An intelligent risk early warning method and system based on multi-source data fusion

Abstract

The application discloses a kind of intelligent risk early warning method and system based on multi-source data fusion, belong to construction risk assessment and early warning technical field, the method is: obtaining multi-source information, obtains early warning control variable based on multi-source information and preset quantization processing mode;Early warning control variable and preset grading early warning standard are used to build initial early warning cloud picture;Based on initial early warning cloud picture, two-dimensional matrix distribution and preset two-dimensional normal cloud model obtain three-dimensional early warning cloud picture;Improved prediction model is obtained by using posterior probability support vector method to optimize and correct three-dimensional early warning cloud picture;Risk early warning result is obtained based on improved evidence fusion method and improved prediction model.The application provides a kind of intelligent risk early warning method and system based on multi-source data fusion, realizes to fuse multi-source data as the basis to combine historical data using posterior probability support vector method and improved evidence fusion method to build corrected three-dimensional early warning model, significantly improve the accuracy, effectiveness and applicability of risk early warning.

Description

Technical Field

[0001] This invention relates to the field of construction risk assessment and early warning technology, and in particular to an intelligent risk early warning method and system based on multi-source data fusion. Background Technology

[0002] During subway foundation pit construction, the project typically involves deep excavation and complex support structures, presenting numerous potential risks. Therefore, early warning of foundation pit construction risks is a crucial aspect of risk management. It promptly alerts decision-makers to take appropriate safety measures, effectively preventing accidents such as foundation pit collapse and support structure failure. Early warning mechanisms are the foundation of risk warning research, with early warning standards serving as a vital basis for assessing the safety status of foundation pit projects. By establishing reasonable early warning standards, key parameters such as deformation and displacement of the foundation pit can be monitored in real time. Once these parameters approach or reach the warning standards, an alarm can be issued promptly, ensuring the safety of foundation pit construction. In engineering practice, national standards, local standards, or empirical knowledge are the most common and widely used forms of standard construction. However, due to their strong subjectivity and regional bias, these standards often influence early warning judgments, leading to results that do not match reality.

[0003] Existing risk early warning methods for subway foundation pit construction suffer from the following problems: early warning standards lack uniformity and adaptability; the same monitoring data from different projects may correspond to different early warning levels; early warning methods based on single monitoring items cannot fully reflect the complex and ever-changing conditions at the construction site, easily leading to erroneous judgments; traditional DS evidence theory is prone to producing results contradicting the facts when dealing with highly conflicting evidence; while risk early warning based on a single monitoring method, a single monitoring item, or a single monitoring point is a common approach, it often results in false alarms and missed alarms in engineering practice, causing significant resource waste and potential safety hazards. These problems lead to insufficient accuracy and reliability of existing early warning methods, failing to provide effective protection for the safety of subway foundation pit construction. Therefore, the application of multi-source data fusion in foundation pit engineering is crucial, and there is an urgent need to provide a risk early warning method for subway foundation pit construction based on multi-source data fusion to solve the above-mentioned technical problems and provide effective and accurate early warning assessment data for the actual construction process. Summary of the Invention

[0004] This invention provides an intelligent risk early warning method and system based on multi-source data fusion. It can solve the technical problems of existing early warning methods based on single monitoring items, which cannot fully reflect the complex and ever-changing situation at the construction site and are prone to erroneous judgments. It also addresses the technical problems of traditional DS evidence theory, which is prone to producing results that contradict the facts when dealing with highly conflicting evidence. The invention achieves the construction of a modified three-dimensional early warning model based on the fusion of multi-source data and historical data, using the posterior probability support vector method and an improved evidence fusion method. This significantly improves the accuracy, effectiveness, and applicability of risk early warning.

[0005] This invention provides an intelligent risk early warning method based on multi-source data fusion, comprising:

[0006] Acquire multi-source information, and obtain early warning control quantities based on multi-source information and preset quantization processing methods;

[0007] An initial early warning cloud map is constructed based on the early warning control quantity and the preset graded early warning standards;

[0008] A three-dimensional early warning cloud map is obtained based on the initial early warning cloud map, two-dimensional matrix distribution, and a preset two-dimensional normal cloud model;

[0009] The posterior probability support vector method is used to optimize and correct the three-dimensional early warning cloud map to obtain an improved prediction model;

[0010] Risk warning results are obtained based on improved evidence fusion methods and improved prediction models.

[0011] This invention provides an intelligent risk early warning method based on multi-source data fusion. It constructs a two-dimensional initial early warning cloud map based on fused multi-source data to obtain relevant risk early warning data after fusion. Then, a three-dimensional early warning cloud map is constructed. Historical data is combined with posterior probability support vector method and improved evidence fusion method to correct the ambiguity and randomness in the transition area between adjacent two-dimensional early warning levels. An improved integration and fusion method is used to obtain the risk change trend and early warning status of foundation pit construction, improving the accuracy and effectiveness of risk early warning. Furthermore, a more unified early warning prediction standard is established, enhancing the adaptability and uniformity of intelligent risk early warning.

[0012] Furthermore, an initial early warning cloud map is constructed based on the early warning control quantity and the preset graded early warning standard, including: obtaining the early warning risk level based on the early warning control quantity and the preset graded early warning standard; obtaining the cumulative change value and change rate value based on the early warning risk level, and constructing a two-dimensional matrix distribution based on the cumulative change value and change rate value; and constructing the initial early warning cloud map using the traditional empirical method based on the two-dimensional matrix distribution.

[0013] The above scheme defines the early warning control quantity to make each monitoring item dimensionless, thereby constructing a preliminary early warning distribution prediction model that can be applied to different application situations. It also uses traditional empirical methods to visualize multi-source data and their corresponding early warning risk levels, thus preparing initial data for the subsequent construction of the three-dimensional model.

[0014] Furthermore, a three-dimensional early warning cloud map is obtained based on the initial early warning cloud map and a preset two-dimensional normal cloud model, including: obtaining a set of early warning state intervals based on the initial early warning cloud map; obtaining cloud digital features and early warning clouds based on the set of early warning state intervals and the preset two-dimensional normal cloud model; and constructing a three-dimensional early warning cloud map based on the cloud digital features and early warning clouds according to a preset construction method.

[0015] The above scheme describes the fuzziness and randomness of the transition area of ​​the initial early warning cloud map by constructing a three-dimensional early warning cloud map. Each early warning cloud has its corresponding membership degree, which can well express the duality of the early warning level transition. It realizes the idea of ​​solving the problem of fuzziness and randomness in the transition area between adjacent early warning levels based on the cloud model.

[0016] Furthermore, the posterior probability support vector method is used to optimize and correct the 3D early warning cloud map to obtain an improved prediction model. This includes: constructing a training set based on historical risk warning data, and constructing a target separating hyperplane based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set; using the posterior probability support vector method to obtain a target decision surface based on the positive and negative sample sets, and obtaining an optimized posterior model based on a preset activation function and the target decision surface; obtaining historical cumulative change values ​​and historical change rates based on historical risk warning data; obtaining historical cloud membership degrees based on historical cumulative change values, historical change rates, and the 3D early warning cloud map; obtaining a target sample set based on historical cumulative change values, historical change rates, and historical cloud membership degrees, and obtaining a posterior probability distribution based on the target sample set and the optimized posterior model; and optimizing and correcting the 3D early warning cloud map based on the posterior probability distribution and preset posterior probability principles to obtain an improved prediction model.

[0017] The above scheme adopts an improved support vector method with strong adaptability to learn early warning cases of subway foundation pit projects in different regions, corrects the results of the empirical knowledge classification, and obtains a more objective early warning level partition. This solves the problem that the early warning levels corresponding to each block in the graded early warning are not universal, and establishes a more unified early warning level classification standard.

[0018] Furthermore, a training set is constructed based on historical risk warning data, and a target separating hyperplane is constructed based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set. This includes: constructing a training set based on historical risk warning data; obtaining a minority feature set based on the training set, and obtaining nearest neighbor data based on the minority feature set and a preset similarity calculation method; obtaining an optimized training set based on the nearest neighbor data and a preset synthetic minority class oversampling method; and constructing a target separating hyperplane based on the optimized training set and a preset classification rule to obtain a positive sample set and a negative sample set.

[0019] In the above scheme, the optimal separating hyperplane for machine learning is constructed through the training set to correctly classify the samples into positive and negative classes according to the classification rules, and place them on both sides of the hyperplane respectively, thus solving the problems of small sample data and nonlinear samples. At the same time, the SMOTE method is used to optimize the posterior probability support vector method, so that the number of samples of each warning level in the warning cases reaches a balanced level.

[0020] Furthermore, the three-dimensional early warning cloud map is optimized and corrected based on the posterior probability distribution and the preset posterior probability principle to obtain an improved prediction model, including: constructing an initial prediction model based on the three-dimensional early warning cloud map; obtaining a training prediction model based on the initial prediction model using a preset hierarchical cross-validation method and a preset network search method; and optimizing and correcting the prediction model based on the posterior probability distribution, the preset posterior probability principle, and the training prediction model to obtain an improved prediction model.

[0021] In the above scheme, the three-dimensional early warning cloud map is corrected and updated based on the objective records of historical cases of foundation pit risk early warning during construction. By using the support vector posterior probability output theory and based on engineering statistical cases, the cloud membership degree of each region to different early warning levels is calculated. The posterior probability support vector method is used to generate the partition correction results, which solves the problem that the early warning levels corresponding to each block in the graded early warning are not universal. A more unified early warning level classification standard is established to provide more accurate and effective risk early warning results.

[0022] Furthermore, risk warning results are obtained based on improved evidence fusion methods and improved prediction models, including: obtaining a membership matrix set based on the improved prediction model; constructing an initial confidence level by introducing global uncertainty based on the membership matrix set; obtaining a two-dimensional distribution of the initial confidence level based on the dual control indicators and the initial confidence level; obtaining first-type high-conflict evidence and second-type high-conflict evidence based on the two-dimensional distribution of the initial confidence level; obtaining a first confidence level based on traditional evidence theory and first-type high-conflict evidence; and obtaining a target confidence level based on the first confidence level, second-type high-conflict evidence, and improved evidence fusion methods, so as to obtain risk warning results based on the target confidence level.

[0023] The above scheme is based on the ideas of model-based early warning and quantitative analysis. It constructs an intelligent early warning model for construction risks based on multi-source monitoring projects, and adopts an improved evidence fusion method to effectively integrate monitoring data from different information sources to obtain more accurate early warning results. It introduces global uncertainty to construct an initial confidence level (BPA) to form evidence, which objectively and effectively represents uncertainty information and realizes the universality of the construction method. It reduces conflicts and the impact of unreliable evidence on the synthesis results, and improves the rationality and accuracy of the final prediction results. This provides strong support for the research on risk early warning of complex multi-attribute decision-making problems such as subway foundation pit collapse.

[0024] Furthermore, the first level of trust is obtained based on traditional evidence theory and the first type of high-conflict evidence, including: obtaining the degree of conflict and the degree of difference based on traditional evidence theory and the first type of high-conflict evidence; obtaining the improved conflict factor based on the degree of conflict and the degree of difference according to the preset conflict improvement method, and obtaining the first level of trust based on the improved conflict factor and the initial level of trust.

[0025] In the above scheme, two representative indicators, conflict degree α and difference degree β, are selected for joint measurement, and a new comprehensive conflict factor is constructed by combining two-dimensional coordinate mapping. This effectively solves the first type of conflict problem, takes the conflict factor more comprehensively, and effectively weakens the differences between evidence bodies.

[0026] Furthermore, a target trust level is obtained based on a first trust level, a second type of high-conflict evidence, and an improved evidence fusion method, in order to obtain risk warning results based on the target trust level. This includes: obtaining a comprehensive conflict factor based on the first trust level; obtaining dynamic high-conflict and dynamic low-conflict based on a preset conflict perception threshold and the comprehensive conflict factor; performing a weighted average based on dynamic high-conflict and a preset weighted average rule to obtain first corrected data; obtaining second corrected data based on dynamic low-conflict and the improved evidence fusion method; and obtaining a target trust level based on the first and second corrected data, in order to obtain risk warning results based on the target trust level.

[0027] The above scheme addresses the problem of invalid high-conflict evidence caused by defects in the fusion rules by using a hybrid fusion rule based on improved DS and weighted average. It sets thresholds for different cases and improves the synthesis rules by introducing weighted average, discount factor and other methods, thereby improving the rationality and accuracy of the synthesis results.

[0028] In the above scheme, deep learning prediction data from multiple sources are used as evidence for risk warning judgment. A preliminary judgment of each evidence is obtained by constructing a two-dimensional initial trust level (BPA). Then, the high conflict of the evidence itself is adjusted by the first type of conflict adjustment, and the failure of the fusion rule is corrected by the second type of conflict correction. Finally, by sensing and fusing the final trust of each information source at four warning levels, the final warning level of the subway foundation pit risk is obtained by applying the principle of maximum trust. The two types of conflict adjustment greatly improve the effective fusion of multi-source data and the accuracy of risk warning.

[0029] This invention provides an intelligent risk early warning method based on multi-source data fusion. It constructs a two-dimensional initial early warning cloud map based on fused multi-source data to obtain relevant risk early warning data after fusion, and then constructs a three-dimensional early warning cloud map. Historical data is combined with posterior probability support vector method and improved evidence fusion method to correct the ambiguity and randomness in the transition area between adjacent two-dimensional early warning levels. Based on two-dimensional cloud BPA theory and DS evidence theory, an improved two-dimensional BPA-DS multi-source information fusion method for foundation pit construction risk early warning is proposed. This method effectively constructs initial confidence levels, effectively fuses highly conflicting multi-source information, and obtains the risk change trend and early warning status of foundation pit construction through improved evidence fusion method, thereby improving the accuracy and effectiveness of risk early warning. Simultaneously, a relatively unified early warning prediction standard is established, enhancing the adaptability and uniformity of intelligent risk early warning, and providing strong support for risk early warning research on the complex multi-attribute decision problem of subway foundation pit construction collapse.

[0030] This invention also provides an intelligent risk early warning system based on multi-source data fusion, used to implement the aforementioned intelligent risk early warning method based on multi-source data fusion, comprising: a monitoring module, a prediction module, and a correction module, wherein:

[0031] The monitoring module is used to acquire multi-source information and obtain early warning control quantities based on the multi-source information and preset quantization processing methods;

[0032] The prediction module is used to construct an initial early warning cloud map based on the early warning control quantity and the preset graded early warning standards;

[0033] The correction module is used to optimize and correct the three-dimensional early warning cloud map using the posterior probability support vector method to obtain an improved prediction model, and to obtain risk warning results based on the improved evidence fusion method and the improved prediction model.

[0034] This invention provides an intelligent risk early warning system based on multi-source data fusion. First, a monitoring module acquires multi-source information and performs data preprocessing. Then, a prediction module constructs a two-dimensional initial early warning cloud map based on the fused multi-source data to obtain relevant risk early warning data after fusion. Subsequently, a three-dimensional early warning cloud map is constructed. The correction module combines historical data and uses the posterior probability support vector method and an improved evidence fusion method to correct the ambiguity and randomness in the transition area between adjacent two-dimensional early warning levels. An improved integration and fusion method is used to obtain the risk change trend of foundation pit construction and its early warning status, thereby improving the accuracy and effectiveness of risk early warning. Furthermore, a more unified early warning prediction standard is established to enhance the adaptability and uniformity of intelligent risk early warning. Attached Figure Description

[0035] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of a smart risk warning method based on multi-source data fusion provided in this embodiment;

[0037] Figure 2 This is a schematic diagram showing the causes and proportions of foundation pit collapse provided in this embodiment;

[0038] Figure 3 This is a schematic diagram of a two-dimensional matrix partitioning diagram for risk classification and early warning based on experience knowledge provided in this embodiment;

[0039] Figure 4 This is a schematic diagram of the three-dimensional early warning cloud map construction process provided in this embodiment;

[0040] Figure 5 This is a schematic diagram of the three-dimensional early warning cloud map provided in this embodiment;

[0041] Figure 6 This is a schematic diagram of the classification hyperplane of the support vector method provided in this embodiment;

[0042] Figure 7 This is a schematic diagram of a three-dimensional early warning cloud map correction process provided in this embodiment;

[0043] Figure 8 This is a schematic diagram of a new sample synthesis process based on Smote provided in this embodiment;

[0044] Figure 9 This is a schematic diagram of a data augmentation process based on SMOTE provided in this embodiment;

[0045] Figure 10 This embodiment provides a schematic diagram of a data change process based on an improved cloud membership-support vector model.

[0046] Figure 11 This is a schematic diagram of the sample membership distribution provided in this embodiment;

[0047] Figure 12 This is a schematic diagram of the mesh search parameter optimization process when the kernel is poly, as provided in this embodiment;

[0048] Figure 13 This is a schematic diagram of the grid search parameter optimization process when the kernel is rbf, as provided in this embodiment;

[0049] Figure 14 This is a schematic diagram of the grid search parameter optimization process when the kernel is sigmoid, as provided in this embodiment;

[0050] Figure 15 This is a schematic diagram of the two-dimensional initial trust level (BPA) distribution provided in this embodiment;

[0051] Figure 16 This is a schematic diagram of the improved conflict factor provided in this embodiment;

[0052] Figure 17 This embodiment provides a schematic diagram of a risk warning process based on an improved two-dimensional BPA-DS.

[0053] Figure 18 This is a schematic diagram of the location of the research survey area provided in this embodiment;

[0054] Figure 19 This is a schematic diagram of an intelligent risk warning model based on multi-source data fusion provided in this embodiment. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] Example 1:

[0057] Please see Figure 1 This embodiment provides an intelligent risk early warning method based on multi-source data fusion, including:

[0058] S1. Obtain multi-source information, and obtain early warning control quantities based on multi-source information and preset quantization processing methods;

[0059] S2. Construct an initial early warning cloud map based on early warning control quantities and preset graded early warning standards;

[0060] S3. Obtain a three-dimensional early warning cloud map based on the initial early warning cloud map, two-dimensional matrix distribution, and a preset two-dimensional normal cloud model;

[0061] S4. The posterior probability support vector method is used to optimize and correct the three-dimensional early warning cloud map to obtain an improved prediction model.

[0062] S5. Obtain risk warning results based on improved evidence fusion methods and improved prediction models.

[0063] This embodiment provides an intelligent risk warning method based on multi-source data fusion. It constructs a two-dimensional initial warning cloud map based on the fusion of multi-source data to obtain relevant risk warning data after fusion. Then, a three-dimensional warning cloud map is constructed, and it is corrected by combining historical data with the posterior probability support vector method and the improved evidence fusion method to solve the ambiguity and randomness problems in the transition area between two-dimensional adjacent warning levels, enhance its adaptability and uniformity, establish a more unified warning level classification standard, and improve the accuracy and effectiveness of risk warning.

[0064] In the specific implementation process, in order to analyze and summarize the risk factors of foundation pit collapse, reveal the relationship between the frequency of occurrence of different factors and collapse, and provide an objective basis for subsequent research on risk warning indicator system, this embodiment of the invention first statistically analyzed 78 cases of subway foundation pit collapse. The collected collapse case data mainly came from the China Emergency Information Website, relevant literature records, news reports, and engineering projects in which the author participated. Some typical cases are shown in Table 1.1 Typical Cases of Subway Foundation Pit Collapse Accidents.

[0065] Table 1.1 Typical Cases of Subway Foundation Pit Collapse Accidents

[0066]

[0067]

[0068] Case studies revealed that the causes of foundation pit collapses during actual construction are multifaceted, primarily categorized into two main types, as shown in Table 1.2, "Classification of Risk Factors for Subway Foundation Pits." One type is subjective factors, including design, construction, and management factors. For example, excessively deep foundation pits may increase lateral earth pressure, while excessively steep slopes may reduce soil stability; insufficient depth of support piles and weak connections in the support structure can lead to support failure; and lax quality control by the supervision unit can result in quality hazards in key components such as the support structure and drainage facilities. The other type is objective factors, including geological and natural factors. For instance, if the foundation pit is located in a high-water-bearing area, long-term water accumulation and soil saturation can significantly reduce the soil's shear strength; natural disasters such as earthquakes, typhoons, and rainstorms can cause soil structure damage and support structure failure.

[0069] Table 1.2 Classification of Risk Factors in Subway Foundation Pit

[0070]

[0071]

[0072] In practice, although obtaining data on subway foundation pit collapses is difficult and its completeness cannot be guaranteed, analyzing and summarizing a large amount of data can yield reliable empirical data that reflects the basic patterns of foundation pit collapses and has certain reference value. This embodiment analyzes several historical cases, and the distribution of the causes of foundation pit collapses, i.e., the proportion of risk factors, is as follows: Figure 2 As shown in the figure, statistical results indicate that non-standard construction operations and lax construction process management are the most frequent risk factors, accounting for 62.8% and 57.6% respectively. Following these are complex geological structures, the impact of construction activities, inadequate drainage measures, high water-bearing areas, lax construction quality control, and non-standard support structure construction, each accounting for 35% or more. Finally, improper selection of support structures, active fault zones, natural disasters, and improper design of foundation pit dimensions are all below 10%. Furthermore, calculating the sum of the proportions of each factor reveals that foundation pit collapse is a result of multiple factors, usually reflected in on-site monitoring information. Therefore, during construction, on-site monitoring efforts and information processing capabilities should be strengthened. By understanding current monitoring information and future trends, a more comprehensive understanding of foundation pit changes can be achieved, thereby minimizing the possibility of foundation pit collapse accidents. Considering that single monitoring data cannot fully perceive risk information, and that different monitoring methods are used for different data, resulting in multi-source heterogeneous data, commonly used data monitoring methods include instrument monitoring, drone aerial photography, and video recognition. This embodiment studies and calculates the risk early warning system for subway foundation pit construction from the perspective of multi-source information fusion. The advantages of multi-source information fusion are: by integrating monitoring data obtained through multiple methods, it can more comprehensively reflect the current state and changing trends of the foundation pit, thereby improving the accuracy of the early warning; at the same time, multi-source information can mutually verify and supplement each other, reducing the errors and blind spots that may exist in a single information source, and enhancing the reliability of the early warning. This chapter will then focus on how to construct a risk early warning index system and standards for subway foundation pit construction based on multi-source information, providing a direction and foundation for further in-depth research on early warning methods.

[0073] Optionally, an initial early warning cloud map is constructed based on the early warning control quantity and the preset graded early warning standard, including: obtaining the early warning risk level based on the early warning control quantity and the preset graded early warning standard; obtaining the cumulative change value and change rate value based on the early warning risk level, and constructing a two-dimensional matrix distribution based on the cumulative change value and change rate value; and constructing the initial early warning cloud map using the traditional empirical method based on the two-dimensional matrix distribution.

[0074] In the specific implementation process, to achieve uniformity in early warning levels from different information sources, and considering the trend consistency between early warning results and actual monitoring values ​​(the ratio of actual monitoring values ​​to control values), this paper defines K as the ratio of actual monitoring values ​​to control values, i.e., K is used as the early warning control quantity. K1 and K2 are used to represent the cumulative change value and rate of change value of each risk factor, respectively. Dimensionless representation is applied to each monitoring item, eliminating the need to consider the specific numerical value of the control value. This allows the same two-dimensional cloud model to be used regardless of the number of monitoring items selected, avoiding the inefficiency caused by repeatedly constructing different cloud models. The specific early warning indicators for foundation pit construction risks and their corresponding K values ​​are shown in Table 1.3 below.

[0075] Table 1.3 Risk Warning Indicators and Corresponding K Values ​​for Foundation Pit Construction

[0076]

[0077]

[0078] Control values ​​determined based on monitoring standards, references, and actual engineering projects are crucial for assessing the safety of engineering structures and their surrounding environment. In this embodiment, 70%, 85%, and 100% of the control values ​​are used as the grading criteria according to a preset hierarchical early warning standard. For ease of subsequent modeling, the right boundary value is set to 1.7. Therefore, the early warning risk levels are divided into four levels, from lowest to highest: Level 1 (blue), Level 2 (yellow), Level 3 (orange), and Level 4 (red), as shown in Table 1.3: Level 1 (blue) represents a safe state, requiring continued basic safety protection measures; Level 2 (yellow) represents a monitoring state, requiring strengthened daily inspections and maintenance to ensure normal operation of equipment and facilities; Level 3 (orange) represents an alarm state, requiring sufficient attention and corresponding control measures; Level 4 (red) represents a dangerous state, with extremely high risk, requiring immediate action and activation of the highest-level emergency response plan. To facilitate problem analysis, the hierarchical early warning standards described in Table 1.4 are visualized. Using the ratio of cumulative change value to corresponding control value (K1) as the horizontal axis and the ratio of rate of change value to corresponding control value (K2) as the vertical axis, after analysis and judgment based on the graded early warning standards, the early warning results are distributed into a two-dimensional matrix, as shown below. Figure 3 As shown, when constructing the initial early warning cloud map based on the two-dimensional matrix distribution using the traditional empirical method, the initial early warning cloud map of the two-dimensional cloud model is mapped into the two-dimensional space according to the two-dimensional matrix distribution and the traditional empirical method, that is, the two-dimensional cloud model, to lay the foundation for the subsequent construction of the three-dimensional distribution model.

[0079] Table 1.4 Risk Classification and Early Warning Standards for Foundation Pit Construction Based on Empirical Knowledge

[0080]

[0081] Optionally, a three-dimensional early warning cloud map is obtained based on the initial early warning cloud map and a preset two-dimensional normal cloud model, including: obtaining a set of early warning state intervals based on the initial early warning cloud map; obtaining cloud digital features and early warning clouds based on the set of early warning state intervals and the preset two-dimensional normal cloud model; and constructing a three-dimensional early warning cloud map based on the cloud digital features and early warning clouds according to a preset construction method.

[0082] In the specific implementation process, when constructing a 3D early warning cloud map, it is necessary to address the issues of ambiguity and randomness in the transition areas between adjacent early warning levels, as well as the lack of universality in the early warning levels corresponding to different areas in the map. The specific steps are as follows:

[0083] (1) Based on the actual situation, the ratios K1 and K2 are divided into several warning levels, and each level corresponds to an interval number [L]. min L max ];

[0084] (2) For the warning levels of ratios K1 and K2, each interval number is selected and transformed into cloud digital features (Ex, En, He) using a preset two-dimensional normal cloud model formula. The cloud model can be described based on the cloud digital features, namely, expectation Ex, entropy En, and hyperentropy He. Expectation Ex is the mean of cloud droplets, entropy En is the deviation from expectation Ex, reflecting the ambiguity of information; He is the entropy of En, reflecting the randomness of information and affecting the degree of cloud droplet aggregation. The two-dimensional cloud model can be described using two sets of cloud digital features (Ex, En). x He x ) and (Ey, En y He y This can be described using the concept of a two-dimensional random function, capable of simultaneously expressing the randomness and ambiguity arising from the dual control indicators of cumulative value and rate of change. Assume F is a two-dimensional random function following a normal distribution, with expectation (Ex, Ey) and expectation (En). x En y ) is the standard deviation, generate random numbers (x) i ,y i ); with (En x En y ) is the expectation, (He x He y ) is the standard deviation, generate random numbers (Px) i Py i The cloud droplet Drop(x) satisfies the preset two-dimensional normal cloud model formula. i y i m i The cloud model constructed from these elements is called a two-dimensional normal cloud model, and its expression is as follows:

[0085]

[0086] In the formula, (x i y i ) represents the cloud droplet coordinates, m i Let m be the membership degree. To obtain the membership degree m of the target cloud... i First, it is necessary to calculate the two-dimensional cloud numerical characteristics (Ex, En) corresponding to the cumulative value and rate of change, which are the two control indicators. x He x ) and (Ey, En y He y In the construction risk early warning indicator system, the monitoring indicators used to determine the early warning level are denoted as Ti (i = 1, 2, 3, ..., 8), where T1 to T8 correspond to groundwater level, pile horizontal displacement, column settlement, ground surface settlement, support axial force, foundation pit convergence, wall top horizontal displacement, and wall top vertical displacement, respectively. Each monitoring indicator Ti corresponds to a pair of dual control indicators, and the corresponding early warning state is denoted as T. i (k j ), where j = 1, 2. Each warning state can correspond to a specific double-limit interval, denoted as [t ij (L), t ij (R)], the bilimit interval is converted into a two-dimensional cloud model (Ex, En) by the following formula. x He x ) and (Ey, En y He y ):

[0087]

[0088] In the formula, the constant He takes values ​​in the range of [0, En], which is used to describe the uncertainty of the current index; the cloud digital characteristics of the dual control index of this study are calculated from Table 1.4 and are shown in Table 1.5, which solves the uncertainty in the transition region.

[0089] Table 1.5 Cloud Digital Characteristics of Dual Control Indicators

[0090]

[0091] (3) Cloud droplets are generated based on a preset two-dimensional normal cloud model formula according to the two-dimensional matrix distribution, thus generating an early warning cloud; to address the issue of ambiguity and randomness in the early warning level zoning based on empirical knowledge not considering the transition region, this embodiment uses a two-dimensional cloud model to characterize the early warning zoning, such as... Figure 4 As shown in (a), the x-axis represents the ratio K1 of the cumulative change value to the corresponding control value, the y-axis represents the ratio K2 of the rate of change value to the corresponding control value, and the z-axis represents the membership degree. Each region in the matrix corresponds to a two-dimensional cloud, which is called an early warning cloud. Figure 4 As shown in (b), once all regions correspond to early warning clouds, a three-dimensional early warning cloud map based on the cloud model can be obtained;

[0092] (4) Repeat steps (2) and (3) to represent all regions in the two-dimensional matrix diagram as early warning clouds, thus forming a three-dimensional early warning cloud map. The vertical axis of the cloud droplets represents the membership degree. The cloud droplets overlap between different early warning clouds, reflecting the fuzziness and randomness of the transition area of ​​the early warning zone. With the ratio K1 as the x-axis, the ratio K2 as the y-axis, and the membership degree of the two-dimensional cloud as the z-axis, the three-dimensional early warning cloud map of the foundation pit construction risk is finally obtained, as follows: Figure 5 As shown.

[0093] Optionally, the 3D early warning cloud map is optimized and corrected using the posterior probability support vector method to obtain an improved prediction model. This includes: constructing a training set based on historical risk warning data, and constructing a target separating hyperplane based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set; obtaining a target decision surface based on the positive and negative sample sets using the posterior probability support vector method, and obtaining an optimized posterior model based on a preset activation function and the target decision surface; obtaining historical cumulative change values ​​and historical change rates based on historical risk warning data; obtaining historical cloud membership degrees based on historical cumulative change values, historical change rates, and the 3D early warning cloud map; obtaining a target sample set based on historical cumulative change values, historical change rates, and historical cloud membership degrees, and obtaining a posterior probability distribution based on the target sample set and the optimized posterior model; and optimizing and correcting the 3D early warning cloud map based on the posterior probability distribution and preset posterior probability principles to obtain an improved prediction model.

[0094] In practical implementation, Support Vector Machine (SVM) is a machine learning method jointly proposed by Vapnik and Corinna Cortes, which excels at solving problems with small sample data and non-linear samples. For linear binary classification tasks, its training samples are represented as (x... i y i ), where x i ∈R n y n ∈{-1, 1}, i = 1, 2, ..., n. The goal of learning is to construct an optimal separating hyperplane, which aims to correctly classify samples into positive and negative classes according to the classification rules and place them on either side of the hyperplane. For the same classification task, there are often multiple different decision boundaries, such as... Figure 6 As shown, the squares and triangles represent the two classes of samples, respectively, and the dashed lines represent different decision planes, namely H1:w. T x+b=-1 and H2:w T x+b=+1 is used to separate the two types of samples. Figure 6 (a) represents multiple classification hyperplanes. Figure 6(b) represents the maximum margin principle: when there exists a decision plane H such that its distances to H1 and H2 are equal, then the optimal decision plane H can be obtained. Where w = (w1, w2, ..., w n ), where b is the normal vector of plane H; and d is the displacement, representing the distance between plane H and the origin (0, 0). The distance between H1 and H2 represents the classification margin d = 2 / ||w||. 2 Solving for the target decision surface means finding a w that maximizes d. Based on the above analysis, the calculation process for finding the optimal hyperplane, i.e., the target decision surface, is described as follows:

[0095] The function of the optimal decision surface H is:

[0096] f(x) = w T x+b;

[0097] Training samples to the optimal decision surface H: The distance is:

[0098]

[0099] Training sample set to the optimal decision surface H: The distance is:

[0100]

[0101] In the formula, S represents the sample space;

[0102] At this point, the optimization objective is:

[0103]

[0104] make Therefore, the optimization objective becomes:

[0105]

[0106] To simplify the differentiation process, the above equation is now transformed into a problem of maximizing the classification margin, i.e., finding... The maximum value is equivalent to finding The minimum value of . Therefore, we need to find a set (w, b) that satisfies the following for all training samples:

[0107]

[0108] The optimization objective now becomes:

[0109]

[0110] Next, by constructing the Lagrange function, the above problem is further transformed into:

[0111]

[0112] In the formula, ai are constrained Lagrange multipliers, and all of them are non-negative;

[0113] Calculate the partial derivatives of w and b in the above equation:

[0114]

[0115] This leads to the dual expression:

[0116]

[0117] According to the duality principle, the process of minimizing the original objective function J(w, b, a) can be transformed into maximizing the corresponding Lagrange dual problem, that is:

[0118]

[0119] Therefore, the above optimization problem is equivalent to solving a convex quadratic programming optimization problem, that is, there exists a global optimal solution. Therefore, b* can be calculated:

[0120]

[0121] Therefore, the optimal decision plane can be described as:

[0122]

[0123] In the formula, each component ai* of the dual variable a* has a strict one-to-one mapping with the training samples. Its multiplier sparsity property makes most ai* approach zero. According to the KKT complementary relaxation condition, the samples corresponding to the non-zero multipliers are the support vectors, which constitute the geometric features of the decision boundary.

[0124] Then, by using the Sigmoid activation function to map the hard output f(x) of the support vectors to the range [0, 1], the posterior probability output of the support vectors can be obtained:

[0125]

[0126] In the formula, P(y=1|f) represents the category y=1. The posterior probability, where A and B are parameters, can be obtained by solving the minimum negative log-likelihood of the parameter set as follows;

[0127]

[0128] In the formula, pi represents p(yi=1|xi), N+ is the number of samples with yi=1, and N- is the number of samples with yi=-1.

[0129] Simultaneously, in practical applications, the 3D early warning cloud map is corrected and updated based on objective records of foundation pit risk early warning cases during construction. Using support vector posterior probability output theory and based on engineering statistical cases, the cloud membership degree of each region to different early warning levels is calculated. The posterior probability support vector method is then used to generate partition correction results. The engineering data processing flow is as follows: Figure 7 As shown, the red text highlights the posterior probability distribution, which can be summarized in the following steps:

[0130] (1) Statistically analyze the cumulative changes, change rates, and corresponding warning levels of foundation pit construction monitoring projects in historical risk warnings. Using a cloud model (2dcloud), convert the ratio of cumulative changes K1 and the ratio of change rates K2 into membership degrees (u1, u2, ..., um). A sample data set {(u1, u2, ..., um), (Ln, Level)} is formed by the membership degrees and the corresponding warning levels. Through training and learning with a large amount of sample data, the optimal posterior probability support vector model for foundation pit risk warning zoning correction can be obtained; (2) Calculate the membership degrees of the regional center points {(u1, 0, ..., 0), (u2, ..., um ... 0, ..., 0), ..., (um, 0, ..., 0)} is fed into the optimal model to obtain the posterior probability distribution {(p11, p12, ..., p1m), (p21, p22, ..., p2m), ... (pm1, pm2, ..., pmm)}. The posterior probability distribution is used as the probability that each region belongs to one of the four warning levels, reflecting the membership degree of each region to the four warning levels in the warning cloud map; (3) According to the principle of maximum posterior probability, SVM is used to correct the warning cloud map, and the level corresponding to each region is (Ln, Level). Among them, Ln is the partition number in the warning cloud map, n = 1-1, 1-2, ..., 3-4, 4; Level is the warning level after Ln correction. In this study, blue, yellow, orange and red are taken.

[0131] Optionally, a training set is constructed based on historical risk warning data, and a target separating hyperplane is constructed based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set, including: constructing a training set based on historical risk warning data; obtaining a minority feature set based on the training set, and obtaining nearest neighbor data based on the minority feature set and a preset similarity calculation method; obtaining an optimized training set based on the nearest neighbor data and a preset synthetic minority class oversampling method; and constructing a target separating hyperplane based on the optimized training set and a preset classification rule to obtain a positive sample set and a negative sample set.

[0132] In practical implementation, when machine learning methods handle classification problems, the smaller the difference in the number of positive and negative samples in the dataset, the better the classification performance. When the proportion of negative samples in an imbalanced dataset is small, the learning process may fail to fully learn the information of negative samples due to the small number of negative samples. The resulting model may learn positive sample features well, but its learning performance is poor for a small number of negative samples with low probability of occurrence, significantly reducing the prediction accuracy for negative samples and even leading to the extreme case of predicting all positive samples. Therefore, this embodiment employs the Synthetic Minority Over-sampling Technique (SMOTE) to achieve inter-class balance optimization through intelligent analysis of the neighborhood relationships of minority class samples, thereby augmenting the training set. The SMOTE method is used to optimize the posterior probability support vector method, ensuring that the number of samples at each warning level in the warning cases reaches a balanced level. The SMOTE method uses each minority class sample as the center point, randomly searches for the remaining minority class samples, forms k nearest neighbors, connects them to the center sample, and performs interpolation to finally form new minority class samples. This method can reduce the impact of data imbalance on inaccurate data classification. The process is described as follows: (1) Assume there are M minority class samples in total, with N feature dimensions, and their feature values ​​are represented as (x1, x2, ..., x...). N For each sample in the minority class set, calculate its similarity to all other samples in the minority class set. Assume the feature values ​​of the i-th and j-th samples in the minority class set are {x...} i :(x i 1,x i 2,......,x i N )} and {x j :(x j 1,x j 2,......,x j N The similarity between the two is dist(x) i ,x j Select the k samples that are closest to i to identify the k nearest neighbors in the feature space. The similarity calculation formula is as follows:

[0133]

[0134] (2) Based on the imbalance of the actual samples, a sampling ratio is set. For each sample xi of the minority class, several samples xt are randomly selected from its k nearest neighbors, and new samples xnew are generated by combining them with the original sample xi according to the following formula:

[0135] xnew =x i +rand(0,1)·(x t -x i )

[0136] If the value of k is too small, the synthesized effect will be similar to that of data generated by random oversampling; if the value of k is too large, it will increase the execution cost and the probability of introducing noise. The value of k is generally set to 5. Figure 8 As shown, k=5 in the k-nearest neighbors, X1 is a minority class sample point, and the five nearest neighbor sample points found are X1, X2, X3, X4, X5, X6, X7, X8, X9, X1, X1, X1, X1, X2 ... 11 ,X 12 ,X 13 ,X 14 and X 15 In X1 and X 11 In the interpolation, dist is the distance between two sample points, and the newly generated sample point x new On the connecting line, rand_dist is from X1 to X. 11 The random distance between them is described by rand(0,1). The target sample, i.e., the new sample, can be described as x. new =X1 + rand_dist * dist. Since the more balanced the data, the better the classification effect, synthesizing samples in a 1:1 ratio yields the best results. This embodiment is based on a real-world engineering early warning case study. The ratio of the collected early warning cases is 66:42:31:19, corresponding to the number of cases for blue, yellow, orange, and red early warnings, respectively. The process is summarized as follows: Engineering data is processed through a cloud model to obtain cloud membership degrees. This step only transforms the membership degrees of the original monitoring data and does not change the number of samples. When the data is augmented using the SMOTE method, the ratio of early warning cases is adjusted to 66:66:66:66, thus achieving a balanced number of samples for each early warning level, which can better adapt to subsequent classification prediction. The support vector posterior probability output model after adding the SMOTE method is as follows. Figure 9 As shown, the red highlighted text indicates the changes in data volume for each warning level before and after data enhancement.

[0137] Optionally, the 3D early warning cloud map is optimized and corrected based on the posterior probability distribution and a preset posterior probability principle to obtain an improved prediction model, including: constructing an initial prediction model based on the 3D early warning cloud map; obtaining a training prediction model based on the initial prediction model using a preset hierarchical cross-validation method and a preset network search method; and optimizing and correcting the posterior probability distribution, the preset posterior probability principle, and the training prediction model to obtain an improved prediction model.

[0138] To address the lack of universality in early warning level zoning during implementation, a posterior probability support vector method with self-learning capabilities is adopted. Based on actual engineering case data, the traditional empirical zoning is optimized and improved. To address the sample imbalance problem inherent in the dataset, the SMOTE method is introduced to balance the current sample categories. Therefore, this embodiment proposes an improved cloud membership-support vector model based on a two-dimensional cloud model, posterior probability support vector method, and SMOTE method. Its architecture is as follows: Figure 10 As shown, the model can optimize and improve the warning level partitions obtained through traditional experience knowledge, so that the risk warning standard has strong adaptability and universality. The specific implementation process is described as follows: (1) Collect actual engineering cases and build a dataset: The actual engineering cases collected in this paper total 218 cases, including warning data of subway foundation pit projects in multiple regions. In order to evaluate the performance of the model on unknown data and ensure that the model does not overfit during training, the dataset is often divided into training set and test set according to the ratio of 70%:30%. Therefore, 60 cases are selected as test set (15 cases per level), which are not used for training and are only used for evaluating the generalization ability of the model. The remaining 158 cases are used as training set. The initial input features include two dimensions (x, y) of cumulative value and rate of change, which are converted to (K1, K2) by ratio, and the output features are (Ln, Level). Ln represents the region where the partition number n is located, and the value range is 1-1, 1-2, ..., 3-4, 4. Level is the warning level of the region where Ln is located, and the value is blue, yellow, orange and red. (2) Solving the uncertainty problem in the transition area of ​​the warning level: Calculating cloud membership. As shown in process (1), the current input feature is (K1, K2). In the dataset of this paper, there are 218 such features. Now, it is necessary to uniformly convert the input features of these 218 samples into cloud membership and calculate the membership degree (u) of (K1, K2) relative to all partitions Ln. n 1, u n 2, ..., u n m ), where u n m represents the cloud membership degree of the current monitoring data relative to partition Ln; since there are 16 clouds in the 3D early warning cloud map, m is set to 16, thus calculating the membership degree of each monitoring data point to the 16 early warning clouds. Traditional early warning partitioning directly determines the early warning level by locating the position of (K1, K2), but due to the membership degree (u n 1, u n 2, ..., u n m() is no longer a fixed value, but a probability distribution that describes the degree of belonging of (K1, K2) in the entire warning zone. This description method can flexibly represent and process fuzzy information. (3) Solve the problem of imbalanced samples in the engineering dataset and perform data augmentation: For the 158 samples in the training set, the number of warning cases in the four categories of blue, yellow, orange and red are 66, 42, 31 and 19 respectively. Obviously, the number of samples in each category is not balanced. According to the distribution of the nearest neighbor samples of the minority class samples, SMOTE data augmentation is used to generate new samples to improve the training and evaluation of the cloud membership-support vector model. Assume that the sample set with a large number of blue warnings is M, and the sample sets with a relatively small number of yellow, orange and red warnings are N1, N2 and N3 respectively. For each sample in N1, N2 and N3, the nearest neighbor point when k=5 is obtained to form three sample sets Q1, Q2 and Q3 respectively. For each sample in Q1, Q2 and Q3, a new sample is generated using the SMOTE algorithm until the data volume of these three warning levels is 66. Thus, the number of warning samples in the four categories of blue, yellow, orange and red is the same. (4) Model training and evaluation to obtain the optimal model: In this embodiment, hierarchical cross-validation and grid search techniques are used to optimize the model parameters. Hierarchical 5-fold cross-validation can increase the diversity of model training samples when the number of cases is limited. At the same time, this method can make the proportion of the target variable in each Fold the same as the proportion of the whole dataset, so that each fold can represent the whole well. The grid search method is used to optimize the parameters of "kernel function, gamma parameter and penalty parameter C". The optimization ranges are [linear, poly, rbf], [0.01, 10] and [0.001, 10], respectively. The search step size of gamma and C is 1. In order to obtain a better warning partition correction effect, the model test accuracy is required to be no less than 95%. (5) Early warning level zoning correction to solve the problem of incompatibility of early warning level zoning: In the three-dimensional early warning cloud map, the center point of each cloud represents the expected value of the early warning level of that area. The 16 center point positions are converted into cloud membership degrees through the cloud model, represented as {(u1, 0, ..., 0), (0, u2, ..., 0), ..., (0, 0, ... ... 16 )}, and use it as input sample, and send it to the optimal model obtained in step (4) for posterior probability estimation, so that the posterior probability distribution of each region for the four warning levels can be obtained, which is expressed as {(p 1 1, p 1 2, ..., p 1 m ), (p 2 1, p 2 2, ..., p 2m ), ......, (p 16 1, p 16 2, ..., p 16 m This reflects the probability distribution of each warning cloud at different levels, and then the warning classification results (Ln, Level) after correction of 16 cloud areas can be obtained, thus solving the problem that the construction risk warning level zoning is not universal for different subway foundation pit projects.

[0139] In the specific implementation process, the engineering case data collected in this embodiment has strong authenticity and reliability. The dataset is divided into three categories, each with inputs and outputs. The first category is the training set, used for model training; the second category is the test set, used for evaluating the model's generalization ability; and the third category is the input dataset used for early warning zoning correction. The initial inputs are the cumulative value x and the rate of change y of each monitoring item, and the final outputs are the zoning number Ln corresponding to the maximum membership degree and the early warning level Level. To reduce the impact of uncertainty in the transition area, the ratio K needs to be converted into membership degree. The model is trained to learn the fuzzy and random information of the transition area. For the above three categories of datasets, the input features of the samples need to be converted into membership degrees, while the output features do not need to be converted and are still described using the early warning level. To clearly describe the membership degree, a radar chart can be used to draw the membership degree distribution. The membership degree distribution chart of some samples in the training set is shown below. Figure 11 As shown in (a), the membership distribution of some samples in the test set is as follows: Figure 11 As shown in (b), the membership distribution map of the partial sample cloud in the early warning zoning correction sample set is as follows. Figure 11 As shown in (c), the figure displays the membership degree distribution of 30 sample groups. The outer coordinates 1, 2, 3, ..., 16 represent the 16 warning zones from 1 to 16. The radius coordinates 0, 0.2, 0.4, 0.6, 0.8, 1.0, 1.2 represent the membership degree within the current zone. The values ​​of the 16 zones along the radius together constitute the membership degree of a given sample. All points of the same color represent the membership degrees (u1, u2, ..., u...) generated by a sample group (K1, K2). 16 Different colors represent the membership degrees generated by different samples.

[0140] In the specific implementation process, this embodiment compares the traditional support vector model and the improved cloud membership-support vector model, allowing C and gamma to be calculated within a certain range (in this study, the search range for C and gamma is between 0.01 and 10, and the search step size is 1). Then, cross-validation is used to find the C and gamma combination with the highest accuracy. Considering that different C and gamma combinations may correspond to the highest accuracy, and that an excessively high penalty parameter C can cause overlearning, thus worsening the model's generalization ability, the C and gamma combination corresponding to the minimum C value is considered the optimal combination. Three combinations can achieve the highest accuracy of 82.74%, namely ① kernel = poly; C = 0.001; gamma = 8.001; (e.g. Figure 12 (As shown) ②kernel=rbf;C=1.001;gamma=1.001 (as shown) Figure 13 (as shown) and ③ kernel = sigmoid; C = 1.001; gamma = 1.001 (as shown) Figure 14 As shown). Figure 12 , 13 As shown in Figure 14, kernel functions selected include poly, rbf, and sigmoid. To clearly visualize the change process, a contour plot is used. In the 3D graph, the X-axis is C, the Y-axis is gamma, and the Z-axis is accuracy. The graph uses color changes to describe accuracy changes; the closer the color is to red, the higher the accuracy, and the closer it is to blue, the lower the accuracy. Specifically: Figure 12 (a) shows the improved cloud membership-support vector optimization process when the kernel is poly, where the optimal point is: C = 1.001, gamma = 2.001, accuracy = 94.94%. Figure 12 (b) shows the traditional support vector optimization process when the kernel is poly, where the optimal point is: C = 0.001, gamma = 8.001, accuracy = 82.74%; Figure 13 (a) shows the improved cloud membership-support vector optimization process when the kernel is rbf, where the optimal point is: C = 4.001, gamma = 5.001, accuracy = 95.84%; Figure 13 (b) shows the traditional support vector optimization process when the kernel is rbf, where the optimal point is: C = 1.001, gamma = 1.001, accuracy = 82.74%; Figure 14 (a) shows the improved cloud membership-support vector optimization process when the kernel is sigmoid, where the optimal point is: C = 2.001, gamma = 21.001, accuracy = 96.46%; Figure 14(b) shows the traditional support vector optimization process when the kernel is sigmoid, where the optimal point is C = 1.001, gamma = 1.001, and accuracy = 92.74%. The poly kernel function can map data from a low-dimensional space to a high-dimensional space, better handling nonlinear problems. However, RBF and sigmoid kernels exhibit oscillations during parameter selection, for example... Figure 13 (b) and 13(b), with alternating light and dark colors, indicate that as C and gamma increase, the accuracy sometimes decreases and sometimes increases, suggesting that the model training process is unstable and difficult to converge. Therefore, according to Figure 12 , 13 The following conclusions were drawn from 14: the accuracy of the poly kernel function is higher than that of rbf and sigmoid; compared with the traditional support vector model, the improved cloud membership-support vector model has higher accuracy.

[0141] In the specific implementation process, this embodiment proposes an improved cloud membership-support vector early warning level zoning correction model based on the risk characteristics of soft soil foundation pits and the experience-based graded early warning zoning method. By converting the two-dimensional early warning zoning into a three-dimensional early warning cloud map, it not only characterizes the fuzziness and randomness of the transition area between adjacent early warning levels, but also solves the problem that the traditional early warning level zoning does not have universality. Based on the engineering case dataset, cloud membership is used to express uncertain information. A balanced class dataset is obtained through SMOTE data augmentation technology. The best model is obtained by cross-validation and grid search. Based on the maximum a posteriori probability principle, the traditional early warning zoning is corrected, and the actual sample distribution is compared to demonstrate the feasibility of the corrected model. It has the following technical effects: (1) A relatively practical and complete foundation pit construction risk early warning mechanism is constructed, covering monitoring, prediction, alarm and correction functions. The causes of foundation pit construction collapse are analyzed, and the relationship between multiple factors, multiple sources of information and multiple monitoring data of collapse is discussed. Various early warning information processing methods are analyzed. Based on the idea of ​​model early warning and quantitative analysis, a risk early warning information processing model framework based on multi-source information is proposed. (2) Based on the types of safety monitoring indicators for underground engineering, a construction principle and standard for an early warning indicator system with strain indicators as the main component and stress and other indicators as supplementary components were proposed. According to the risk characteristics of soft soil subway foundation pits and the principle of indicator selection, based on three common monitoring methods, eight typical early warning indicators covering strain, stress and other categories were selected. Based on cumulative change and rate of change, an early warning indicator system for construction risks of soft soil subway foundation pits was constructed. (3) Based on empirical knowledge such as specifications, literature, and engineering records, the level of early warning indicators was unified through actual monitoring values ​​and control values. The right boundary value of the early warning level was set and a four-level early warning classification standard was divided. The dual control indicators were spatially mapped, and the monitoring items of different information sources were uniformly visualized. Furthermore, an early warning level zoning method based on empirical knowledge was proposed. (4) Based on a two-dimensional cloud model, the two-dimensional early warning zoning was converted into a three-dimensional early warning cloud map, expressing the fuzziness and randomness of the transition area between adjacent early warning levels. To address the imbalanced sample problem inherent in the dataset, SMOTE technology was introduced to balance the engineering case data. Simultaneously, based on the support vector method with self-learning capabilities, an improved cloud membership-support vector early warning level partitioning correction model was proposed, establishing an adaptive and unified method for partitioning and correcting the early warning level of soft soil iron foundation pit construction risks.

[0142] Optionally, risk warning results can be obtained based on improved evidence fusion methods and improved prediction models, including: obtaining a membership matrix set based on the improved prediction model; introducing global uncertainty based on the membership matrix set to construct an initial confidence level; obtaining a two-dimensional distribution of the initial confidence level based on dual control indicators and the initial confidence level; obtaining first-type high-conflict evidence and second-type high-conflict evidence based on the two-dimensional distribution of the initial confidence level; obtaining a first confidence level based on traditional evidence theory and first-type high-conflict evidence; and obtaining a target confidence level based on the first confidence level, second-type high-conflict evidence, and improved evidence fusion methods, so as to obtain risk warning results based on the target confidence level.

[0143] In its implementation, this embodiment, based on the aforementioned early warning standards and combining numerical simulation and deep learning models, uses a real-world engineering project as a background to predict information from various monitoring sources. It also employs a multi-source information fusion method to achieve the theoretical research and practical application of intelligent early warning for foundation pit construction risks. The application of multi-source information fusion in foundation pit engineering is crucial. To ensure safety during subway foundation pit construction, it is typically necessary to set up numerous monitoring points at the construction site to collect various monitoring information in real time. This embodiment processes multi-source data from multiple aspects, including type, fusion method, and application scenario. Specifically, it includes: (1) Types of multi-source information: Monitoring data: By setting up monitoring points in and around the foundation pit, data such as displacement, settlement, stress, and water level can be obtained, which can reflect the actual state of the foundation pit in real time; Geological survey data: Covering soil layer distribution, physical and mechanical properties, groundwater conditions, etc., it is the basic data for foundation pit design and construction, and is used to evaluate the stability of the strata; Design drawings and parameters: Including detailed design information such as the shape, size, and support form of the foundation pit, which clarifies the goals and requirements of the project; Construction process data: Data such as construction progress, construction process parameters, and material properties can be used to monitor whether the construction quality and progress meet the requirements. (2) Fusion methods: Data layer fusion: Directly fuse raw data from different sensors or data sources, such as unifying the analysis of raw signals from displacement sensors and stress sensors; Feature layer fusion: First extract features from each source data, and then fuse these features, such as extracting trend features from monitoring data and soil layer features from geological data for comprehensive analysis; Decision layer fusion: Each data source is processed independently and preliminary decisions are made, and then these decisions are fused, such as combining the early warning results of monitoring data analysis with the judgment of construction experts to decide whether to take emergency measures. (3) Application scenarios: Real-time monitoring and early warning: By fusing multiple monitoring data and geological information, the safety status of the foundation pit can be judged more accurately and early warnings can be issued in a timely manner; Optimized design: Based on the fusion analysis of actual data and geological conditions during the construction process, the original design can be optimized and adjusted, such as adjusting the parameters of the support structure; Construction process control: By fusing construction data and monitoring data, the impact of construction on the foundation pit and the surrounding environment can be grasped in real time, and the construction process and progress can be adjusted in a timely manner; Risk assessment: By integrating multi-source information, the risks of foundation pit projects at different construction stages can be comprehensively assessed, providing a basis for risk management. To improve accuracy: Avoiding the limitations of single-source information, resulting in a more comprehensive and accurate understanding of foundation pit engineering. To enhance reliability: The mutual corroboration and supplementation of information from multiple sources improves the reliability of decision-making and judgment. To achieve intelligent management: Supporting intelligent management of foundation pit engineering, facilitating automated monitoring, analysis, and decision-making.

[0144] Optionally, a first level of trust is obtained based on traditional evidence theory and the first type of high-conflict evidence, including: obtaining the degree of conflict and the degree of difference based on traditional evidence theory and the first type of high-conflict evidence; obtaining an improved conflict factor based on the degree of conflict and the degree of difference according to a preset conflict improvement method, and obtaining a first level of trust based on the improved conflict factor and the initial level of trust.

[0145] In the specific implementation process, the membership matrix ui of the i-th information source is constructed from the cloud membership using the improved prediction model mentioned above, as follows:

[0146]

[0147] In the formula, n=3, indicating that there are 3 information sources for this foundation pit project; m=4, indicating that there are 4 risk warning levels in the early warning standard, and A1, A2, A3, and A4 represent blue, yellow, orange, and red warnings, respectively. uij is the cloud membership degree of the i-th information source at the j-th warning level, i=1, 2, 3, j=1, 2, 3, 4. The cloud membership degree is further defined as follows:

[0148] θ i =1-max(u ij )

[0149] m i (Θ)=θ i

[0150]

[0151] In the formula, θi represents the degree of uncertainty of the i-th information source belonging to the j-th warning level; mi(Θ) is the global uncertainty; and mi(Aj) is the initial confidence level (BPA) of the i-th information source in the j-th warning level. The initial confidence level (BPA) is constructed in two dimensions based on the dual control indicators. The two-dimensional BPA matrix is ​​constructed as shown in the following formula:

[0152]

[0153] And map it onto a two-dimensional cloud early warning system, such as Figure 15 As shown in the figure, the sum of the values ​​of all regions is 1. m(Ajs) represents the BPA score of a certain information source in the s-th region of the j-th warning level. Therefore, the BPA of a certain information source in the j-th warning level is equal to the sum of all BPA components under that warning level.

[0154] When adjusting the first type of high-conflict evidence, since in the traditional DS evidence theory, the conflict factor only considers the degree of conflict as a measure. At present, scholars have conducted a lot of research on conflict factors, such as the conflict coefficient, the Pignistic probability distance and the Jousselme distance. These indicators describe the conflict caused by the evidence itself from different perspectives. However, a single indicator can only reflect one aspect of the conflict. In order to more comprehensively characterize its contradiction, two representative indicators, the degree of conflict α and the degree of difference β, are selected for joint measurement. A new comprehensive conflict factor is constructed through two-dimensional coordinate mapping, thereby effectively solving the first type of conflict problem. Specifically, it includes: (1) Calculating the degree of conflict α: Based on the traditional DS evidence theory, the degree of conflict is represented by the conflict coefficient K, which describes the overall degree of conflict between the evidence. The smaller the degree of conflict, the smaller the conflict. In the construction risk warning, for the evidence composed of different information sources, let αij be the degree of conflict between the i-th information source and the j-th information source, as shown in the following: α ij =∑ i≠ j m i (A)m j (A), where mi(A) and mj(A) represent the initial confidence level BPA of the i-th and j-th information sources when the warning level is A, respectively; (2) Calculate the difference β: The difference can be quantified by the distance in the vector space, which can effectively reflect the similarity of different evidence. The smaller the distance, the greater the similarity of the evidence and the smaller the conflict. This paper adopts the Euclidean distance representation, let βij be the difference between the i-th and j-th information sources, as shown in the following: (3) Improved Conflict Factor: Based on the calculation methods of conflict degree α and difference degree β, both have the mathematical meaning of vector and distance, and both exhibit the same monotonic change trend. The smaller the value, the smaller the contradiction between the evidence. To further optimize, the properties of two-dimensional space are introduced, mapping the conflict degree α to the x-axis to reflect the overall conflict intensity between the evidence; mapping the difference degree β to the y-axis to reflect the local differences between the evidence; the comprehensive state of evidence conflict is intuitively represented by the point (α, β) in the coordinate system, that is, the distance from (α, β) to (0, 0) is represented as the comprehensive conflict factor. The improved conflict factor is as follows: Figure 16 As shown. Let δ represent the credibility of the i-th information source, used to adjust the initial confidence level of the evidence. The calculation formula is as follows: (4) Type I Conflict Adjustment: By applying the improved conflict factor to the initial trust level (BPA) corresponding to the information source, the trust level after Type I conflict adjustment can be obtained. The calculation formula is as follows:

[0155] m adjust,i (A)=(1-δ)·m i (A)

[0156] madjust,j (A)=(1-δ)·m j (A)

[0157] In the formula, madjust,i(A) and madjust,i(A) represent the temporary confidence levels of the i-th and j-th information sources for the warning level A, respectively.

[0158] Optionally, a target trust level is obtained based on a first trust level, a second type of high-conflict evidence, and an improved evidence fusion method, in order to obtain a risk warning result based on the target trust level. This includes: obtaining a comprehensive conflict factor based on the first trust level; obtaining dynamic high conflict and dynamic low conflict based on a preset conflict perception threshold and the comprehensive conflict factor; performing a weighted average based on dynamic high conflict and a preset weighted average rule to obtain first corrected data; obtaining second corrected data based on dynamic low conflict and the improved evidence fusion method; and obtaining a target trust level based on the first and second corrected data, in order to obtain a risk warning result based on the target trust level.

[0159] In practical implementation, the first type of high-conflict evidence adjustment takes the conflict factor more comprehensively, effectively weakening the differences between evidence bodies, but it does not solve the problem of high-conflict evidence failure caused by the defects of the fusion rule. Traditional DS fusion rules are prone to producing results that contradict the facts when facing high-conflict evidence. This is because when synthesizing conflict evidence, the conflicting parts are allocated to each evidence body according to a certain proportion, which may lead to the unreasonable increase or decrease of the credibility of some evidence bodies. The conflict factor should be less than 1. When it is very close to 1, the traditional DS fusion rule will produce the opposite result. To solve the problem of the failure of the traditional DS fusion rule in the high-conflict state, this section proposes a hybrid fusion rule based on the combination of improved DS and weighted averaging to make up for this deficiency. A threshold ζ is set as the critical point for conflict perception to dynamically distinguish between high and low conflict. When the comprehensive conflict factor K>ζ, the state is high conflict, and the weighted averaging rule is adopted, that is, weighted averaging is performed according to the similarity between the current evidence body and other evidence bodies; otherwise, it automatically switches to the improved DS rule, that is, fusion first and then normalization. The threshold value is related to the actual engineering project. This study mainly focuses on intelligent early warning of construction risks in subway foundation pits. Its multi-source information comes from machine learning predictions of monitoring data. In this paper, the threshold is set to 0.9. When δ < 0.9, an improved DS rule is adopted, as shown in the following formula:

[0160]

[0161] In the formula, Based on the improved DS rule, the target confidence level of the ith and jth information sources for the warning level A after the second type of conflict correction is calculated. When A = Θ, it is required that mi(Θ) ≤ 0.1 and max((mrevised(Ak)) - max(mrevised(Ak)) ≥ 0.2, where k ≠ p. That is, the difference between the maximum and second largest values ​​of the mass function must exceed 0.2 for the fusion result to be considered valid; otherwise, the fusion is invalid. When δ ≥ 0.9, a weighted average rule is adopted, as shown in the following formula:

[0162]

[0163] In the formula, ωi is the weighted average coefficient, which is also the weight of the i-th information source. di is the similarity between the i-th information source and other information sources. Based on the weighted average rule, the target confidence level of the warning level A for the second type of conflict is calculated by the i-th and j-th information sources. Similarly, when A = Θ, it is required that mi(Θ) ≤ 0.1 and max((mrevised(Ak)) - max(mrevised(Ak)) ≥ 0.2, where k ≠ p. That is, the difference between the maximum and second largest values ​​of the mass function must exceed 0.2 for the fusion result to be considered valid; otherwise, the fusion is invalid.

[0164] In the specific implementation process, regarding the construction of the initial confidence level (BPA), since the BPA is a preliminary judgment of evidence and a description of uncertain information, and cloud membership has the advantage of being able to accurately express the uncertainty of information, this embodiment adopts the idea of ​​cloud membership to construct the BPA. However, when the sum of cloud memberships is not 1, it does not meet the definition of BPA and cannot be directly used as a BPA. Therefore, this embodiment constructs the initial confidence level BPA based on cloud membership by introducing global uncertainty to form evidence, thereby achieving the universality of the construction method while objectively and effectively representing uncertain information. When dealing with the problem of high-conflict evidence fusion failure, it is divided into two categories according to the source of conflict. The first category is the conflict between evidence bodies, caused by the inconsistency or contradiction of the evidence itself, such as the huge difference in the trust allocation of different evidence sources for the same event. The second category is the defect of the evidence fusion rule, the limitation of the fusion rule itself in high-conflict scenarios, such as the DS rule amplifying conflict due to normalization. Regarding the first category of high conflict... Before evidence synthesis, the original evidence is preprocessed to reduce conflict. This is achieved by discounting the evidence—weighting it according to its reliability and accuracy—to reduce the impact of unreliable evidence on the synthesis result. Regarding the second type of high-conflict problem, traditional DS synthesis rules may produce unreasonable results when dealing with highly conflicting evidence. Therefore, this embodiment improves the synthesis rules by introducing weighted averages and discount factors, thereby enhancing the rationality and accuracy of the synthesis result.

[0165] In the specific implementation process, based on the two-dimensional cloud BPA theory and DS evidence theory, this embodiment proposes a risk early warning method for foundation pit construction based on an improved two-dimensional BPA-DS (2dBPA-DS) multi-source information fusion method. This method effectively constructs initial trust levels and effectively fuses highly conflicting multi-source information, providing strong support for risk early warning research on the complex multi-attribute decision-making problem of subway foundation pit collapse. The process is as follows: Figure 17 As shown, the specific steps include: Based on monitoring data obtained from three monitoring methods—instrument monitoring, computer vision, and UAV aerial photography—the spatiotemporal features of the support axial force, pile horizontal displacement, surface settlement, column settlement, groundwater level, foundation pit convergence, wall horizontal displacement, and vertical displacement are first extracted to obtain the corresponding SelfC-BiGRU predicted values. These predicted values ​​are used as input evidence for multi-source information fusion to proactively perceive risk status and issue warnings. After converting cloud membership into initial confidence level (BPA), a basic probability distribution (BPA) is constructed based on a two-dimensional cloud generator of dual control indicators constructed from cumulative deformation and deformation rate to obtain... The initial state of the evidence is described; new conflict coefficients are obtained by mapping the degree of conflict and the degree of difference to a two-dimensional space, and the BPA is adjusted to the intermediate confidence level mass. New conflict coefficients are generated according to the projection principle of the two-dimensional coordinate system and the relevant properties of the two-dimensional space, thus completing the adjustment of the first type of high-conflict evidence; by setting a threshold, the degree of conflict is perceived by judging the size of the new conflict coefficient and the threshold. If it is greater than the threshold, it is regarded as high-conflict evidence and integrated according to the weighted average rule to avoid fusion failure; if it is less than the threshold, it is integrated according to the improved DS fusion rule; thus completing the correction of the second type of conflict evidence, and finally obtaining the final warning result according to the maximum confidence principle.

[0166] This embodiment considers prediction data from multiple information sources and obtains early warning results by fusing multi-source prediction data. This not only solves the bias caused by early warning of a single monitoring project but also allows for early analysis of potential incidents and the provision of warning prompts. The overall early warning level calculation process mainly consists of three steps: ① collecting prediction values ​​from multiple sources using the SelfC-BiGRU (GicAdam) model; ② fusing multi-source prediction values ​​using an improved two-dimensional BPA-DS method, including two-dimensional BPA transformation and intermediate confidence level calculation; ③ calculating the final confidence level and obtaining the risk early warning result for the multi-source monitoring project based on the maximum confidence level principle. This improved early warning method in this embodiment obtains a more comprehensive representation of multiple future trends in construction risks by collecting prediction values ​​from multiple information sources. These trends are then used as multiple input evidences for multi-source information fusion and fused using the improved two-dimensional BPA-DS method to obtain a comprehensive representation of the risk status, thereby achieving intelligent early warning of subway foundation pit construction risks through multi-source information fusion.

[0167] Example 2:

[0168] This embodiment provides an intelligent risk early warning system based on multi-source data fusion, used to implement the aforementioned intelligent risk early warning method based on multi-source data fusion, including: a monitoring module, a prediction module, and a correction module, wherein:

[0169] The monitoring module is used to acquire multi-source information and obtain early warning control quantities based on the multi-source information and preset quantization processing methods;

[0170] The prediction module is used to construct an initial early warning cloud map based on the early warning control quantity and the preset graded early warning standards;

[0171] The correction module is used to optimize and correct the three-dimensional early warning cloud map using the posterior probability support vector method to obtain an improved prediction model, and to obtain risk warning results based on the improved evidence fusion method and the improved prediction model.

[0172] In the specific implementation process, the monitoring module realizes its monitoring function. During the construction of the subway foundation pit, the safety status of the construction site is monitored through multiple sources, including advanced geological forecasting, on-site inspection, instrument monitoring, drone aerial photography, and computer vision, and monitoring results are generated. A monitoring implementation plan for the foundation pit is formulated based on geological survey data and construction plans. The main purpose is to monitor the safety risks existing in various stages such as retaining structure, dewatering, earthwork excavation, and support erection, and to fully grasp the dynamic changes during the foundation pit construction process to guide the entire construction process. The prediction module realizes its prediction and alarm functions. The prediction function refers to the analysis and prediction of the deformation of the foundation pit and its surrounding environment during the subway foundation pit construction process using the intelligent risk early warning system based on multi-source data fusion provided in this embodiment, including the analysis of the current state and the prediction of the future state. The alarm function refers to the system's ability to automatically activate an alarm procedure and simultaneously generate a notification to the responsible party when measured values ​​reach or exceed preset warning limits during subway foundation pit construction, based on real-time monitoring of various parameters. It also clarifies warning standards, indicator systems, and warning limits according to engineering geological conditions, using methods such as direct comparison to obtain warning information at each stage of construction. Through the implementation of the warning system, multiple channels such as SMS and email are used to constantly alert all participating parties to safety concerns, promptly mobilizing emergency response efforts and reducing the likelihood of safety accidents. The correction module implements its correction function. After monitoring and prediction during subway foundation pit construction and triggering an alarm, it promptly adjusts unsafe and unbalanced states during construction, analyzes the causes of alarms immediately, and activates emergency plans. If the warning and control measures are effectively implemented and the monitored values ​​gradually stabilize, the monitoring density can be reasonably reduced. If there is no convergence within a certain period and there is a continuing increasing trend, the alarm should be upgraded until the state converges. The significance of correction lies in reducing the risk and loss of accidents through effective countermeasures before they occur. This constitutes a complete early warning mechanism for subway foundation pit construction risks.

[0173] In the specific implementation process, when the monitoring module acquires multi-source information, the monitoring objects mainly include the support system and the surrounding environment. The monitoring items involved cover stress, strain, and other categories. Risk monitoring is mainly carried out through on-site inspections, instrument monitoring, computer vision, and drone aerial photography. However, traditional instrument monitoring methods are easily affected by environmental, personnel, and weather factors, resulting in low monitoring efficiency, delayed data acquisition, and an inability to reflect the safety status of the foundation pit project in a timely manner. Therefore, this embodiment uses computer vision and drone aerial photography to supplement and enhance monitoring of high-risk points to ensure the normal progress of foundation pit construction. Among them, computer vision is simple, convenient, and highly accurate. In this embodiment, it is used to supplement the horizontal and vertical displacements of the wall top that were not measured in time due to damage to the instrument monitoring equipment. In addition, since the foundation pit is irregularly shaped, which is prone to construction safety accidents, drone aerial photography, which is more flexible and can perform global three-dimensional imaging, is used to enhance the monitoring of the overall convergence of the foundation pit. During drone aerial photography, high-resolution images of specific areas within the foundation pit are acquired using drones equipped with optical photography equipment (DJI Mavic Pro model). Simultaneously, image processing technology is used to obtain real-time data on the deformation of the foundation pit, enabling dynamic monitoring and early warning of the construction process. The entire process—from image acquisition to point cloud construction, data simplification, and detection / early warning—is implemented using drones. First, based on drone image acquisition methods and point cloud modeling theory, a 3D point cloud model of the foundation pit is constructed using Pix4Dmapper software. To accelerate point cloud data processing while ensuring model accuracy, mainstream algorithms such as bounding box methods, random sampling, and curvature sampling are used to simplify the point cloud data. Finally, deep mining is performed based on the fitted results of the pit's edge point cloud data to obtain convergence deformation values, which are then used to analyze the foundation pit's safety status. Utilizing drone imagery for foundation pit safety monitoring provides an effective way to improve project management efficiency, optimize resource allocation, reduce operating costs, and prevent safety accidents.

[0174] In the specific implementation process, since information fusion must be for the same research object, each research object is composed of multiple individuals. For example, the support system refers to the retaining structure, concrete support, steel support and other individuals. Each individual is equipped with a large number of monitoring points. The data of each monitoring point is different. The stress and strain of different locations are also different. Therefore, it is inaccurate to judge the early warning status by information fusion for the entire support system. For another example, the surrounding environment is composed of the ground surface around the foundation pit. A large number of monitoring points are set up on the ground surface. The settlement and water level of each monitoring point are different. The collapse, cracks and other damage at different locations are also different. Therefore, it is also inaccurate to judge the early warning status by information fusion for the entire surrounding environment. On the other hand, if each monitoring point is treated as a research object for information fusion according to the layout of the monitoring points, this will result in an extremely large amount of data and very low timeliness. Therefore, the fusion measurement area division criteria in this paper are (1) information fusion must be for the same research object; (2) according to the principles of location proximity and merging, the early warning fusion measurement area should be reasonably planned and determined. Based on the above principles, taking the monitoring area as the fusion object and early warning object, the improved two-dimensional BPA-DS evidence fusion model is used for multi-source information fusion construction risk early warning. Taking the early warning monitoring area as the research object, the risk warning status of the monitoring area at different times is obtained. Based on the on-site monitoring point layout map, according to the principles of location proximity and merging, the early warning fusion monitoring area is re-planned and rationally determined, taking the monitoring area as the fusion object and early warning object, in order to achieve efficient early warning fusion and alarm indication. In practical application, the final early warning fusion monitoring area division is as follows: Figure 18 As shown, the black highlighted area represents the early warning measurement zone division result of the support system, and the gray highlighted area represents the early warning measurement zone division result of the surrounding environment. This area includes 8 early warning measurement zones: SS-MK1, SS-MK2, SS-MK3, SS-MK4 and SE-MK1, SE-MK2, SE-MK3, SE-MK4.

[0175] Example 3:

[0176] This embodiment provides an intelligent risk early warning platform based on multi-source data fusion. It is implemented using the aforementioned intelligent risk early warning system based on multi-source data fusion. The improved intelligent risk early warning method based on multi-source data fusion provided in the above embodiment is integrated, encapsulated, and packaged to form a lightweight, convenient, independently runnable and installable application, including: a prediction interface, an early warning interface, and a details interface. Specifically: the optimized early warning standards are written into the background program as default settings, eliminating the need for additional operation by the user; numerical simulation data and measured data obtained from experiments and numerical models are loaded into the system via Excel, and the optimal prediction model for the current dataset is obtained by adjusting parameters; the predicted multi-source data is input into the early warning module to obtain a comprehensive description of the future foundation pit risk status from multiple information sources. The interface mainly includes a user manual, software version, and developer information.

[0177] In the specific implementation process, addressing the problems of lack of uniformity in the initial trust construction and failure of high-conflict evidence fusion in traditional DS fusion methods, this embodiment proposes an intelligent early warning model for subway foundation pit construction risks based on an improved 2dBPA-DS. Based on actual engineering cases, data from three information sources—instrument monitoring, computer vision, and UAV aerial photography—are collected. The entire foundation pit is divided into several areas, with the fusion measurement area as the unit. Through historical measured data and future predicted data, combined with the actual situation of the construction site, the accuracy and effectiveness of the improved method in this chapter are verified from both theoretical and practical perspectives. Furthermore, based on computer programming languages ​​and combined with the improved early warning method, an intelligent early warning information system platform for subway foundation pit construction risks is constructed. The main technical effects obtained are as follows: (1) An improved 2dBPA-DS intelligent early warning method for foundation pit construction risk based on multi-source information fusion is proposed. Not only is the initial confidence level constructed, but also the high-conflict failure problem is adjusted and corrected. The initial confidence level is expressed by using two-dimensional cloud BPA. By introducing conflict degree and difference degree and performing two-dimensional projection, new conflict factors are generated. Unreliable evidence is discounted to achieve the adjustment of the first type of conflict. New synthesis rules are established by improving the DS fusion strategy and weighted average strategy to achieve the correction of the second type of conflict. (2) The limitations of the calculation method of early warning level of single monitoring project are analyzed, and a calculation method of early warning level suitable for multi-source information fusion is proposed. Based on the sameness of the research object of information fusion, and with actual engineering cases as the background, the standard end and the irregular expansion end are re-planned according to the principles of location proximity and merging, and their respective early warning fusion measurement areas are determined. The measurement area is used as the basic unit of data fusion and early warning monitoring at the same time, realizing the efficient integration and accurate prompting of early warning information. (3) Based on actual engineering cases, and using historical measured data from the previous month, the accuracy of the improved early warning method is verified by comparing single information source, traditional DS fusion method, and improved DS fusion method. Based on the predicted data for the next five months, the improved DS fusion method is used to obtain the trend of construction risk changes and its early warning status. Combined with the construction site conditions, the effectiveness of the improved early warning method is verified. (4) Based on the proposed optimized graded early warning standard, constitutive model and stratum seepage law, construction deformation dynamic prediction model and multi-source information risk intelligent early warning model, a subway foundation pit construction risk intelligent early warning information system platform is constructed. It integrates the main functions of standard, prediction, and early warning. Complex calculations can be completed through simple parameter settings and mouse clicks. It forms a lightweight, convenient, independently runnable and installable application, which greatly reduces the workload of on-site operators.

[0178] In practical implementation, the platform user interface mainly includes three panels: input, output, and operation. The input panel is mainly used for parameter setting, while the operation panel is mainly used for controlling data import, running calculations, exporting results, clearing current results, and exiting the system. The prediction interface only requires setting simple model parameters and loading training data. Through the calculation function, it can obtain loss changes, error changes, prediction curves, and evaluation results. It can also export prediction results to a local Excel file. Based on the early warning method proposed earlier, an improved gradient optimization SelfC-BIGRU foundation pit construction deformation prediction program is developed. The prediction interface typically includes a parameter setting window, a control panel, a prediction result graph window, an evaluation index window, and a prediction result window. Based on the proposed optimized hierarchical early warning standard, constitutive model and stratum seepage law, construction deformation prediction model, and multi-source information fusion risk early warning model, a subway foundation pit construction risk intelligent early warning information system is constructed. The early warning interface can achieve information fusion from three information sources. By inputting the predicted values ​​for the next three days obtained from the prediction interface, the calculation function can calculate the corresponding target confidence level and its early warning status. Similarly, the early warning results can be exported to a local Excel file. The program is developed based on the improved two-dimensional BPA-DS risk warning model proposed above. The intelligent warning interface typically includes an input data window, a results display window, and a control panel. It usually displays the predicted values ​​for the next three days, with different colors used to display the warning results for the first, second, and third days.

[0179] Example 4:

[0180] This embodiment provides an intelligent risk warning model based on multi-source data fusion, implemented using an intelligent risk warning method based on multi-source data fusion from the above embodiments, such as... Figure 19 As shown, it includes the following steps:

[0181] Establish a deformation prediction model based on multi-source information, Y (n) pre =F pre (X (n) act ), where X (n) act The actual value of the nth information source is obtained through the prediction model F. pre Then we obtain the predicted value Y of the nth information source. (n) pre ;

[0182] Establish a risk early warning model based on multi-source information prediction, Y warn =F warn (Y (1) pre ,Y (2) pre,......,Y (n) pre ), where Y (n) pre This is the prediction result from the nth information source, processed by the early warning model F. warn The final warning result Y was then obtained. warn .

[0183] In the process of establishing a risk warning model based on multi-source information prediction, the posterior probability support vector method and the improved evidence fusion method are used to optimize and correct the model acquisition, so as to establish an emerging warning information processing method based on statistics and artificial intelligence.

[0184] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent risk early warning based on multi-source data fusion, characterized in that, include: Acquire multi-source information during the construction process of the subway foundation pit, and obtain early warning control quantities based on the multi-source information and a preset quantization processing method; An initial early warning cloud map is constructed based on the aforementioned early warning control quantity and the preset graded early warning standard; Specifically: the warning risk level is obtained based on the aforementioned warning control quantity and the preset graded warning standard; Based on the aforementioned early warning risk level, the cumulative change value and the rate of change value are obtained, and a two-dimensional matrix distribution is constructed based on the cumulative change value and the rate of change value. An initial early warning cloud map is constructed based on the aforementioned two-dimensional matrix distribution using traditional empirical methods. A three-dimensional early warning cloud map is obtained based on the initial early warning cloud map, the two-dimensional matrix distribution, and the preset two-dimensional normal cloud model; specifically: a set of early warning state intervals is obtained based on the initial early warning cloud map; cloud digital features and early warning clouds are obtained based on the set of early warning state intervals and the preset two-dimensional normal cloud model. A three-dimensional early warning cloud map is constructed based on the cloud digital features and the early warning cloud according to a preset construction method; The 3D early warning cloud map is optimized and corrected using the posterior probability support vector method to obtain an improved prediction model. Specifically: a training set is constructed based on historical risk early warning data, and a target separating hyperplane is constructed based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set; the target decision surface is obtained based on the positive sample set and the negative sample set using the posterior probability support vector method, and an optimized posterior model is obtained based on a preset activation function and the target decision surface; historical cumulative change values ​​and historical change rates are obtained based on historical risk early warning data. Historical cloud membership is obtained based on the historical cumulative change value, historical change rate, and the three-dimensional early warning cloud map; A target sample set is obtained based on the historical cumulative change value, historical change rate and historical cloud membership, and a posterior probability distribution is obtained based on the target sample set and the optimized posterior model. Based on the posterior probability distribution and the preset posterior probability principle, the three-dimensional early warning cloud map is optimized and corrected to obtain an improved prediction model. Risk warning results are obtained based on the improved evidence fusion method and the improved prediction model; Specifically: the dual control indicators include cumulative change value and rate of change value; A membership matrix set is obtained based on the improved prediction model; an initial confidence level is constructed by introducing global uncertainty based on the membership matrix set. A two-dimensional distribution of initial trust is obtained based on the dual control indicators and the initial trust level; a first type of high-conflict evidence and a second type of high-conflict evidence are obtained based on the initial trust level; a first trust level is obtained based on traditional evidence theory and the first type of high-conflict evidence. A target trust level is obtained based on the first trust level, the second type of high-conflict evidence, and the improved evidence fusion method, and a risk warning result is obtained based on the target trust level.

2. The intelligent risk early warning method based on multi-source data fusion as described in claim 1, characterized in that, The process of constructing a training set based on historical risk warning data, and then constructing a target separating hyperplane based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set, includes: A training set was constructed based on historical risk warning data; A minority feature set is obtained based on the training set, and nearest neighbor data is obtained based on the minority feature set and a preset similarity calculation method. An optimized training set is obtained based on the nearest neighbor data and a preset synthetic minority class oversampling method; Based on the preset classification rules of the optimized training set, a target separating hyperplane is constructed to obtain positive and negative sample sets.

3. The intelligent risk early warning method based on multi-source data fusion as described in claim 1, characterized in that, The optimization and correction of the three-dimensional early warning cloud map based on the posterior probability distribution and a preset posterior probability principle to obtain an improved prediction model includes: An initial prediction model is constructed based on the aforementioned three-dimensional early warning cloud map; Based on the initial prediction model, a training prediction model is obtained using a preset hierarchical cross-validation method and a preset network search method; The improved prediction model is obtained by optimizing and correcting the posterior probability distribution, the preset posterior probability principle, and the trained prediction model.

4. The intelligent risk early warning method based on multi-source data fusion as described in claim 1, characterized in that, The method of obtaining the first level of trust based on traditional evidence theory and the first type of high-conflict evidence includes: Based on traditional evidence theory and the first type of high-conflict evidence, the degree of conflict and difference is obtained; Based on the conflict degree and the difference degree, an improved conflict factor is obtained according to a preset conflict improvement method, and a first trust degree is obtained based on the improved conflict factor and the initial trust degree.

5. The intelligent risk early warning method based on multi-source data fusion as described in claim 1, characterized in that, The step of obtaining a target trust level based on the first trust level, the second type of high-conflict evidence, and an improved evidence fusion method, and then obtaining a risk warning result based on the target trust level, includes: The comprehensive conflict factor is obtained based on the first level of trust. Dynamic high conflict and dynamic low conflict are obtained based on the preset conflict perception threshold and the comprehensive conflict factor. The first corrected data is obtained by performing a weighted average based on the dynamic high-conflict and preset weighted average rules. The second corrected data is obtained based on the dynamic low-conflict and improved evidence fusion method described above; The target trust level is obtained based on the first and second corrected data, and the risk warning result is obtained based on the target trust level.

6. An intelligent risk early warning system based on multi-source data fusion, characterized in that, A method for implementing an intelligent risk warning method based on multi-source data fusion as described in any one of claims 1 to 5, comprising: a monitoring module, a prediction module, and a correction module, wherein: The monitoring module is used to acquire multi-source information during the construction of the subway foundation pit, and to acquire early warning control quantities based on the multi-source information and a preset quantification processing method. The prediction module is used to construct an initial early warning cloud map based on the early warning control quantity and a preset graded early warning standard. Specifically, it obtains the early warning risk level based on the early warning control quantity and the preset graded early warning standard; it obtains the cumulative change value and change rate value based on the early warning risk level, and constructs a two-dimensional matrix distribution based on the cumulative change value and change rate value; it constructs the initial early warning cloud map using a traditional empirical method based on the two-dimensional matrix distribution. It is also used to obtain a three-dimensional early warning cloud map based on the initial early warning cloud map and a preset two-dimensional normal cloud model. Specifically, it obtains a set of early warning state intervals based on the initial early warning cloud map; it obtains cloud digital features and early warning clouds based on the set of early warning state intervals and the preset two-dimensional normal cloud model; and it constructs a three-dimensional early warning cloud map based on the cloud digital features and early warning clouds according to a preset construction method. The correction module is used to optimize and correct the three-dimensional early warning cloud map using the posterior probability support vector method to obtain an improved prediction model. Specifically: a training set is constructed based on historical risk early warning data, and a target separating hyperplane is constructed based on the training set according to a preset classification rule to obtain a positive sample set and a negative sample set; a target decision surface is obtained based on the positive sample set and the negative sample set using the posterior probability support vector method, and an optimized posterior model is obtained based on a preset activation function and the target decision surface; historical cumulative change values ​​and historical change rates are obtained based on historical risk early warning data; historical cloud membership degrees are obtained based on the historical cumulative change values, historical change rates, and the three-dimensional early warning cloud map; a target sample set is obtained based on the historical cumulative change values, historical change rates, and historical cloud membership degrees, and a posterior probability score is obtained based on the target sample set and the optimized posterior model. The method involves optimizing and correcting the three-dimensional early warning cloud map based on the posterior probability distribution and a preset posterior probability principle to obtain an improved prediction model. It is also used to obtain risk warning results based on an improved evidence fusion method and the improved prediction model. Specifically: the dual control indicators include cumulative change values ​​and change rate values; a membership matrix set is obtained based on the improved prediction model; global uncertainty is introduced based on the membership matrix set to construct an initial confidence level; a two-dimensional distribution of the initial confidence level is obtained based on the dual control indicators and the initial confidence level; a first type of high-conflict evidence and a second type of high-conflict evidence are obtained based on the two-dimensional distribution of the initial confidence level; a first confidence level is obtained based on traditional evidence theory and the first type of high-conflict evidence; a target confidence level is obtained based on the first confidence level, the second type of high-conflict evidence, and the improved evidence fusion method, so as to obtain risk warning results based on the target confidence level.

Images