Coal mine typical disaster risk heterogeneous information collaborative fusion and hierarchical prediction method

Abstract

The application discloses a coal mine typical disaster risk heterogeneous information collaborative fusion and hierarchical prediction method, relates to the technical field of coal mine disaster risk, and comprises the following steps: response trajectory extraction is carried out on various types of heterogeneous features participating in hierarchical prediction in a plurality of labeled coal mine typical disaster samples by constructing a feature decision response contrast set, and based on the trigger offset degree calculation result of the disaster risk level variation node, it is judged whether there is a decision importance difference in the heterogeneous features participating in hierarchical prediction; in the case that it is judged that there is a decision importance difference, the feature decision scale value of each heterogeneous feature participating in hierarchical prediction is calculated according to the time leading property, trigger strength and risk boundary penetration ability in the response trajectory. The application solves the problem that the decision importance of heterogeneous features is averaged in hierarchical prediction, so that the key risk features are difficult to highlight, and realizes the hierarchical prediction effect that the feature importance is quantifiable, the expression is controllable, and the weight is dynamically adjusted.

Description

Technical Field

[0001] This invention relates to the field of coal mine disaster risk technology, specifically to a method for the collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks. Background Technology

[0002] Collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks refers to the comprehensive utilization of heterogeneous information from different sources, formats, and dimensions in the coal mine production environment, targeting typical disaster types such as gas explosions, coal dust explosions, water hazards, and roof falls. This includes data from gas concentration sensor data, geological monitoring data, video surveillance images, work logs, and historical accident records. Through data preprocessing, feature extraction, spatiotemporal alignment, and semantic mapping, structured fusion is achieved. Based on this, a cross-modal collaborative computing mechanism is established to achieve information supplementation, mutual verification, and coupled modeling among multi-source data, thereby forming a comprehensive and dynamic perception of disaster risk status. In existing technologies, multi-source information is typically collected and initially cleaned through a data integration platform. Then, a rule engine or machine learning-based model is used to extract features and train risk level classification. The prediction portion employs traditional support vector machines, random forests, or more recently used deep learning models for risk assessment. Based on set thresholds, risks are categorized into low, medium, high levels, or emergency warning states. Finally, the tiered prediction results are pushed to the monitoring center and relevant personnel through an early warning system, enabling early detection and response to coal mine disaster risks. The entire process typically includes five core stages: heterogeneous data collection, fusion processing, feature modeling, risk prediction, and result delivery. These stages work together collaboratively to improve the accuracy of coal mine disaster prediction and response efficiency.

[0003] The existing technology has the following shortcomings: In the process of collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks, when the system needs to use multi-source heterogeneous features such as gas disturbance, roadway subsidence, image recognition results, and personnel work trajectories for hierarchical prediction, these features have different decision-making importance in disaster evolution judgment. Since the fused features have not undergone importance analysis and dominance identification before being input into the model, the model treats all heterogeneous features as equally important during the training phase, thus weakening key risk features in the overall feature set. In this situation, the model cannot correctly identify the dominant disaster factors, and the prediction results tend to be averaged. Existing collaborative fusion and hierarchical prediction technologies for typical coal mine disaster risks cannot dynamically adjust the influence weight of each heterogeneous feature in the hierarchical prediction model based on the differences in decision-making importance among the heterogeneous features participating in the hierarchical prediction. This results in the model being unable to accurately highlight the risk contribution of key disaster features when the risk approaches the hierarchical threshold, leading to blurred prediction boundaries, biased risk level classification, and ultimately, the failure to identify high-risk areas in a timely manner, affecting the accuracy of disaster early warning and the efficiency of coal mine safety response.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for the collaborative fusion and hierarchical prediction of heterogeneous information on typical disaster risks in coal mines, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks, specifically including the following steps: S1. By constructing a feature decision response comparison set, response trajectories are extracted from various heterogeneous features participating in graded prediction in multiple labeled typical coal mine disaster samples. Based on the calculation results of the trigger offset degree of disaster risk level change nodes, it is determined whether there are differences in decision importance among the heterogeneous features participating in graded prediction. S2. In the case of determining that there are differences in decision importance, calculate the feature decision scale value of each heterogeneous feature participating in the hierarchical prediction based on the time precedence, triggering intensity and risk boundary penetration capability in the response trajectory, so as to determine the feature importance of each heterogeneous feature. S3. Based on the feature importance of each heterogeneous feature, perform expression tensor mapping on the original sequences of each type of heterogeneous feature, and project each heterogeneous feature to the corresponding channel of the expression tensor according to the feature importance, so as to generate a fusion expression tensor in which the expression channel depth is related to the feature importance. S4. By guiding the fusion expression tensor into the hierarchical prediction path, setting a decision priority focusing channel, prioritizing the activation of heterogeneous feature channels whose feature importance ranking is at the front index position of the feature decision sequence, and constructing a focusing path based on feature importance control, a risk classification judgment process dominated by features is established. S5. After obtaining the hierarchical prediction results, construct a feature contribution offset map based on the prediction output offset, and adjust the channel activation ratio in the fusion expression tensor in combination with the feature importance to dynamically adjust the influence weight of each heterogeneous feature in the hierarchical prediction process.

[0007] Preferably, S1 specifically includes the following steps: S101. By organizing the time series records of multiple labeled typical coal mine disaster samples, the original data of heterogeneous features participating in the graded prediction at each disaster stage are aligned with a unified time benchmark, and a feature decision response comparison set containing time markers, feature value sequences and risk level markers is constructed. S102. Based on the feature decision response comparison set, identify the nodes of disaster risk level change in the time series, and calculate the slope of numerical change, starting point position and change magnitude of various heterogeneous features participating in the graded prediction before and after the change nodes. After calculation, generate the trigger offset degree of each type of heterogeneous feature, and extract the response trajectory accordingly. S103. By normalizing the trigger offset degree corresponding to various heterogeneous features participating in hierarchical prediction and calculating the difference between each pair, when the difference between the normalized trigger offset degree of any two heterogeneous features exceeds the preset judgment threshold, it is determined that there is a difference in decision importance among the heterogeneous features participating in hierarchical prediction.

[0008] Preferably, S102 specifically is: By arranging the time series in the characteristic decision response comparison set according to the risk level label order, the jump points of the risk level label are used to determine the disaster risk level change nodes, and the time index corresponding to the disaster risk level change node is used as the benchmark to extract continuous raw data segments from the fixed time span before and after the time index to form the pre-change sequence and post-change sequence for calculation. In the pre-change and post-change sequences, continuous data of heterogeneous features participating in hierarchical prediction are calculated respectively. The slope of numerical change is determined by the rate of change of the difference between adjacent data points in the pre-change and post-change sequences. The starting point position is determined by the time point when the continuous data first produces an offset. The magnitude of change is determined by the maximum difference range between the pre-change and post-change sequences. The set of input parameters for triggering the offset degree is constructed using three continuously changing parameters. The trigger offset degree is generated by combining the set of input parameters in a preset order and mapping them to a unified numerical domain. The trigger offset degree is then arranged along the time index to form a response trajectory, which is used as the input for subsequent processing.

[0009] Preferably, S2 specifically includes the following steps: S201. When it is determined that there are differences in decision importance, the response trajectory of each heterogeneous feature participating in the graded prediction is analyzed by time index analysis to identify the time position in the response trajectory where the degree of trigger offset first continuously increases. The time index corresponding to the time position is calculated with the time index of the disaster risk level change node. The difference is used as the time precursor to indicate the degree of advance of the response of the heterogeneous feature in the risk evolution process. S202. Determine the maximum value of the trigger offset in the response trajectory, and use the maximum value as the trigger intensity. Obtain the risk level mark corresponding to the occurrence of the value, and count the value range of the risk level mark in each continuous time period before and after the risk level mark. If the value range spans at least two different risk level intervals, the heterogeneous feature has the ability to penetrate risk boundaries, which is used to reflect the depth of its influence on the multi-level risk boundary. S203. Convert time-leading characteristics, triggering intensity, and risk boundary penetration capability into feature decision scale values ​​according to preset mapping rules. Normalize and sort all feature decision scale values, and determine the feature importance of each heterogeneous feature participating in hierarchical prediction based on the sorting results.

[0010] Preferably, S203 is as follows: By combining time-leadingness, triggering intensity, and risk boundary penetration capability into a three-dimensional input vector according to a preset order, and performing linear scaling on each component of the three-dimensional input vector according to a preset mapping rule, the three-dimensional input vector forms a mapping result in a unified numerical domain. This mapping result serves as the feature decision scale value for each heterogeneous feature participating in hierarchical prediction. By performing normalization on all feature decision scale values, all feature decision scale values ​​fall within a set continuous closed interval. Based on the normalized feature decision scale values, a feature decision sequence is generated from the position of maximum value to the position of minimum value, so that the feature decision sequence forms a sequential index with a single sorting direction. By assigning feature importance based on sequential index position in the feature decision sequence, the position with the first sequential index is assigned the maximum weight value in the preset weight domain, the position with the last sequential index is assigned the minimum weight value in the preset weight domain, and features in the middle index position are assigned equal-interval weight values ​​between the maximum and minimum weight values ​​according to the index order.

[0011] Preferably, S3 is as follows: Establish a correspondence between the original sequences of various heterogeneous features and their corresponding feature importance values. By traversing the feature importance values ​​of each type of heterogeneous feature, construct a feature importance ranking table according to the values ​​from high to low, and allocate channel index positions in the expression tensor based on the feature importance ranking table to form a mapping rule between feature importance and expression tensor channels. According to the mapping rules, the original sequences of each type of heterogeneous features are projected to the specified channel positions of the expression tensor according to their feature importance. During the projection process, the original sequences are normalized, and the normalized sequence values ​​are filled into the time series dimension of the corresponding channel in the expression tensor according to the channel index to form a primary tensor channel set with feature importance expression features. In the primary tensor channel set, hierarchical splicing is performed on all filled channels along the channel dimension of the expression tensor in order of feature importance, so that the channel arrangement structure is consistent with the feature importance, and a fused expression tensor with the expression channel depth correlated with the feature importance is generated.

[0012] Preferably, S4 is as follows: By inputting the fusion expression tensor into the input node of the hierarchical prediction path according to the time series dimension, the channel dimension structure of the fusion expression tensor is parsed in the input node, and the position identification of all channels of the fusion expression tensor is performed according to the index order in the feature decision sequence. The decision priority focusing channel attribute is marked on the channel whose channel index position is at the beginning of the feature decision sequence, so as to complete the guidance of the fusion expression tensor to the decision priority focusing channel in the hierarchical prediction path. In the decision priority focusing channel, channel activation weight enhancement parameters are set, and according to the correspondence between channel index and feature importance, the amplified expression tensor value is injected into the channel to enhance the channel activation intensity. This activation enhancement behavior is recorded in the channel weight control matrix, so that heterogeneous feature channels with feature importance ranking at the front index position of the feature decision sequence are preferentially activated in the hierarchical prediction path. By reading the activation order in the channel weight control matrix, a focusing path controlled by feature importance is constructed in the channel dimension. This focusing path is then used as the input index control sequence of the feedforward layer in the hierarchical prediction path, ensuring that the response order of the prediction path is consistent with the feature decision sequence. This establishes a risk classification judgment process dominated by feature importance.

[0013] Preferably, S5 is as follows: After obtaining the hierarchical prediction results, the numerical offset between the predicted output and the true label at each time index position is extracted. This numerical offset is then mapped to the channel corresponding to each heterogeneous feature participating in the hierarchical prediction according to the channel structure of the fusion expression tensor. By statistically analyzing the mean and variance of the predicted output offset on each channel, a feature contribution offset map is constructed to express the degree of influence of each heterogeneous feature on the current predicted output offset. Based on the contribution intensity of each channel in the feature contribution offset map and the feature importance of the corresponding channel in the fusion representation tensor, a weighted fusion calculation is performed to generate a channel activation ratio adjustment factor table, and the channel activation ratio update magnitude of each channel is determined based on the channel activation ratio adjustment factor table. In the channel dimension of the fusion expression tensor, the original activation ratio is adjusted by gain or suppression according to the update magnitude of the channel activation ratio, and the adjusted fusion expression tensor is re-input into the hierarchical prediction path to realize the dynamic adjustment of the influence weight of each heterogeneous feature participating in hierarchical prediction in the hierarchical prediction process.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention improves the expression efficiency and decision participation of multi-source heterogeneous features in the risk classification prediction of typical coal mine disasters by constructing a collaborative fusion mechanism that integrates feature response difference identification, feature importance measurement, expression tensor mapping, prediction path focusing control, and dynamic feedback adjustment. By introducing indicators such as time precedence, triggering intensity, and risk boundary penetration capability, the invention can accurately characterize the response dominance of different heterogeneous features in the disaster evolution process and determine feature importance accordingly. It achieves ordered binding of features and channels in the expression tensor structure, giving key risk features higher expression priority. This avoids the problem of weakened dominant information caused by the averaging of heterogeneous features in traditional models, thus improving the sensitivity and response strength of the prediction model to high-risk signs.

[0015] 2. The prediction path focusing mechanism and feature contribution offset feedback adjustment mechanism introduced in this invention support the model to dynamically update the activation ratio of each channel according to the offset between the prediction result and the actual risk during operation. This achieves adaptive optimization of the allocation of heterogeneous feature influence, thereby further enhancing the model's ability to judge the critical range of risk level, reducing risk level classification errors, and improving the model's discrimination accuracy and stability in high-risk identification scenarios. This mechanism supports the rapid response and accurate classification of typical coal mine disasters by the disaster risk early warning system, providing a model support system with intelligent adjustment capabilities for disaster prevention and control, and enhancing the safety response efficiency and practicality of the intelligent risk identification system in mines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart illustrating the method for collaborative fusion and hierarchical prediction of heterogeneous information on typical disaster risks in coal mines according to the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] This invention provides, for example Figure 1 The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks, as shown, specifically includes the following steps: S1. By constructing a feature decision response comparison set, response trajectories are extracted from various heterogeneous features participating in graded prediction in multiple labeled typical coal mine disaster samples. Based on the calculation results of the trigger offset degree of disaster risk level change nodes, it is determined whether there are differences in decision importance among the heterogeneous features participating in graded prediction. In this embodiment, S1 specifically includes the following steps: S101. By organizing the time series records of multiple labeled typical coal mine disaster samples, the original data of heterogeneous features participating in the graded prediction at each disaster stage are aligned with a unified time benchmark, and a feature decision response comparison set containing time markers, feature value sequences and risk level markers is constructed. When constructing a feature-based decision-response comparison set, it is first necessary to organize the data records of multiple labeled typical coal mine disaster samples. These sample data usually come from historical monitoring data warehouses of real disaster events. Each sample contains heterogeneous information in multiple dimensions, such as gas concentration sensor readings, roadway displacement data, video image recognition results, and personnel positioning trajectories. These data have problems such as different sampling frequencies and inconsistent timestamps. Therefore, it is necessary to first select a unified time reference, such as constructing a continuous time axis in seconds, and use methods such as time interpolation, sliding window synchronization, or event anchor alignment to resample and synchronize all data to this time axis. After alignment, the original values ​​of each type of heterogeneous feature at the unified time are encapsulated together with the corresponding risk level label to form a composite record containing time stamps, feature value sequences, and risk level labels, thus constituting the feature-based decision-response comparison set. For example, in an actual gas anomaly, the gas concentration rises continuously at a certain point in time, the roadway settlement data changes abruptly at the same time, and the location of the workers shows that their trajectory stops. At this time, the three types of features form synchronous data on a unified time axis, and together with the risk level mark at that point in time, they constitute a sample. This organization method can more realistically reflect the correspondence between each feature and the risk evolution, which helps to identify the importance of features and response trajectories in the future.

[0020] The labeled typical coal mine disaster samples refer to the collection of data records in historical data for which risk levels of typical coal mine disasters have been annotated manually or through models. These records cover the entire process of disasters before, during, and after their occurrence. Heterogeneous features involved in the graded prediction refer to different types of information sources used to determine risk levels, including structured sensor data (such as gas concentration and wind speed), event tags extracted from image data (such as smoke and fire), and worker behavior data based on positioning systems. These features vary in form and source. Raw data for each disaster stage refers to unprocessed multi-source monitoring data collected at different time periods during the disaster, used to reflect the dynamic evolution of the disaster. The feature-based decision response comparison set is a data structure composed of uniformly aligned time stamps, multiple heterogeneous feature value sequences, and risk level labels for each time point. It supports subsequent trigger offset calculations and feature-based decision modeling, serving as the foundation for dynamic response analysis and importance identification.

[0021] S102. Based on the feature decision response comparison set, identify the nodes of disaster risk level change in the time series, and calculate the slope of numerical change, starting point position and change magnitude of various heterogeneous features participating in the graded prediction before and after the change nodes. After calculation, generate the trigger offset degree of each type of heterogeneous feature, and extract the response trajectory accordingly. S103. By normalizing the trigger offset degree corresponding to various heterogeneous features participating in hierarchical prediction and calculating the difference between each pair, when the difference between the normalized trigger offset degree of any two heterogeneous features exceeds the preset judgment threshold, it is determined that there is a difference in decision importance among the heterogeneous features participating in hierarchical prediction.

[0022] To determine whether there are significant differences in decision importance among heterogeneous features participating in hierarchical prediction, it is necessary to normalize the trigger offsets generated by various heterogeneous features under the same time index. The purpose of normalization is to map trigger response values ​​from features of different sources, dimensions, and amplitudes to the same numerical range. Standardization methods such as max-min scaling or Z-score transformation are commonly used to ensure that the response values ​​of different features are comparable in numerical dimensions. After normalization, the absolute difference between the normalized trigger offsets of any two heterogeneous features is calculated. By iterating through combinations and comparing, the differences between all feature pairs are obtained. To quantify the significance of the differences, a preset judgment threshold, set through empirical statistics or dynamic training, is introduced. When the difference in the normalized trigger offset of any pair of heterogeneous features exceeds this threshold, it can be determined that there is a hierarchical division of response intensity between the feature group, thus identifying a difference in decision importance. In practical implementation, for example, if the difference in the normalized offset between gas disturbance and the work trajectory far exceeds the threshold, it indicates that gas disturbance is more sensitive to changes in risk level during that period and should be given higher weight in subsequent predictions.

[0023] Normalization refers to mapping the trigger offset of multiple heterogeneous features to a unified standard scale, making different features numerically equivalent and enabling horizontal comparison. This process eliminates the influence of unit differences and improves the stability and accuracy of subsequent calculations. Normalization methods can be linear or nonlinear, depending on the feature distribution, and can be combined with weighting factors to highlight the performance of key response features. A preset judgment threshold is a numerical boundary used to determine whether there is a response difference between two heterogeneous features. This threshold can be set based on the distribution of trigger offset differences in historical samples or obtained through outlier analysis and stability screening. In specific scenarios, to avoid oversensitivity or misjudgment, the preset judgment threshold can be dynamically adjusted according to the environment; for example, the judgment boundary can be relaxed in the early stages of disaster evolution, while the threshold judgment accuracy can be improved as the risk approaches high. By combining normalization with threshold comparison, feature combinations with significantly different decision-making impact can be effectively identified, providing a basis for subsequent feature importance ranking and weight allocation.

[0024] In this embodiment, S102 specifically refers to: By arranging the time series in the characteristic decision response comparison set according to the risk level label order, the jump points of the risk level label are used to determine the disaster risk level change nodes, and the time index corresponding to the disaster risk level change node is used as the benchmark to extract continuous raw data segments from the fixed time span before and after the time index to form the pre-change sequence and post-change sequence for calculation. In constructing the response trajectory, the time series needs to be segmented using changes in risk level as key nodes. First, all records in the feature decision response comparison set are arranged chronologically according to the risk level marker order, ensuring that the risk level marker for each time point unfolds linearly according to the time evolution process. By detecting abrupt changes in continuous data within the risk level markers—that is, the time positions where the risk level value shows discontinuous changes from low to high or from high to low—the nodes of disaster risk level change can be accurately determined. These nodes represent moments when the risk state undergoes significant changes and are key anchor points for judging the feature's response capability. To facilitate the quantification of the response of heterogeneous features before and after abrupt changes in risk level, a fixed-length time span is set before and after the time index corresponding to the risk level change node, for example, selecting a time window of 5 minutes before and 5 minutes after the change point. Continuous values ​​of each type of heterogeneous feature within this time window are extracted from the original aligned data, serving as the pre-change sequence and post-change sequence. The data segments formed in this way can be used for subsequent calculations of the changing trend and response intensity of features during disaster evolution. The risk level labeling order refers to the continuous sequence of labels corresponding to risk levels after being sorted by time; the jump point is the position where the risk level label value first changes abruptly; the fixed time span is used to standardize the comparison range of the analysis window before and after the change; the pre-change sequence and the post-change sequence represent the evolution path of feature values ​​before and after the change, respectively, which can provide the dynamic behavior characteristics of each heterogeneous feature before and after the risk change. In this way, a clear and continuous data foundation can be provided for calculating the degree of trigger offset, further enhancing the accuracy and temporal correlation of feature response analysis.

[0025] In the pre-change and post-change sequences, continuous data of heterogeneous features participating in hierarchical prediction are calculated respectively. The slope of numerical change is determined by the rate of change of the difference between adjacent data points in the pre-change and post-change sequences. The starting point position is determined by the time point when the continuous data first produces an offset. The magnitude of change is determined by the maximum difference range between the pre-change and post-change sequences. The set of input parameters for triggering the offset degree is constructed using three continuously changing parameters. When performing dynamic response calculations for each type of heterogeneous feature participating in graded prediction in the pre- and post-change sequences, it is necessary to extract core parameters reflecting the trend and intensity of feature changes. First, by calculating the numerical differences between adjacent data points in the pre- and post-change sequences, the rate of change at each moment is obtained. The average of these rates is then calculated to reflect the slope of the feature's numerical change within that time period; a larger slope indicates a more drastic change. Next, analyzing the location in the entire time series where the feature value first deviates significantly from the previous stable range identifies the starting point where the heterogeneous feature begins to respond to risk changes. The position of this location on the time axis helps measure the leading nature of the feature's response. Then, by calculating the maximum numerical difference between the pre- and post-change sequences, the magnitude of change of the heterogeneous feature during the risk level change process is determined. This parameter reflects the intensity of the feature's response during disaster evolution. Combining these three representative change indicators—the slope of numerical change, the starting point location, and the magnitude of change—forms a trigger offset input parameter set. This parameter set serves as the basis for subsequently generating trigger offset values ​​and can fully express the changing trends, response timing, and response intensity of different heterogeneous features during disaster risk evolution. The slope of numerical changes measures the intensity of the response, the starting point reflects the timing of the response, and the magnitude of change reflects the strength of information contribution. The integration of these three factors forms a stable system of input variables, which can be used for standardized evaluation and importance comparison of different response characteristics. For example, in a high-risk evolution triggered by a gas leak, the slope of the gas concentration characteristic increases rapidly, the offset point appears earlier, and the numerical increase is large, while the slope of the personnel trajectory change is slow, the offset is later, and the magnitude is small. This suggests that the former has a higher response dominance in this evolution.

[0026] The trigger offset degree is generated by combining the set of input parameters in a preset order and mapping them to a unified numerical domain. The trigger offset degree is then arranged along the time index to form a response trajectory, which is used as the input for subsequent processing.

[0027] To quantify the response intensity of different heterogeneous features during changes in disaster risk levels, the previously extracted trigger offset input parameter set needs to be uniformly encoded. First, combining the parameters in a preset order involves concatenating the slope of numerical changes, starting point positions, and magnitudes of change according to a fixed arrangement rule. This ensures that the parameters maintain a consistent logical relationship during the combination process, guaranteeing repeatability and comparability in subsequent mapping. Next, to eliminate unit differences in the parameters at their original scale, the combined data needs to be uniformly mapped to a standardized numerical domain, such as a closed interval between zero and one. The mapping process for a unified numerical domain typically employs normalization or minimax scaling strategies and can be combined with weighting factors to enhance the expression of the response contribution of key parameters. The single index generated through this numerical mapping process is the trigger offset, used to quantify the response intensity of a heterogeneous feature at a specific moment of risk change. Subsequently, the trigger offsets generated at multiple time index points are arranged chronologically to form a continuous response trajectory. The response trajectory reflects the evolution of a feature's response to changes in disaster risk over time, serving as a crucial input for subsequently assessing the differences in feature importance and constructing a feature-driven prediction process. Combining features in a pre-defined order ensures consistent parameter structures before mapping, while a unified numerical domain provides a mathematical basis for horizontal comparisons. The trigger offset degree, as a single-point intensity indicator, is highly operable, and the response trajectory reflects the temporal behavior pattern of the feature. For example, in a real-world case, the response trajectory of the gas concentration feature continuously rises before the risk increases, with a higher amplitude than other features. The corresponding trigger offset degree curve exhibits a significant slope and leading characteristic, indicating that this feature has a stronger dominant role in disaster evolution.

[0028] S2. In the case of determining that there are differences in decision importance, calculate the feature decision scale value of each heterogeneous feature participating in the hierarchical prediction based on the time precedence, triggering intensity and risk boundary penetration capability in the response trajectory, so as to determine the feature importance of each heterogeneous feature. In this embodiment, S2 specifically includes the following steps: S201. When it is determined that there are differences in decision importance, the response trajectory of each heterogeneous feature participating in the graded prediction is analyzed by time index analysis to identify the time position in the response trajectory where the degree of trigger offset first continuously increases. The time index corresponding to the time position is calculated with the time index of the disaster risk level change node. The difference is used as the time precursor to indicate the degree of advance of the response of the heterogeneous feature in the risk evolution process. When extracting temporal precedence, it is necessary to first perform time index analysis on the response trajectory generated by each heterogeneous feature participating in the graded prediction. This involves scanning the trigger offset value in the response trajectory point by point along a unified time axis. During the scanning process, the position where the trigger offset degree enters the continuously rising range from the stable range is identified. This position is the time position where the trigger offset degree first begins to rise continuously. The identification method can be determined by judging the numerical increment of adjacent data points. When multiple consecutive time points show positive increments, the starting position of the continuous rise can be determined. Subsequently, the time index of this time position is compared with the time index of the disaster risk level change node to calculate the time precedence, which is used to quantify the advance magnitude of the response generated by this heterogeneous feature during the risk evolution process. For example, in a gas anomaly event, the trigger offset degree of the gas concentration feature may begin to rise continuously at multiple time points before the risk level change occurs, while the trigger offset degree of the roadway subsidence feature may only respond near the change node. By calculating the time index difference, it can be found that the time precedence of the gas concentration feature is greater than that of the roadway subsidence feature, thus indicating that this feature has an earlier perception capability for disaster evolution.

[0029] Time index analysis is a sequence analysis method based on a unified time axis. By traversing the trigger offset values ​​point by point in the response trajectory, it identifies characteristic patterns of numerical changes over time, thus characterizing the dynamic response process of heterogeneous features to disaster evolution. The time position of the first sustained increase in the trigger offset in the response trajectory refers to the first time point when the trigger offset shows an upward trend across multiple consecutive time points, representing the time node when the heterogeneous feature first exhibits a significant response to risk changes. Time leadingness refers to the difference between the time index corresponding to the time position of the first sustained increase and the time index of the disaster risk level change node, measuring the ability of the heterogeneous feature to perceive risk trends before risk level changes occur. The larger the time leadingness, the more obvious the leading response of the feature in risk evolution. By comprehensively using these technical features, the differences in response speed and leadingness of different heterogeneous features in the risk evolution process can be objectively judged, providing quantifiable temporal basis for subsequent calculation of feature decision scale values.

[0030] S202. Determine the maximum value of the trigger offset in the response trajectory, and use the maximum value as the trigger intensity. Obtain the risk level mark corresponding to the occurrence of the value, and count the value range of the risk level mark in each continuous time period before and after the risk level mark. If the value range spans at least two different risk level intervals, the heterogeneous feature has the ability to penetrate risk boundaries, which is used to reflect the depth of its influence on the multi-level risk boundary. When calculating trigger strength and risk boundary penetration capability based on the response trajectory, it is necessary to first find the maximum value of the trigger offset in the response trajectory and use this maximum value as the trigger strength, representing the peak response strength of the heterogeneous feature during risk level changes. To further analyze the depth of influence of this peak value during risk level changes, it is necessary to locate the risk level marker corresponding to the occurrence of the maximum value, and collect a continuous risk level marker sequence within a fixed duration interval before and after the time point where the risk level marker is located. By statistically analyzing the value range of this risk level marker sequence, it can be determined whether the response of the heterogeneous feature is associated with multi-level risk intervals. When the value range spans at least two different risk level intervals, it indicates that the heterogeneous feature not only responds to a single level during risk evolution but also exhibits influence at a broader risk level scale, thus possessing risk boundary penetration capability. Taking gas concentration characteristics as an example, if the maximum value of the trigger offset occurs near the transition of the risk level from medium risk to high risk, and the value of the risk level marker in the consecutive time periods before and after the transition includes both medium risk and high risk ranges, it indicates that the gas concentration not only has a high response intensity in this risk evolution, but also has a deep involvement in the risk level transition. It is a typical characteristic type with risk boundary penetration capability.

[0031] Trigger strength refers to the maximum value of the trigger offset in the response trajectory, used to quantify the extreme response capability of a heterogeneous feature when it affects changes in risk level. The value range spans at least two different risk level intervals, meaning that within a specified time period before and after the occurrence of the maximum trigger offset, the risk level marker values ​​include markers from both lower and higher risk level intervals, thus forming a crossing relationship between level intervals. Risk boundary penetration capability is a capability index defined based on this level interval crossing phenomenon, used to describe whether a heterogeneous feature can maintain a continuous response during risk level boundary transitions, reflecting the feature's penetration and correlation depth in the risk rise and fall process. Heterogeneous features with risk boundary penetration capability can typically continuously generate responses across risk level boundaries and are key input factors that require higher feature importance in subsequent classification predictions.

[0032] S203. Convert time-leading characteristics, trigger strength, and risk boundary penetration capability into feature decision scale values ​​according to preset mapping rules. Normalize and sort all feature decision scale values, and determine the feature importance of each heterogeneous feature participating in hierarchical prediction based on the sorting results, so that it has measurable weight to participate in subsequent prediction judgment.

[0033] Converting time-leading characteristics, trigger strength, and risk boundary penetration capability into feature decision scale values ​​according to preset mapping rules, and then normalizing and ranking all feature decision scale values, aims to construct a unified standard that can accurately measure the influence of various heterogeneous features participating in graded prediction during risk evolution, even when differences in decision importance exist. Since time-leading characteristics reflect a feature's ability to perceive risk warnings in advance, trigger strength measures the magnitude of a feature's response to changes in disaster levels, and risk boundary penetration capability describes whether a feature's influence spans multiple risk level intervals, these three attributes differ significantly in their dimensions, distribution, and interpretation. Without mapping and normalization within a unified numerical domain, it is impossible to organically combine these three into comparable feature decision scale values, nor can a fair and effective ranking basis be formed for all heterogeneous features. This process not only allows for the structured expression of the original response information but also provides a logical foundation for subsequently allocating explicit weights in the graded prediction path, thereby enabling data-driven identification of key features, highlighting dominant factors, and improving the targeting and accuracy of graded prediction.

[0034] In this embodiment, S203 specifically refers to: By combining time-leadingness, triggering intensity, and risk boundary penetration capability into a three-dimensional input vector according to a preset order, and performing linear scaling on each component of the three-dimensional input vector according to a preset mapping rule, the three-dimensional input vector forms a mapping result in a unified numerical domain. This mapping result serves as the feature decision scale value for each heterogeneous feature participating in hierarchical prediction. To transform time-leading characteristics, trigger strength, and risk boundary penetration capability into a unified discriminative index, the three parameters must first be combined into a three-dimensional input vector according to a preset order. This order can be set based on empirical weights, for example, prioritizing time response capability, then reaction strength, and finally assessing the boundary penetration range. After combination, each component of the three-dimensional input vector undergoes linear scaling, mapping it from its original numerical range to a unified numerical domain. The unified numerical domain must ensure comparability between features and is often set as a closed numerical interval, such as a linear segment between zero and one. The linear scaling process can be based on the maximum and minimum values ​​of each feature's historical data, using a scaling factor to uniformly compress the differences in feature dimensions to the same standard scale, avoiding comparison distortion caused by differences in the original dimensions or amplitudes of different parameters. The mapping result is the standard expression of the heterogeneous feature participating in hierarchical prediction under the three-dimensional index, reflecting its comprehensive response capability. This comprehensive expression is defined as the feature decision scale value. Feature decision scale values ​​serve as the fundamental quantitative basis for subsequently determining the relative importance of heterogeneous features in hierarchical prediction, possessing good interpretability and universality. This approach ensures that the model fully considers the behavioral differences of input features during the risk response process when evaluating their impact, effectively improving the specificity of prediction and classification.

[0035] By performing normalization on all feature decision scale values, all feature decision scale values ​​fall within a set continuous closed interval. Based on the normalized feature decision scale values, a feature decision sequence is generated from the position of maximum value to the position of minimum value, so that the feature decision sequence forms a sequential index with a single sorting direction. To ensure the comparability of heterogeneous features participating in hierarchical prediction across feature decision scale values, all feature decision scale values ​​must first be normalized to map them to a uniform numerical range. This range is typically set as a continuous closed interval between zero and one to avoid value dispersion that could lead to sorting confusion or weight imbalance. Normalization can be achieved through linear scaling, using the maximum and minimum values ​​of the feature decision scale values ​​in the current dataset for interval mapping. After normalization, the features are arranged sequentially from the largest to the smallest value, generating an ordered feature decision sequence. This sequence has a single sorting direction, ensuring that each feature's position in the sequence is unique. This method forms an ordered index, facilitating the subsequent direct assignment of feature importance values ​​according to the index position. For example, if a heterogeneous feature has the largest normalized scale value among all features, it will be at the first position in the feature decision sequence, and its ordered index will be first; while the feature with the smallest value will be at the last position in the sequence. This ordered index not only reflects the comprehensive ability of different heterogeneous features to respond to disaster risk but also provides a stable input basis for further weighting. By constructing a feature decision sequence with a single sorting direction, the accuracy and controllability of feature weight evaluation can be effectively improved, and the uncertainty of the model in the weight allocation process can be reduced.

[0036] By assigning feature importance based on the sequential index position in the feature decision sequence, the position with the first sequential index is assigned the maximum weight value in the preset weight domain, the position with the last sequential index is assigned the minimum weight value in the preset weight domain, and features in the middle index positions are assigned weight values ​​at equal intervals between the maximum and minimum weight values ​​according to the index order, so that the importance of all features has a directly comparable numerical expression and can be used for subsequent prediction and judgment.

[0037] To ensure that each heterogeneous feature participating in the hierarchical prediction has a quantifiable weight in the subsequent risk prediction process, weights need to be assigned based on its sequential index position in the feature decision sequence. First, a numerical range is predefined to represent the feature weights, i.e., the predefined weight domain. This range has upper and lower limits, typically set as a closed interval, such as from 0.1 to 1.0, to ensure the allocation result is within the bounded number domain. The heterogeneous feature corresponding to the first position in the feature decision sequence is assigned the maximum weight value in the predefined weight domain, representing the highest decision participation of this feature in the prediction judgment; the feature corresponding to the last position is assigned the minimum weight value in the predefined weight domain, representing the lowest reference value of this feature in the prediction. All features located at intermediate index positions are assigned equally spaced weight values ​​between the maximum and minimum weight values ​​according to their ranking position in the feature decision sequence, based on the quotient of the total span of the weight domain and the total number of participating features. For example, when there are five features and the predefined weight domain is from 0.1 to 1.0, the weights can be divided into a set of equally spaced values ​​and assigned to each feature sequentially. This index-based mapping method enables feature importance to be expressed in a directly comparable numerical form, without relying on the model's internal self-learning structure. This helps to establish explicit weight guidance in subsequent fusion processing and risk assessment, thereby enhancing the dominance and interpretability of the prediction.

[0038] S3. Based on the feature importance of each heterogeneous feature, perform expression tensor mapping on the original sequences of each type of heterogeneous feature, and project each heterogeneous feature to the corresponding channel of the expression tensor according to the feature importance, so as to generate a fusion expression tensor in which the expression channel depth is related to the feature importance. In this embodiment, S3 specifically refers to: Establish a correspondence between the original sequences of various heterogeneous features and their corresponding feature importance values. By traversing the feature importance values ​​of each type of heterogeneous feature, construct a feature importance ranking table according to the values ​​from high to low, and allocate channel index positions in the expression tensor based on the feature importance ranking table to form a mapping rule between feature importance and expression tensor channels. When establishing the correspondence between the original sequences of various heterogeneous features and their corresponding feature importance, the normalized value of the feature importance can be used as the sorting criterion to rank all heterogeneous features from largest to smallest, generating a feature importance ranking table. For example, among the four features of gas disturbance, roadway settlement, image recognition results, and personnel work trajectories, if their corresponding feature importance values ​​are 0.92, 0.75, 0.65, and 0.41 respectively, the ranking results are gas disturbance, roadway settlement, image recognition results, and personnel work trajectories. Subsequently, the channel index positions of the representation tensor are assigned according to the ranking results, with the channel numbers corresponding to the feature importance from highest to lowest. For example, channel 0 is assigned to gas disturbance, channel 1 to roadway settlement, and so on. This process constructs a clear mapping rule, ensuring that each type of heterogeneous feature has a fixed and weight-aware channel position in the representation tensor, avoiding disordered feature order, and enabling the subsequent prediction process to explicitly identify the importance hierarchy of features.

[0039] The original sequences of various heterogeneous features refer to multiple types of data record sequences collected over a unified time dimension. Each type of sequence corresponds to a data source with different physical meanings, such as changes in gas concentration values ​​or target label streams output from image recognition. The feature importance ranking table is an ordered table arranged from largest to smallest feature decision scale value, reflecting the strength of each type of heterogeneous feature's influence on disaster risk prediction. The channel index position in the expression tensor represents the dimensional position of each independent feature sequence within the tensor structure; the channel arrangement directly affects the interpretation priority of feature representations. The mapping rule between feature importance and expression tensor channels is a one-to-one correspondence, ensuring that different features occupy distinct spatial positions in the expression tensor according to their importance, thus enabling the model to differentiate the information contributions of different features when receiving input. The construction of this mapping rule provides the model with structured priors, helping to improve the expression strength and prediction weights of key features.

[0040] According to the mapping rules, the original sequences of each type of heterogeneous features are projected to the specified channel positions of the expression tensor according to their feature importance. During the projection process, the original sequences are normalized, and the normalized sequence values ​​are filled into the time series dimension of the corresponding channel in the expression tensor according to the channel index to form a primary tensor channel set with feature importance expression features. In the implementation process, the channel number corresponding to each type of heterogeneous feature can be determined first by using the established mapping rules between feature importance and expression tensor channels. For example, the gas disturbance with the highest feature importance is assigned to channel 0, and the second most important feature, roadway settlement, is assigned to channel 1. Then, the original sequences of each type of heterogeneous feature are normalized to ensure their values ​​are distributed within a uniform range; for example, the gas disturbance concentration sequence is normalized to between 0 and 1. Next, based on the channel number, these normalized sequence values ​​are sequentially filled into the corresponding channel positions in the expression tensor according to the time dimension. The core of this process is to ensure that important features occupy clear positions in the tensor structure and to present their dynamic changes on a standardized scale, thereby improving their recognition and participation in the model processing stage. In this way, a tensor structure with preliminary feature discrimination capabilities is formed, providing a data foundation for subsequent deep fusion.

[0041] The designated channel position in the representation tensor refers to the fixed-dimensional channel number assigned to each type of heterogeneous feature in the multidimensional data tensor. This is used to separate information from features from different sources in the high-dimensional structure, avoiding feature mixing. Each channel position maps to only one heterogeneous feature, ensuring the consistency and traceability of the tensor structure. The primary tensor channel set, which expresses feature importance, is a set of tensors filled with several normalized feature sequences according to specific channel indices. This set not only preserves the original temporal information of various heterogeneous features but also presents the weight level of each feature in the prediction judgment through channel distribution. As an intermediate representation before fusion, the primary tensor channel set plays a crucial role in establishing a preliminary perceptual structure of feature importance for the prediction model, which helps improve the model's ability to focus on key disaster information and its response efficiency.

[0042] In the primary tensor channel set, all filled channels are hierarchically concatenated along the channel dimension of the expression tensor according to the order of feature importance, so that the channel arrangement structure is consistent with the feature importance, and a fused expression tensor with the expression channel depth correlated with the feature importance is generated, which can be directly called as the input data source for subsequent hierarchical prediction paths.

[0043] To construct a fusion representation tensor that correlates channel depth with feature importance, the channel hierarchy must first be reorganized within the primary tensor channel set. Specifically, this involves sequentially concatenating all filled feature channels along the channel dimension of the representation tensor, in descending order of feature importance. For example, the gas disturbance channel, ranked first in importance, is placed at the beginning, followed by the roadway subsidence channel, ranked second, and so on, down to the least important feature channel. During the concatenation process, the original time series data order within each channel is not altered; only the channel's position within the tensor structure is adjusted. This operation allows the channel index itself to reflect the influence of features in the hierarchical judgment, enabling the prediction model to prioritize channels with higher information weights in subsequent calculations, thereby improving the accuracy and response speed of disaster risk identification.

[0044] Hierarchical concatenation refers to a structural rearrangement process along the channel dimension of an expression tensor. Its goal is to linearly combine different heterogeneous features according to their importance into a channel set with sequential information. Unlike ordinary concatenation, hierarchical concatenation emphasizes the semantic carrying of order. A fusion expression tensor where channel depth is correlated with feature importance means that the order of each channel in the channel dimension is determined by feature importance; the depth order indicates the priority of risk assessment. This fusion expression tensor not only preserves the original dynamic information of each heterogeneous feature but also embeds the importance of features into the data source in a structured form. This allows subsequent hierarchical prediction paths to directly perceive the weight differences of each feature, enabling the model to focus on key information calculations and accurately characterize risk boundaries.

[0045] S4. By guiding the fusion expression tensor into the hierarchical prediction path, setting a decision priority focusing channel, prioritizing the activation of heterogeneous feature channels whose feature importance ranking is at the front index position of the feature decision sequence, and constructing a focusing path based on feature importance control, a risk classification judgment process dominated by features is established. In this embodiment, S4 specifically refers to: By inputting the fusion expression tensor into the input node of the hierarchical prediction path according to the time series dimension, the channel dimension structure of the fusion expression tensor is parsed in the input node, and the position identification of all channels of the fusion expression tensor is performed according to the index order in the feature decision sequence. The decision priority focusing channel attribute is marked on the channel whose channel index position is at the beginning of the feature decision sequence, so as to complete the guidance of the fusion expression tensor to the decision priority focusing channel in the hierarchical prediction path. Inputting the fused representation tensor into the input node of the hierarchical prediction path along the time-series dimension means that the tensor structure must first be maintained in a two-dimensional time-channel arrangement, carrying the representation channels of all heterogeneous features within each time slice. As the starting point of the hierarchical prediction path, the input node needs channel resolution capabilities, which can be achieved by setting a nested index scanning mechanism to deconstruct and analyze the channel dimension of the fused representation tensor. By retrieving the index order position of each heterogeneous feature in the feature decision sequence, the structure number of each channel in the fused representation tensor is traversed, and its order mapping relationship in the feature decision sequence is established. When the index order of a certain channel is at the beginning of the feature decision sequence, its corresponding channel is marked as the decision priority focusing channel, for example, by setting a priority focusing channel attribute identifier and recording it in the path guidance table. The purpose of this operation is to ensure that subsequent hierarchical prediction paths can prioritize the representation channels with higher feature importance, thereby improving the sensitivity and accuracy of the model in the early identification and decision-making hierarchical stages. In actual implementation, a loop structure combined with Boolean logic flags can be used to classify channels according to predetermined thresholds and generate an input mapping table to control the model's guidance behavior.

[0046] The input nodes of the hierarchical prediction path are the starting units in the model structure that receive external tensor inputs and expand them into internal computation modules. They are typically implemented as structural layers with multi-channel parallel receiving capabilities, used to structurally deconstruct the channel dimensions of the expressed tensors. The feature decision sequence is a feature index table formed by sorting the feature decision scale values ​​of each heterogeneous feature. It defines the priority order in which each channel should be processed in the prediction path. Channels at the beginning of the feature decision sequence refer to those with higher index numbers in the sequence; their corresponding heterogeneous features have higher feature decision scale values, and therefore need to obtain higher priority in path execution. The index positions at the beginning of the feature decision sequence are used to precisely define which channels are in the priority activation range. Generally, the first N positions are set as high-weight channel intervals, and channels at these positions are assigned the attribute of decision priority focus channels, used for subsequent construction of the focus path and control of channel activation behavior. This structural setting ensures that the entire prediction path can guide resources to focus on key features in the initial stage, improving the response efficiency and judgment accuracy to changes in risk classification.

[0047] In the decision priority focusing channel, channel activation weight enhancement parameters are set, and according to the correspondence between channel index and feature importance, the amplified expression tensor value is injected into the channel to enhance the channel activation intensity. This activation enhancement behavior is recorded in the channel weight control matrix, so that heterogeneous feature channels with feature importance ranking at the front index position of the feature decision sequence are preferentially activated in the hierarchical prediction path. Setting channel activation weight enhancement parameters in the decision-priority focusing channels means purposefully increasing the weight of the channel activation signals. This can be achieved by weighting and amplifying the values ​​of the channels marked as priority focusing channels in the fusion representation tensor. The weighting process introduces an adjustable gain factor before the channel activation function, which is automatically set based on the channel's index position in the feature decision sequence. For example, multiplying the top three channels by a factor greater than 1 makes their activation values ​​more likely to reach the trigger condition during model computation. Subsequently, the system writes the mapping relationship between channel index and feature importance into the channel weight control matrix and records the specific value of the amplification factor to ensure that the model can trace and reuse this activation enhancement behavior during training or inference. This mechanism can be implemented by constructing weighted layers or weight adjustment modules to enhance the representational strength of signal channels, improve the expressive power of high-importance features in the model, and thus guide the model to focus more attention on features that are more sensitive to changes in risk level.

[0048] Channel activation weight enhancement parameters are numerical factors used to weight the expression tensor values ​​within specific channels, making these channels more easily triggered during the activation phase of the neural network. This parameter can be a constant, a learnable weight, or a dynamic function based on the feature decision sequence index. In practical applications, it is used to adjust the response intensity of high-weight channels. The channel weight control matrix is ​​a structured data record table used to maintain the weight adjustment state of each expression channel, recording information including channel index, activation enhancement factor, and feature importance value. This matrix can be accessed by the model at different training stages, ensuring that each iteration follows a consistent activation enhancement mechanism and avoiding instability in prediction results due to activation differences. The channel weight control matrix can also provide a basis for dynamic adjustments to the model in different scenarios, such as reconfiguring the activation strategy after data drift or feature weight reassessment, further enhancing the model's controllability and interpretability regarding feature-dominated paths.

[0049] By reading the activation order in the channel weight control matrix, a focusing path controlled by feature importance is constructed in the channel dimension. This focusing path is then used as the input index control sequence of the feedforward layer in the hierarchical prediction path, ensuring that the response order of the prediction path is consistent with the feature decision sequence. This establishes a risk classification judgment process dominated by feature importance.

[0050] By reading the activation order in the channel weight control matrix, the priority of each expression channel in the prediction path can be clearly defined, thereby constructing a focused path for feature importance control. In implementation, the system extracts the channel indices and their corresponding feature importance order recorded in the channel weight control matrix, forming an ordered sequence of channel activation indices. This index sequence serves as the priority processing order of input channels in each layer of the feedforward network structure, and is called the input index control sequence of the feedforward layer. When the model performs a hierarchical prediction task, the feedforward layer reads the data of the corresponding channels sequentially according to this sequence, ensuring that the response order of the prediction path remains strictly consistent with the feature decision sequence. For example, if the feature importance of three channels is 0.9, 0.6, and 0.3, the corresponding index control sequence will guide the feedforward layer to process the channel with an importance of 0.9 first, and then process the remaining channels in sequence. This approach can introduce an index guidance mechanism into the input layer or intermediate fusion layer of the neural network, ensuring that the model prioritizes the input channels most sensitive to changes in risk level during feature processing. Through this path guidance mechanism, the model's efficiency in focusing on dominant features is improved, thereby optimizing the accuracy and stability of hierarchical judgment.

[0051] The feature importance-controlled focusing path is a channel sequence structure built along the channel dimension of the representation tensor. Its core purpose is to prioritize channels with high feature importance in the processing flow through a ranking mechanism, thereby enhancing the model's responsiveness to key features. The input index control sequence of the feedforward layer in the hierarchical prediction path refers to the specific index queue that controls the input order of each feedforward processing unit in the channel dimension. This sequence directly originates from the feature importance ranking result in the channel weight control matrix. Maintaining consistency between the response order of the prediction path and the feature decision sequence means that the model strictly follows the feature importance ranking during the inference phase, processing channels sequentially to avoid information confusion or priority bias, and ensuring semantic consistency between feature representation and model response. The risk grading judgment process, dominated by feature importance, is a model inference mechanism built around the strength of features in risk evolution. Essentially, it uses structural design to extend the ranking advantage of features at the input end to the prediction output, enhancing the hierarchical model's ability to identify and respond to high-risk signals.

[0052] S5. After obtaining the hierarchical prediction results, construct a feature contribution offset map based on the prediction output offset, and adjust the channel activation ratio in the fusion expression tensor in combination with the feature importance to dynamically adjust the influence weight of each heterogeneous feature in the hierarchical prediction process.

[0053] In this embodiment, S5 specifically refers to: After obtaining the hierarchical prediction results, the numerical offset between the predicted output and the true label at each time index position is extracted. This numerical offset is then mapped to the channel corresponding to each heterogeneous feature participating in the hierarchical prediction according to the channel structure of the fusion expression tensor. By statistically analyzing the mean and variance of the predicted output offset on each channel, a feature contribution offset map is constructed to express the degree of influence of each heterogeneous feature on the current predicted output offset. After obtaining the tiered prediction results, the model's predicted output needs to be compared point-by-point with the true labels under the corresponding time index, calculating the numerical offset between the two at each time index position. This numerical offset can be obtained through simple numerical difference calculation and is used to measure the model's accuracy in predicting risk levels at different time points. Subsequently, based on the channel structure of the fusion representation tensor, the heterogeneous features corresponding to each channel are associated with the prediction offset at that time index. By statistically analyzing the mean and variance of the prediction offset for each channel throughout the entire time period, it is possible to identify which channels exhibit more significant prediction errors, thereby constructing a feature contribution offset map. This map, with channels as the dimension and the statistical characteristics of prediction errors as the data source, reflects the potential impact of different heterogeneous features on the model's prediction bias. This process can be achieved through a time sliding window and channel attribution mapping, providing a basis for subsequent channel activation adjustments and facilitating a feedback optimization mechanism based on actual output performance.

[0054] The numerical offset between the predicted output and the true label at each time index refers to the numerical difference between the predicted output value and the corresponding true label at each time point predicted by the model. This difference characterizes the degree of deviation in the prediction; it is a fundamental parameter for evaluating prediction accuracy. The feature contribution offset map is a graphical representation structure constructed by mapping the prediction offset to channels in the fused representation tensor and performing statistical analysis. It is used to quantify the impact of heterogeneous features carried by different channels on prediction bias. This map considers not only the average intensity of the offset but also the stability of offset fluctuations. The mean and variance together reflect the relative weight of each feature in error formation, serving as the core data foundation for subsequent activation adjustment and dynamic weight allocation.

[0055] Based on the contribution intensity of each channel in the feature contribution offset map and the feature importance of the corresponding channel in the fusion representation tensor, a weighted fusion calculation is performed to generate a channel activation ratio adjustment factor table, and the channel activation ratio update magnitude of each channel is determined based on the channel activation ratio adjustment factor table. After constructing the feature contribution offset map, the contribution intensity of each channel in the map can be further analyzed jointly with the feature importance of the corresponding channel in the fused expression tensor. Specifically, firstly, the offset statistics of each channel in the feature contribution offset map are extracted, such as the mean of the prediction error, as the contribution intensity. Then, the feature importance corresponding to that channel is extracted from the feature decision sequence. The two are then weighted and fused according to a preset ratio to generate a channel activation ratio adjustment factor for each channel. The weighted fusion can be performed using a linear combination method, dynamically adjusting the activation ratio in the expression tensor based on the influence of different channels on the error and their position in the overall decision. Subsequently, these adjustment factors are used to construct a complete channel activation ratio adjustment factor table to guide whether each channel needs to have its expression intensity enhanced or suppressed in subsequent predictions. This method achieves reverse adjustment of channel weights from the prediction offset results, thereby optimizing the overall expressive power of the prediction path and improving the model's accuracy in identifying risk levels.

[0056] The contribution intensity of each channel in the feature contribution offset map refers to a numerical metric obtained through statistical analysis, used to measure the role of each channel in the prediction result offset. It is typically expressed based on the mean or variance of the prediction error and measures the degree of influence of each channel on the prediction error. Weighted fusion calculation refers to integrating multi-dimensional information using mathematical weighting when considering multiple influencing factors (such as contribution intensity and feature importance), thereby deriving a comprehensive index for subsequent adjustments. In this scenario, the output of the weighted fusion calculation is the channel activation ratio adjustment factor, used to indicate how the current channel should adjust its activation weight in the expression tensor in the next prediction round. The channel activation ratio adjustment factor table is an ordered data structure composed of adjustment factors corresponding to all channels, providing clear guidance for updating the channel activation ratio, enabling the model's expression tensor to dynamically adapt to the actual prediction error.

[0057] In the channel dimension of the fusion expression tensor, the original activation ratio is adjusted by gain or suppression according to the update magnitude of the channel activation ratio, and the adjusted fusion expression tensor is re-input into the hierarchical prediction path to realize the dynamic adjustment of the influence weight of each heterogeneous feature participating in hierarchical prediction in the hierarchical prediction process.

[0058] In the channel dimension of the fused expression tensor, to dynamically optimize the weights of the heterogeneous features participating in the hierarchical prediction, the original activation ratio of each channel can be updated based on the previously generated channel activation ratio adjustment factor table. Specifically, all channels in the fused expression tensor are traversed, and their update magnitude values ​​in the adjustment factor table are compared. Multiplicative gain or attenuation methods are used to perform numerical scaling on the activation value of the current channel, thereby enhancing or weakening the channel's influence on the final prediction result. The adjusted activation values ​​remain within a uniform numerical domain, ensuring that the continuity of the data structure is not disrupted. Subsequently, the adjusted fused expression tensor is reintroduced as input into the hierarchical prediction path, allowing the model to dynamically optimize the channel expression based on the prediction error of the current round, thus forming a closed-loop updateable prediction control mechanism. Through this feedback-driven adjustment strategy, the model can more accurately identify key features and suppress interfering features, thereby improving its ability to judge the evolution trend of disaster risk levels.

[0059] The fusion expression tensor is a multidimensional data structure formed by mapping multiple heterogeneous feature sequences to different channels and then combining them. The channel dimension represents each heterogeneous feature participating in hierarchical prediction. The channel activation ratio refers to the relative strength of each channel's contribution to the overall expression in the expression tensor. The update magnitude originates from the previously constructed channel activation ratio adjustment factor, which guides whether to enhance or weaken the participation of each channel. Gain or suppression adjustment refers to multiplying the original channel activation value by a scaling factor greater than or less than 1 to amplify or reduce the expression intensity of the channel, thereby achieving dynamic control of the weight level. Re-inputting the adjusted fusion expression tensor into the hierarchical prediction path means that the model input structure is updated. This structure carries a new feature expression weight configuration, thereby driving the prediction path to focus on the more influential features in subsequent judgments, establishing a linkage mechanism between feedback adjustment and prediction judgment.

[0060] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0061] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0062] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0063] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0066] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks, characterized in that, Specifically, the following steps are included: S1. By constructing a feature decision response comparison set, response trajectories are extracted from various heterogeneous features participating in graded prediction in multiple labeled typical coal mine disaster samples. Based on the calculation results of the trigger offset degree of disaster risk level change nodes, it is determined whether there are differences in decision importance among the heterogeneous features participating in graded prediction. S2. In the case of determining that there are differences in decision importance, calculate the feature decision scale value of each heterogeneous feature participating in the hierarchical prediction based on the time precedence, triggering intensity and risk boundary penetration capability in the response trajectory, so as to determine the feature importance of each heterogeneous feature. S3. Based on the feature importance of each heterogeneous feature, perform expression tensor mapping on the original sequences of each type of heterogeneous feature, and project each heterogeneous feature to the corresponding channel of the expression tensor according to the feature importance, so as to generate a fusion expression tensor in which the expression channel depth is related to the feature importance. S4. By guiding the fusion expression tensor into the hierarchical prediction path, setting a decision priority focusing channel, prioritizing the activation of heterogeneous feature channels whose feature importance ranking is at the front index position of the feature decision sequence, and constructing a focusing path based on feature importance control, a risk classification judgment process dominated by features is established. S5. After obtaining the hierarchical prediction results, construct a feature contribution offset map based on the prediction output offset, and adjust the channel activation ratio in the fusion expression tensor in combination with the feature importance to dynamically adjust the influence weight of each heterogeneous feature in the hierarchical prediction process.

2. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 1, characterized in that, S1 specifically includes the following steps: S101. By organizing the time series records of multiple labeled typical coal mine disaster samples, the original data of heterogeneous features participating in the graded prediction at each disaster stage are aligned with a unified time benchmark, and a feature decision response comparison set containing time markers, feature value sequences and risk level markers is constructed. S102. Based on the feature decision response comparison set, identify the nodes of disaster risk level change in the time series, and calculate the slope of numerical change, starting point position and change magnitude of various heterogeneous features participating in the graded prediction before and after the change nodes. After calculation, generate the trigger offset degree of each type of heterogeneous feature, and extract the response trajectory accordingly. S103. By normalizing the trigger offset degree corresponding to various heterogeneous features participating in hierarchical prediction and calculating the difference between each pair, when the difference between the normalized trigger offset degree of any two heterogeneous features exceeds the preset judgment threshold, it is determined that there is a difference in decision importance among the heterogeneous features participating in hierarchical prediction.

3. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 2, characterized in that, S102 specifically refers to: By arranging the time series in the characteristic decision response comparison set according to the risk level label order, the jump points of the risk level label are used to determine the disaster risk level change nodes, and the time index corresponding to the disaster risk level change node is used as the benchmark to extract continuous raw data segments from the fixed time span before and after the time index to form the pre-change sequence and post-change sequence for calculation. In the pre-change and post-change sequences, continuous data of heterogeneous features participating in hierarchical prediction are calculated respectively. The slope of numerical change is determined by the rate of change of the difference between adjacent data points in the pre-change and post-change sequences. The starting point position is determined by the time point when the continuous data first produces an offset. The magnitude of change is determined by the maximum difference range between the pre-change and post-change sequences. The set of input parameters for triggering the offset degree is constructed using three continuously changing parameters. The trigger offset degree is generated by combining the set of input parameters in a preset order and mapping them to a unified numerical domain. The trigger offset degree is then arranged along the time index to form a response trajectory, which is used as the input for subsequent processing.

4. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 1, characterized in that, S2 specifically includes the following steps: S201. When it is determined that there are differences in decision importance, the response trajectory of each heterogeneous feature participating in the graded prediction is analyzed by time index analysis to identify the time position in the response trajectory where the degree of trigger offset first continuously increases. The time index corresponding to the time position is calculated with the time index of the disaster risk level change node. The difference is used as the time precursor to indicate the degree of advance of the response of the heterogeneous feature in the risk evolution process. S202. Determine the maximum value of the trigger offset in the response trajectory, and use the maximum value as the trigger intensity. Obtain the risk level mark corresponding to the occurrence of the value, and count the value range of the risk level mark in each continuous time period before and after the risk level mark. If the value range spans at least two different risk level intervals, the heterogeneous feature has the ability to penetrate risk boundaries, which is used to reflect the depth of its influence on the multi-level risk boundary. S203. Convert time-leading characteristics, triggering intensity, and risk boundary penetration capability into feature decision scale values ​​according to preset mapping rules. Normalize and sort all feature decision scale values, and determine the feature importance of each heterogeneous feature participating in hierarchical prediction based on the sorting results.

5. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 4, characterized in that, S203 specifically refers to: By combining time-leadingness, triggering intensity, and risk boundary penetration capability into a three-dimensional input vector according to a preset order, and performing linear scaling on each component of the three-dimensional input vector according to a preset mapping rule, the three-dimensional input vector forms a mapping result in a unified numerical domain. This mapping result serves as the feature decision scale value for each heterogeneous feature participating in hierarchical prediction. By performing normalization on all feature decision scale values, all feature decision scale values ​​fall within a set continuous closed interval. Based on the normalized feature decision scale values, a feature decision sequence is generated from the position of maximum value to the position of minimum value, so that the feature decision sequence forms a sequential index with a single sorting direction. By assigning feature importance based on sequential index position in the feature decision sequence, the position with the first sequential index is assigned the maximum weight value in the preset weight domain, the position with the last sequential index is assigned the minimum weight value in the preset weight domain, and features in the middle index position are assigned equal-interval weight values ​​between the maximum and minimum weight values ​​according to the index order.

6. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 1, characterized in that, S3 specifically refers to: Establish a correspondence between the original sequences of various heterogeneous features and their corresponding feature importance values. By traversing the feature importance values ​​of each type of heterogeneous feature, construct a feature importance ranking table according to the values ​​from high to low, and allocate channel index positions in the expression tensor based on the feature importance ranking table to form a mapping rule between feature importance and expression tensor channels. According to the mapping rules, the original sequences of each type of heterogeneous features are projected to the specified channel positions of the expression tensor according to their feature importance. During the projection process, the original sequences are normalized, and the normalized sequence values ​​are filled into the time series dimension of the corresponding channel in the expression tensor according to the channel index to form a primary tensor channel set with feature importance expression features. In the primary tensor channel set, hierarchical splicing is performed on all filled channels along the channel dimension of the expression tensor in order of feature importance, so that the channel arrangement structure is consistent with the feature importance, and a fused expression tensor with the expression channel depth correlated with the feature importance is generated.

7. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 1, characterized in that, S4 specifically refers to: By inputting the fusion expression tensor into the input node of the hierarchical prediction path according to the time series dimension, the channel dimension structure of the fusion expression tensor is parsed in the input node, and the position identification of all channels of the fusion expression tensor is performed according to the index order in the feature decision sequence. The decision priority focusing channel attribute is marked on the channel whose channel index position is at the beginning of the feature decision sequence, so as to complete the guidance of the fusion expression tensor to the decision priority focusing channel in the hierarchical prediction path. In the decision priority focusing channel, channel activation weight enhancement parameters are set, and according to the correspondence between channel index and feature importance, the amplified expression tensor value is injected into the channel to enhance the channel activation intensity. This activation enhancement behavior is recorded in the channel weight control matrix, so that heterogeneous feature channels with feature importance ranking at the front index position of the feature decision sequence are preferentially activated in the hierarchical prediction path. By reading the activation order in the channel weight control matrix, a focusing path controlled by feature importance is constructed in the channel dimension. This focusing path is then used as the input index control sequence of the feedforward layer in the hierarchical prediction path, ensuring that the response order of the prediction path is consistent with the feature decision sequence. This establishes a risk classification judgment process dominated by feature importance.

8. The method for collaborative fusion and hierarchical prediction of heterogeneous information on typical coal mine disaster risks according to claim 1, characterized in that, S5 specifically refers to: After obtaining the hierarchical prediction results, the numerical offset between the predicted output and the true label at each time index position is extracted. This numerical offset is then mapped to the channel corresponding to each heterogeneous feature participating in the hierarchical prediction according to the channel structure of the fusion expression tensor. By statistically analyzing the mean and variance of the predicted output offset on each channel, a feature contribution offset map is constructed to express the degree of influence of each heterogeneous feature on the current predicted output offset. Based on the contribution intensity of each channel in the feature contribution offset map and the feature importance of the corresponding channel in the fusion representation tensor, a weighted fusion calculation is performed to generate a channel activation ratio adjustment factor table, and the channel activation ratio update magnitude of each channel is determined based on the channel activation ratio adjustment factor table. In the channel dimension of the fusion expression tensor, the original activation ratio is adjusted by gain or suppression according to the update magnitude of the channel activation ratio, and the adjusted fusion expression tensor is re-input into the hierarchical prediction path to realize the dynamic adjustment of the influence weight of each heterogeneous feature participating in hierarchical prediction in the hierarchical prediction process.

Images