically, a system, device or equipment equipped with a readable storage medium may be provided, on which a software program code for implementing the functions of any of the above-mentioned embodiments is stored, and a computer or processor of the system, device or equipment reads and executes the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the above-mentioned embodiments, so the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0096] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0097] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for constructing a working face mine pressure prediction model based on domain generalization small sample learning, characterized in that: The following steps are involved: S1, multiple mine pressure sensors are arranged at different coal mine working faces to collect mine pressure data from the source domain and the target domain, and the data are preprocessed in a consistent manner to construct a small sample data set containing source domain and target domain data; S2, constructing a working face mine pressure prediction model, the model comprising a feature extraction network and an attention weighted mine pressure prediction network; The feature extraction network includes a block unit, a feature encoding unit, a position encoding unit, a transformer network modeling unit, and a feature flattening and output unit; the block unit is used to perform block processing on the original mine pressure data, and divide the data into multiple continuous segments with overlaps to retain the correlation between adjacent segments; the feature encoding unit uses a multi-layer perceptron to extract key features of each segment and generate a feature vector; the position encoding unit encodes the position of each segment; the transformer network modeling unit is used to perform relational modeling on the features, effectively capture global features and local details, and enhance the accuracy and correlation of feature expression; the feature flattening and output unit is used to flatten the output result of the transformer network modeling unit, convert it into a feature vector of a specific length, and pass it to the subsequent fully connected layer; The attention-weighted mine pressure prediction network uses a long short-term memory network and a dual attention mechanism to perform efficient feature fusion and prediction. S3, based on the feature extraction network, constructs a domain generalization training strategy based on the generative adversarial network. Based on the small sample data set in S1, the feature extraction network is trained through adversarial learning between the generator and the discriminator to obtain the final working face mine pressure prediction model.

2. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 1, characterized in that: The preprocessing in S1 includes denoising and standardization; The denoising process includes two processes: data filtering and data denoising. The data filtering process uses the arithmetic mean filtering method to reduce random errors. The input N working face mine pressure time series data are represented as source domain data. and target domain data , thereby obtaining the sampling value, minimizing the sum of the squares of the deviations between it and each mine pressure sampling value; Where N is the sample size; is the sampling value of the mine pressure at time n; is the arithmetic mean of the sampling values; the filtered mine pressure source domain data S1 and target domain data T1 are obtained; Data denoising is to use wavelet threshold denoising method to compress the signal and extract useful mine pressure data from the noise signal; assuming that the noisy signal is , then the formula is expressed as: in f(k) is a useful signal; e(k) is the noise coefficient function; ϵ is the standard deviation of the noise coefficient; extract the useful mine pressure source domain data S2 and target domain data T2 from the noise signal; The standardization process is to normalize the acquired mine pressure source domain data S2 and target domain data T2.

3. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 1, characterized in that: The position encoding unit creates a position index for each working surface position of the source domain and the target domain based on the block processing of the block unit, and then calculates the position encoding according to the feature vector length of each position of the feature encoding unit, and the encoding is implemented by sine and cosine functions, wherein each dimension corresponds to a combination of a sine and a cosine value; for each position of the source domain and the target domain, the position encoding is specifically calculated using the following formula: in is the position index, j is the current dimension index, dim is the dimension index; then these calculated sine and cosine values ​​are interlaced to fill the position encoding vector; finally, the position encoding vector is added to the corresponding working face mine pressure feature vector.

4. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 1, characterized in that: The transformer network modeling unit designs a multi-head self-attention layer and a feed-forward layer; the input data comes from the feature encoding unit and the position encoding unit, and first passes through a multi-head self-attention layer, in which 8 heads are set, and the dimension of each head is 8; the multi-head self-attention captures the relationship between different features by dividing the input data into multiple heads and calculating the attention weight of each head respectively, and finally splices the outputs of each head together to enhance the model's ability to represent information; Next, the output passes through a feed-forward layer, which consists of two fully connected layers. The output dimension of the first layer is 64, and the output dimension of the second layer is 20; the activation function uses ReLU, and finally, it is normalized by the layer.

5. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 1, characterized in that: In the above S3, the domain generalization training strategy based on the generative adversarial network is used to construct the generator G derived from the aforementioned feature extraction network, and combined with the feature mapping module to achieve effective data conversion; The feature mapping module includes a feature enhancement layer and feature map layers The feature enhancement layer consists of two fully connected layers. The output dimension of the first layer is 128, and the output dimension of the second layer is 64. The ReLU activation function is used; the feature mapping layer uses a linear transformation to transform the feature vector of the source domain obtained by the feature extraction network And the feature vector of the target domain The features are mapped to the target space to maintain the similarity between the source domain and the target domain data. The expression formula is as follows: The generator G shares the parameters of the source domain and the target domain. In the generator, the working face mine pressure data of the source domain and the target domain are input separately and processed by the shared feature mapping module, so that the generated features of the two are as close as possible. The goal of the generator is to minimize the distance between the generated features of the source domain and the target domain. The optimization objective function formula is as follows: in, and Represent the generated features of the source domain and the target domain respectively; The constructed discriminator D is used to effectively distinguish the working face mine pressure data of the source domain and the target domain. The discriminator first receives the feature data output by the generator and transforms its shape into 8x16 by flattening. Next, the feature data is processed by two fully connected layers. The output dimension of the first layer is 64 and the output dimension of the second layer is 32. Both use the ReLU activation function to enhance the nonlinear representation ability of the feature. Subsequently, the processed feature data is passed to the output layer, which contains a single neuron and uses Sigmoid The activation function performs binary classification, and the output value represents the similarity between the source domain and the target domain features, which is mathematically expressed as the following formula: in, D(X) is the output of the discriminator, i.e. the input of the generator G, σ for Sigmoid function, and are the weight and bias of the output layer respectively; the goal of the discriminator is to improve the discrimination ability of the model by maximizing the difference between the source domain features and the target domain features. Its optimization objective function is as follows: Among them Table Expected Function During the training process, the generator and discriminator are optimized alternately to achieve an effective fusion of source domain and target domain features.

6. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 1, characterized in that: The attention-weighted mine pressure prediction network includes a long-short-term memory network encoding module, a guided co-attention weighting module and a soft attention weighting module; first, the long-short-term memory network encoding module is responsible for modeling the time dependency of sequence data, capturing the long-term and short-term characteristics of mine pressure data to enhance the model's understanding of the law of time change; then, the guided co-attention weighting module introduces guiding weights in the feature selection process, focusing attention on feature intervals with more predictive significance, thereby improving feature extraction efficiency; Finally, the soft attention weighting module dynamically weights the extracted features by assigning different weight values.

7. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 6, characterized in that: The LSTM network encoding module is set to a single layer with 128 hidden units to balance model complexity and computational efficiency. , the output of the LSTM layer obtains N hidden states ,Finally, the output dimension after the LSTM layer is 128, forming a new feature representation.

8. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 6, characterized in that: The guided co-attention weighted module first receives the feature representation from the long short-term memory network and the shared features of the generator output , by calculating the attention weight, the shared features are guided to the weighted process of the feature value. The calculation formula of the attention weight is as follows: in, and is the weight matrix obtained through learning. Through these weights, the features are weighted to obtain the weighted feature representation Such as the formula: By guiding the attention weighting, the information of the source domain and the target domain is effectively fused.

9. The method for constructing a working face mine pressure prediction model based on domain generalization small sample learning according to claim 6, characterized in that: The soft attention weighting module takes as input the feature output processed by the guided co-attention weighting module. , these features will be passed to the decoding network for further processing; the module first calculates the attention weight of the features of each time step to other time steps, and the weight is calculated by the following formula: in, is the learned weight matrix, Represents the time step i Attention weights for all input features; the features obtained by weighted summation are expressed as formula: in, is the output feature after processing by the guided co-attention weighted module i Dimension C provides a more stable and accurate input for the decoding network R, ensuring that the model can still effectively learn the shared knowledge between the source domain and the target domain under small sample conditions.

10. A method for predicting working face mine pressure based on domain generalization small sample learning, characterized by: The working face mine pressure prediction model constructed by the construction method according to any one of claims 1 to 9 is deployed in the monitoring system of the mining area, and includes the following processes: Obtain mine pressure data in real time through sensors; The mine pressure data is preprocessed by denoising and standardization; The preprocessed mine pressure data is input into the deployed working face mine pressure prediction model to output real-time mine pressure dynamic prediction.

Images