A mine pressure prediction method based on size model mixing

Abstract

The application discloses a mine pressure prediction method based on a large-small model hybrid, which comprises the following steps: obtaining original data collected by at least one sensor underground, and processing the original data to obtain multi-modal spatio-temporal features; determining a target scene type based on the multi-modal spatio-temporal features through a large language model; determining a corresponding scheduling model set from a small model library based on the target scene type and the multi-modal spatio-temporal features; and obtaining a target mine pressure prediction result based on the multi-modal spatio-temporal features and the target scene type through the scheduling model set and the large language model. The application improves the adaptability and accuracy of the mine pressure prediction result, enhances the explainability and engineering usability, provides a clear reference basis for underground dispatchers, and improves the trustworthiness and decision efficiency of the mine pressure prediction result.

Description

Technical Field

[0001] This invention relates to the field of mine pressure prediction technology, and in particular to a mine pressure prediction method, device and storage medium based on a hybrid large and small model. Background Technology

[0002] With the deepening of intelligent mine construction, mine pressure prediction technology, as a key means to ensure mining safety, relies on intelligent analysis methods of sensor data such as support pressure and microseismic monitoring. Among them, the underground mining environment is highly dynamic and complex. The interaction of factors such as working face advance speed, geological structure changes, and equipment operating status makes mine pressure data exhibit significant nonlinear, temporal fluctuations, and spatial heterogeneity characteristics.

[0003] In related technologies, mine pressure prediction models (such as LSTM) struggle to adapt to varying data distributions under different working faces, geological conditions, or sudden events (such as fault crossings or roof pressure). They are unable to proactively detect environmental changes, and errors tend to accumulate over long periods, leading to a decrease in the accuracy of mine pressure prediction results. Furthermore, the mine pressure prediction results obtained from these technologies lack interpretability, making it difficult to gain the trust of personnel and hindering the practical application of these predictions in safety decision-making. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] To address this, the present invention proposes a mine pressure prediction method based on a hybrid large and small model. This method can perform semantic parsing and identify the current target scenario type using a large language model, and then call the target small model from the scheduling model set that matches the target scenario type to predict mine pressure. This improves the adaptability and accuracy of the mine pressure prediction results, enhances interpretability and engineering usability, provides clear reference for underground dispatchers, and improves the trustworthiness and decision-making efficiency of the mine pressure prediction results.

[0006] Another objective of this invention is to provide a mine pressure prediction device based on a hybrid size model.

[0007] To achieve the above objectives, this invention proposes a method for predicting mine pressure based on a hybrid large and small model, the method comprising: Acquire raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features; Based on the aforementioned multimodal spatiotemporal features, the target scene type is determined using a large language model; Based on the target scene type and the multimodal spatiotemporal characteristics, a corresponding set of scheduling models is determined from the small model library; Based on the multimodal spatiotemporal features and the target scene type, the target mining pressure prediction result is obtained through the scheduling model set and the large language model.

[0008] The mineral pressure prediction method based on a hybrid size model in this invention may also have the following additional technical features: In one embodiment of the present invention, the processing of the original data to obtain multimodal spatiotemporal features includes: The original data is resampled using a sliding window mechanism to obtain resampled data. The resampled data is aligned to obtain aligned data; The textual information in the alignment data is embedded into a first feature vector through an encoder, and the numerical information in the alignment data is determined as a second feature vector; The first feature vector and the second feature vector are concatenated to obtain multimodal spatiotemporal features.

[0009] In one embodiment of the present invention, determining the target scene type based on the multimodal spatiotemporal features using a large language model includes: Determine the semantic cue words corresponding to the multimodal spatiotemporal features; The multimodal spatiotemporal features and the semantic prompt words are input into a large language model to obtain the first probability distribution corresponding to the scene type; Based on the first probability distribution, the target scene type is determined.

[0010] In one embodiment of the present invention, determining the corresponding scheduling model set from the small model library based on the target scene type and the multimodal spatiotemporal features includes: The target scene type and the multimodal spatiotemporal features are input into the target model selector to obtain the second probability distribution of the small models in the small model library for scheduling. The labeling order of the target small models used for scheduling is determined based on the second probability distribution; Obtain the weight coefficients corresponding to the target small model, and determine the scheduling model set by the label order of the target small model and the weight coefficients of the target small model.

[0011] In one embodiment of the present invention, obtaining the target mineral pressure prediction result based on the multimodal spatiotemporal features and the target scene type through the scheduling model set and the large language model includes: The target sub-models in the scheduling model set are scheduled sequentially, and the first mineral pressure prediction result of each target sub-model is obtained based on the multimodal spatiotemporal features; The second mineral pressure prediction result is obtained by weighting and summing the weight coefficients of each target small model with the first mineral pressure prediction result. Based on the target scenario type, determine the corresponding set of historical scheduling models and input semantics; Based on the historical scheduling model set and the input semantics, the natural language explanation is generated through the large language model; The explained content and the second mine pressure prediction result are determined as the target mine pressure prediction result.

[0012] In one embodiment of the present invention, the method further includes: Obtain the actual collected value at the time corresponding to the target mine pressure prediction result; Based on the target mine pressure prediction results and the actual collected values, the prediction error is determined; Based on the prediction error, determine whether to adjust the large model and the target small model in the scheduling model set; If it is determined that adjustments need to be made to the large model and the target small model in the scheduling model set, then the adjustments are made to the large model and the target small model in the scheduling model set based on the target mining pressure prediction results and the actual collected values.

[0013] In one embodiment of the present invention, determining whether to adjust the large model and the target small model in the scheduling model set based on the prediction error includes: If the prediction error exceeds a preset threshold, then it is determined that adjustments will be made to the large model and the target small model in the scheduling model set.

[0014] In one embodiment of the present invention, adjusting the large model and the target small model in the scheduling model set based on the target mineral pressure prediction result and the actual collected value includes: Based on the target mineral pressure prediction results and the actual collected values, the loss value is determined through a loss function; Based on the loss value, the network parameters in the target small model are adjusted to obtain the adjusted small model; The large model is fine-tuned based on the target mine pressure prediction results and the actual collected values ​​to obtain the fine-tuned large model.

[0015] Another aspect of the present invention proposes a mining pressure prediction device based on a hybrid size model, the device comprising: The processing module is used to acquire raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features. The first determining module is used to determine the target scene type based on the multimodal spatiotemporal features using a large language model; The second determining module is used to determine the corresponding set of scheduling models from the small model library based on the target scene type and the multimodal spatiotemporal features; The prediction module is used to obtain the target mining pressure prediction result based on the multimodal spatiotemporal features and the target scene type, through the scheduling model set and the large language model.

[0016] This invention discloses a method and apparatus for mine pressure prediction based on a hybrid big-small model approach. The method includes: acquiring raw data collected by at least one underground sensor and processing the raw data to obtain multimodal spatiotemporal features; determining the target scenario type based on the multimodal spatiotemporal features using a large language model; determining the corresponding scheduling model set from a small model library based on the target scenario type and the multimodal spatiotemporal features; and obtaining the target mine pressure prediction result based on the multimodal spatiotemporal features and the target scenario type, using the scheduling model set and the large language model. Therefore, this invention can perform semantic parsing and identify the current target scenario type using a large language model, and call the target small model from the scheduling model set matching the target scenario type for mine pressure prediction. This improves the adaptability and accuracy of the mine pressure prediction results, enhances interpretability and engineering usability, provides clear reference for underground dispatchers, and improves the trustworthiness and decision-making efficiency of the mine pressure prediction results.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a mineral pressure prediction method based on a hybrid size model according to an embodiment of the present invention; Figure 2 This is a structural diagram of a mine pressure prediction device based on a hybrid size model according to an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] The following description, with reference to the accompanying drawings, describes a method and apparatus for predicting mine pressure based on a hybrid size model according to an embodiment of the present invention.

[0022] Figure 1 This is a flowchart of a mineral pressure prediction method based on a hybrid size model according to an embodiment of the present invention.

[0023] like Figure 1 As shown, the method may include the following steps: Step 101: Obtain raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features.

[0024] In one embodiment of the present invention, acquiring raw data collected by at least one sensor downhole may include support pressure, hydraulic system status, coal mining machine current and position, roof type, advance speed, and engineering log.

[0025] In one embodiment of the present invention, after acquiring raw data collected by at least one sensor downhole, the raw data can be processed to obtain multimodal spatiotemporal features in order to construct a unified input space and solve the problems of complex data sources, inconsistent frequencies, and misaligned timing in mine pressure prediction.

[0026] In one embodiment of the present invention, the method for processing raw data to obtain multimodal spatiotemporal features may include the following steps: Step 1011: The original data is resampled using a sliding window mechanism to obtain resampled data; Step 1012: Align the resampled data to obtain aligned data; Step 1013: The text information in the alignment data is embedded into the first feature vector by the encoder, and the numerical information in the alignment data is determined as the second feature vector; Step 1014: Concatenate the first feature vector and the second feature vector to obtain multimodal spatiotemporal features.

[0027] In one embodiment of the present invention, a sliding window mechanism is used to resample the original data to address the differences in data sampling frequency and time delay.

[0028] Furthermore, in one embodiment of the present invention, the method for aligning resampled data to obtain aligned data may include: aligning resampled data through timestamp synchronization, missing data imputation, and normalization to obtain aligned data.

[0029] Furthermore, in one embodiment of the present invention, after obtaining the first feature vector and the second feature vector through the above steps, the first feature vector and the second feature vector can be concatenated to obtain multimodal spatiotemporal features. This is to provide high-quality input data for subsequent scene recognition and mine pressure prediction.

[0030] In one embodiment of the present invention, the above-mentioned It can be:

[0031] in, Indicates in Multimodal spatiotemporal features constructed at every moment; This indicates the characteristics of the time-series data acquired by the support pressure sensor; Indicates the operating status characteristics of equipment (such as coal mining machines and conveyors) (such as current, position, and speed); Indicates geological environmental characteristics (such as roof type and support form); This indicates textual description features (such as safety logs and operating condition records).

[0032] Step 102: Based on multimodal spatiotemporal features, determine the target scene type using a large language model.

[0033] In one embodiment of the present invention, after obtaining the multimodal spatiotemporal features through the above steps, the target scene type can be determined based on the multimodal spatiotemporal features using a large language model.

[0034] Specifically, in one embodiment of the present invention, the method for determining the target scene type based on multimodal spatiotemporal features and a large language model may include the following steps: Step 1021: Determine the semantic prompts corresponding to the multimodal spatiotemporal features; Step 1022: Input the multimodal spatiotemporal features and semantic prompts into the large language model to obtain the first probability distribution corresponding to the scene type; Step 1023: Determine the target scene type based on the first probability distribution.

[0035] In one embodiment of the present invention, semantic cue words corresponding to multimodal spatiotemporal features can be determined based on experience and needs. Also, in another embodiment of the present invention, semantic cue words corresponding to multimodal spatiotemporal features can be determined using a cue word model. The cue word model can be an existing model, such as the HiTime hierarchical multimodal model.

[0036] Furthermore, in one embodiment of the present invention, after determining the semantic prompt words corresponding to the multimodal spatiotemporal features, the multimodal spatiotemporal features and semantic prompt words can be input into a large language model to obtain a first probability distribution corresponding to the scene type. The large language model is obtained by fine-tuning using training data from that domain.

[0037] In one embodiment of the present invention, the first probability distribution corresponding to the scene type obtained through the above steps can be expressed as:

[0038] in, Represented as multimodal spatiotemporal features Belongs to scene type The probability of; Represented as the first Scenario types (such as periodic pressure, fault effects, etc.); Indicates by parameters A large language model for control is used for functions that match multimodal spatiotemporal features and semantic cue words with scene types; This represents the standard normalization function and outputs the probability distribution.

[0039] Furthermore, in one embodiment of the present invention, after obtaining the first probability distribution through the above steps, the target scene type can be determined based on the first probability distribution. Specifically, in one embodiment of the present invention, the scene type corresponding to the highest probability in the first probability distribution can be determined as the target scene type.

[0040] Step 103: Based on the target scene type and multimodal spatiotemporal characteristics, determine the corresponding scheduling model set from the small model library.

[0041] In one embodiment of the present invention, after determining the target scene type and multimodal spatiotemporal characteristics through the above steps, the corresponding scheduling model set can be determined from the small model library based on the target scene type and multimodal spatiotemporal characteristics.

[0042] Specifically, in one embodiment of the present invention, the method for determining the corresponding scheduling model set from the small model library based on the target scene type and multimodal spatiotemporal characteristics may include the following steps: Step 1031: Input the target scene type and multimodal spatiotemporal features into the target model selector to obtain the second probability distribution of the small models in the small model library for scheduling; Step 1032: Determine the labeling order of the target small models for scheduling based on the second probability distribution; Step 1033: Obtain the weight coefficients corresponding to the target small model, and determine the label order and weight coefficients of the target small model as the scheduling model set.

[0043] In one embodiment of the present invention, the aforementioned small model library may contain various lightweight and efficient small models, and each small model is pre-trained based on historical data under different scenarios. In one embodiment of the present invention, the aforementioned small model library may include at least one of LSTM, Mamba, TCN, and iTransformer.

[0044] In one embodiment of the present invention, the target model selector described above can be obtained through training.

[0045] In one embodiment of the present invention, the second probability distribution obtained by the target model selector may include small models in the small model library and their corresponding probabilities.

[0046] In one embodiment of the present invention, after obtaining the second probability distribution through the above steps, the method for determining the labeling order of the target small models for scheduling based on the second probability distribution may include: sorting the probabilities in the second probability distribution in descending order, determining the first preset number of small models as target small models, and labeling the target small models in descending order to obtain the labeling order of the target small models.

[0047] In one embodiment of the present invention, the aforementioned preset quantity can be set as needed, such as 1 or 3. Furthermore, when the preset quantity is 1, a single model can be matched; when the preset quantity is greater than 1, multiple models can be matched.

[0048] In one embodiment of the present invention, after obtaining the target small model through the above steps, the weight coefficients corresponding to the target small model can be obtained, and the labeling order of the target small models and the weight coefficients of the target small models are determined as the scheduling model set. In one embodiment of the present invention, the weight coefficients corresponding to each target small model input by the user can be obtained, and the sum of the weight coefficients of all target small models is 1.

[0049] In one embodiment of the present invention, the set of scheduling models obtained through the above steps can be represented as:

[0050] in, Represented as a set of scheduling models, it can specify the order in which target sub-models are called; Represented as the first A small target model; This represents the weight coefficient of the j-th target mini-model that is scheduled; This indicates the number of target small models invoked in this round of scheduling; and This achieves normalized weighting.

[0051] Step 104: Based on multimodal spatiotemporal features and target scene type, obtain the target mine pressure prediction result through a set of scheduling models and a large language model.

[0052] In one embodiment of the present invention, after obtaining the multimodal spatiotemporal features, target scene type and scheduling model set through the above steps, the target mineral pressure prediction result can be obtained based on the multimodal spatiotemporal features and target scene type, through the scheduling model set and large language model.

[0053] Specifically, in one embodiment of the present invention, the method for obtaining target mineral pressure prediction results based on multimodal spatiotemporal features and target scene type through a set of scheduling models and a large language model may include the following steps: Step 1041: In sequence, the target sub-models in the scheduling model set are called, and the first mineral pressure prediction result of each target sub-model is obtained based on the multimodal spatiotemporal characteristics. Step 1042: The weighted sum of the weight coefficients of each target small model and the first mine pressure prediction result is performed to obtain the second mine pressure prediction result; Step 1043: Based on the target scenario type, determine the corresponding set of historical scheduling models and input semantics; Step 1044: Based on the historical scheduling model set and input semantics, generate natural language explanations using a large language model; Step 1045: Combine the explanation with the second mine pressure prediction result to determine the target mine pressure prediction result.

[0054] In one embodiment of the present invention, target small models in the scheduling model set are scheduled sequentially based on multimodal spatiotemporal features. Obtain the future time steps output by each target small model First mine pressure prediction results .

[0055] Furthermore, in one embodiment of the present invention, the weight coefficients of each target small model are used... Corresponding to the first mine pressure prediction results By performing a weighted summation, the second mine pressure prediction result is obtained. ,in,

[0056] in, Indicates to Predicted value of mine pressure at any given time; Indicates the first Small target model For input multimodal spatiotemporal features The first prediction result of mine pressure; Indicates the first The weight of the prediction results of each target small model in the total output.

[0057] Furthermore, in one embodiment of the present invention, the corresponding historical scheduling model set and input semantics can be determined based on the target scenario type. In one embodiment of the present invention, the input semantics can be the historical output semantics corresponding to the historical scheduling set.

[0058] In one embodiment of the present invention, the explanation content generated by the large language model may include a description of the current working condition, reasons for model selection, prediction trend analysis, and risk warning suggestions. Based on this, the large language model can output corresponding explanation content, enhancing the interpretability and engineering usability of the model, providing clear reference for downhole dispatchers, and improving the reliability of prediction results and decision-making efficiency.

[0059] Furthermore, in one embodiment of the present invention, after obtaining the explanation content and the second mineral pressure prediction result through the above steps, the explanation content and the second mineral pressure prediction result can be determined as the target mineral pressure prediction result. Thus, the scenario-driven target small model combination prediction can be realized through the hybrid architecture of large and small models, solving the problem of accuracy degradation caused by fixed models.

[0060] Furthermore, in one embodiment of the present invention, the above method may further include the following steps: Step 105: Obtain the actual collected value at the time corresponding to the target mine pressure prediction result; Step 106: Determine the prediction error based on the target mine pressure prediction results and the actual collected values; Step 107: Determine whether to adjust the target small model in the large model and scheduling model set based on the prediction error; Step 108: If it is determined that the target small model in the set of large models and scheduling models needs to be adjusted, then the target small model in the set of large models and scheduling models is adjusted based on the target mining pressure prediction results and the actual collected values.

[0061] In one embodiment of the present invention, after obtaining the target mine pressure prediction result, the actual collected value at the corresponding time of the target mine pressure prediction result can be obtained. Based on the target mining pressure prediction results and actual collected values Determine the prediction error .

[0062] Furthermore, in one embodiment of the present invention, the method for determining whether to adjust the target small model in the large model and scheduling model set based on the prediction error may include: if the prediction error exceeds a preset threshold, then determining to adjust the target small model in the large model and scheduling model set; if the prediction error does not exceed the preset threshold, then determining not to adjust the target small model in the large model and scheduling model set.

[0063] Furthermore, in one embodiment of the present invention, after determining the adjustment of the target small model in the set of large models and scheduling models through the above steps, the target small model in the set of large models and scheduling models can be adjusted based on the target mineral pressure prediction results and the actual collected values.

[0064] Specifically, in one embodiment of the present invention, the method for adjusting the target small model in the large model and scheduling model set based on the target mining pressure prediction result and the actual collected value may include the following steps: Step 1081: Based on the target mine pressure prediction results and the actual collected values, determine the loss value through a loss function; Step 1082: Adjust the network parameters in the target small model based on the loss value to obtain the adjusted small model; Step 1083: Fine-tune the large model based on the target mine pressure prediction results and the actual collected values ​​to obtain the fine-tuned large model.

[0065] In one embodiment of the present invention, the content of steps 1081 to 1082 described above can be referred to the prior art, and will not be repeated here.

[0066] In one embodiment of the present invention, the large model can be fine-tuned using the target mine pressure prediction results and the actual collected values ​​to obtain the fine-tuned large model, thereby constructing a complete closed loop from input recognition, model prediction to error feedback, and improving long-term adaptability and generalization ability.

[0067] The mine pressure prediction method based on a hybrid big-small model in this invention can acquire raw data collected by at least one underground sensor and process the raw data to obtain multimodal spatiotemporal features. Based on the multimodal spatiotemporal features, the target scenario type is determined through a large language model. Based on the target scenario type and the multimodal spatiotemporal features, a corresponding set of scheduling models is determined from a small model library. Based on the multimodal spatiotemporal features and the target scenario type, the target mine pressure prediction result is obtained through the set of scheduling models and the large language model. Therefore, this invention can perform semantic parsing and identify the current target scenario type through a large language model, and call the target small model from the set of scheduling models matching the target scenario type for mine pressure prediction. This improves the adaptability and accuracy of the mine pressure prediction results, enhances interpretability and engineering usability, provides clear reference for underground dispatchers, and improves the trustworthiness and decision-making efficiency of the mine pressure prediction results.

[0068] Figure 2 This is a schematic diagram of the structure of the mine pressure prediction device 200 based on a hybrid size model according to an embodiment of the present invention.

[0069] like Figure 2 As shown, the device may include: The processing module 201 is used to acquire raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features. The first determining module 202 is used to determine the target scene type based on multimodal spatiotemporal features and a large language model. The second determining module 203 is used to determine the corresponding set of scheduling models from the small model library based on the target scene type and multimodal spatiotemporal characteristics. The prediction module 204 is used to obtain the target mine pressure prediction result based on multimodal spatiotemporal features and target scene type, through a set of scheduling models and a large language model.

[0070] In one embodiment of the present invention, the processing module 201 is specifically used for: The original data is resampled using a sliding window mechanism to obtain resampled data; The resampled data is aligned to obtain aligned data; The textual information in the aligned data is embedded into the first feature vector by the encoder, and the numerical information in the aligned data is determined as the second feature vector. By concatenating the first and second feature vectors, we obtain the multimodal spatiotemporal features.

[0071] In one embodiment of the present invention, the first determining module 202 is specifically used for: Identify semantic cue words corresponding to multimodal spatiotemporal features; By inputting multimodal spatiotemporal features and semantic cue words into a large language model, the first probability distribution corresponding to the scene type is obtained; Based on the first probability distribution, the target scenario type is determined.

[0072] In one embodiment of the present invention, the second determining module 203 is specifically used for: Input the target scene type and multimodal spatiotemporal features into the target model selector to obtain the second probability distribution of the small models in the small model library for scheduling; The labeling order of the target small models used for scheduling is determined based on the second probability distribution; Obtain the weight coefficients corresponding to the target small model, and determine the scheduling model set by the label order of the target small model and the weight coefficients of the target small model.

[0073] In one embodiment of the present invention, the prediction module 204 is specifically used for: The target sub-models in the scheduling model set are scheduled sequentially, and the first mineral pressure prediction result of each target sub-model is obtained based on the multimodal spatiotemporal characteristics; The second mine pressure prediction result is obtained by weighting and summing the weight coefficients of each target small model with the first mine pressure prediction result. Based on the target scenario type, determine the corresponding set of historical scheduling models and input semantics; Based on the historical scheduling model set and input semantics, natural language explanations are generated through a large language model. The explanation and the second mining pressure prediction result are used to determine the target mining pressure prediction result.

[0074] In one embodiment of the present invention, the above-described apparatus is further configured to: Obtain the actual collected value at the corresponding moment of the target mine pressure prediction result; Based on the target mine pressure prediction results and the actual collected values, the prediction error is determined; The decision to adjust the target small model within the large model and scheduling model set is based on the prediction error. If it is determined that the target small model in the set of large models and scheduling models needs to be adjusted, then the target small model in the set of large models and scheduling models needs to be adjusted based on the target mining pressure prediction results and the actual collected values.

[0075] In one embodiment of the present invention, the above-described apparatus is further configured to: If the prediction error exceeds a preset threshold, then adjustments will be made to the target small model in the large model and scheduling model set.

[0076] In one embodiment of the present invention, the above-described apparatus is further configured to: Based on the target mine pressure prediction results and the actual collected values, the loss value is determined through a loss function; The network parameters in the target small model are adjusted based on the loss value to obtain the adjusted small model; The large model is fine-tuned based on the target mine pressure prediction results and the actual collected values ​​to obtain the fine-tuned large model.

[0077] The mine pressure prediction device based on hybrid large and small models in this invention can perform semantic parsing and identify the current target scenario type through a large language model, and call the target small model in the scheduling model set that matches the target scenario type to perform mine pressure prediction. This improves the adaptability and accuracy of the mine pressure prediction results, enhances interpretability and engineering usability, provides clear reference for underground dispatchers, and improves the trustworthiness and decision-making efficiency of the mine pressure prediction results.

[0078] In this specification, the use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refers to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0079] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for predicting mine pressure based on a hybrid size model, characterized in that, include: Acquire raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features; Based on the aforementioned multimodal spatiotemporal features, the target scene type is determined using a large language model; Based on the target scene type and the multimodal spatiotemporal features, a corresponding set of scheduling models is determined from the small model library; the process includes: The target scene type and the multimodal spatiotemporal features are input into the target model selector to obtain the second probability distribution of the small models in the small model library for scheduling. The labeling order of the target small models used for scheduling is determined based on the second probability distribution; Obtain the weight coefficients corresponding to the target small model, and determine the label order of the target small model and the weight coefficients of the target small model as the scheduling model set; Based on the multimodal spatiotemporal features and the target scene type, the target mineral pressure prediction result is obtained through the scheduling model set and the large language model; the process includes: sequentially scheduling the target small models in the scheduling model set, and obtaining the first mineral pressure prediction result of each target small model based on the multimodal spatiotemporal features; The second mineral pressure prediction result is obtained by weighting and summing the weight coefficients of each target small model with the first mineral pressure prediction result. Based on the target scenario type, determine the corresponding set of historical scheduling models and input semantics; Based on the historical scheduling model set and the input semantics, the natural language explanation content is generated through the large language model; The explained content and the second mine pressure prediction result are determined as the target mine pressure prediction result.

2. The method according to claim 1, characterized in that, The process of processing the original data to obtain multimodal spatiotemporal features includes: The original data is resampled using a sliding window mechanism to obtain resampled data. The resampled data is aligned to obtain aligned data; The textual information in the alignment data is embedded into a first feature vector through an encoder, and the numerical information in the alignment data is determined as a second feature vector; The first feature vector and the second feature vector are concatenated to obtain multimodal spatiotemporal features.

3. The method according to claim 1, characterized in that, The process of determining the target scene type based on the multimodal spatiotemporal features using a large language model includes: Determine the semantic cue words corresponding to the multimodal spatiotemporal features; The multimodal spatiotemporal features and the semantic prompt words are input into a large language model to obtain the first probability distribution corresponding to the scene type; Based on the first probability distribution, the target scene type is determined.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the actual collected value at the time corresponding to the target mine pressure prediction result; Based on the target mine pressure prediction results and the actual collected values, the prediction error is determined; Based on the prediction error, determine whether to adjust the large language model and the target small model in the scheduling model set; If it is determined that adjustments need to be made to the large language model and the target small model in the scheduling model set, then the adjustments are made based on the target mineral pressure prediction results and the actual collected values.

5. The method according to claim 4, characterized in that, The step of determining whether to adjust the large language model and the target small model in the scheduling model set based on the prediction error includes: If the prediction error exceeds a preset threshold, then it is determined that adjustments should be made to the large language model and the target small model in the scheduling model set.

6. The method according to claim 4, characterized in that, The adjustment of the target small model in the large language model and the scheduling model set based on the target mineral pressure prediction result and the actual collected value includes: Based on the target mineral pressure prediction results and the actual collected values, the loss value is determined through a loss function; Based on the loss value, the network parameters in the target small model are adjusted to obtain the adjusted small model; The large language model is fine-tuned based on the target mine pressure prediction results and the actual collected values ​​to obtain the fine-tuned large model.

7. A mine pressure prediction device based on a hybrid size model, characterized in that, include: The processing module is used to acquire raw data collected by at least one downhole sensor and process the raw data to obtain multimodal spatiotemporal features. The first determining module is used to determine the target scene type based on the multimodal spatiotemporal features using a large language model; The second determining module is used to determine the corresponding scheduling model set from the small model library based on the target scene type and the multimodal spatiotemporal features; the process includes: The target scene type and the multimodal spatiotemporal features are input into the target model selector to obtain the second probability distribution of the small models in the small model library for scheduling. The labeling order of the target small models used for scheduling is determined based on the second probability distribution; Obtain the weight coefficients corresponding to the target small model, and determine the label order of the target small model and the weight coefficients of the target small model as the scheduling model set; The prediction module is used to obtain the target mineral pressure prediction result based on the multimodal spatiotemporal features and the target scene type, through the scheduling model set and the large language model; the process includes: sequentially scheduling the target small models in the scheduling model set, and obtaining the first mineral pressure prediction result of each target small model based on the multimodal spatiotemporal features; The second mineral pressure prediction result is obtained by weighting and summing the weight coefficients of each target small model with the first mineral pressure prediction result. Based on the target scenario type, determine the corresponding set of historical scheduling models and input semantics; Based on the historical scheduling model set and the input semantics, the natural language explanation is generated through the large language model; The explained content and the second mine pressure prediction result are determined as the target mine pressure prediction result.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images