Tunnel blasting construction safety evaluation system and method thereof

Abstract

The present application relates to the technical field of tunnel blasting construction safety evaluation, in particular to a tunnel blasting construction safety evaluation system and method thereof. It comprises: a multi-source heterogeneous data fusion unit for generating standardized time series data flow; a time series risk deduction unit for outputting physical risk assessment value; a data trust deficit quantification unit for outputting data trust deficit index; a counteractive incentive modeling unit for outputting counteractive incentive pressure signal to the data trust deficit quantification unit; a coupled decision arbitration unit for outputting high-priority intervention instruction; and a counter-fragile strategy execution unit for receiving the high-priority intervention instruction, executing counter-fragile action to obtain real-time data, and for calibrating and resetting the data trust deficit index. The present application overcomes the defect of traditional safety evaluation system which excessively relies on the surface authenticity of data, and improves the identification ability of hidden risks and the reliability of decision-making in complex construction environment.

Description

Technical Field

[0001] This invention relates to the field of tunnel blasting construction safety evaluation technology, specifically to a tunnel blasting construction safety evaluation system and method. Background Technology

[0002] In related technologies, safety assessments for high-risk projects such as tunnel blasting construction typically rely on artificial intelligence systems to evaluate physical risks. These systems predict and assess the extent to which the physical system approaches the failure boundary of a major safety accident by analyzing information such as construction videos from the work site, microseismic sensor data, and manually filled safety self-inspection forms. However, these traditional AI safety systems generally have a technical weakness: they mostly passively rely on reliable data, and the effectiveness of their assessment capabilities is entirely based on the premise that the input data is true and reliable. In actual engineering projects, management often implements aggressive financial incentives or rush-work reward and punishment plans for schedule or cost considerations. This management model can introduce adversarial incentives, leading to motivated data falsification or data embellishment by frontline personnel, such as submitting abnormally perfect or highly regular data.

[0003] Existing AI systems are not designed to effectively deal with this kind of motivated adversarial interference; when the system receives this embellished false data, it becomes factually blind due to the systematic distortion of the data source; the physical risk assessment values ​​it outputs are no longer of reference value, ultimately causing the AI ​​system to be unable to predict catastrophic security incidents caused by data fraud.

[0004] Therefore, a new technical solution is urgently needed to address the passive reliance of existing AI systems on data authenticity, enabling them not only to assess physical risks but also to proactively quantify and manage the risks inherent in data trust itself, thereby avoiding assessment failures caused by adversarial data interference.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a safety evaluation system and method for tunnel blasting construction. Specifically, the technical solution of this invention is as follows:

[0007] A safety evaluation system for tunnel blasting construction, characterized in that it includes:

[0008] The multi-source heterogeneous data fusion unit is used to receive and standardize multi-source heterogeneous data from all work sites to generate a standardized time-series data stream.

[0009] a time-series risk deduction unit configured to receive the standardized time-series data stream, dynamically deduce an implicit risk propagation chain, and output a physical risk assessment value;

[0010] a data trust deficit quantification unit configured to receive the standardized time-series data stream in parallel, quantify a data trust deficit of the construction data, and output a data trust deficit indicator;

[0011] a counteractive incentive modeling unit configured to receive and analyze business instructions of a project management layer, identify a root cause of the data trust deficit, and output a counteractive incentive pressure signal to the data trust deficit quantification unit;

[0012] a coupled decision arbitration unit configured to receive the physical risk assessment value and the data trust deficit indicator, make a decision based on a risk-trust dual assessment and an anti-fragile decision closed-loop mechanism, and output a high-priority intervention instruction;

[0013] an anti-fragile strategy execution unit configured to receive the high-priority intervention instruction, execute an anti-fragile action to obtain field real data, and feed the field real data back to the data trust deficit quantification unit for calibration and resetting of the data trust deficit indicator.

[0014] Further, the time-series risk deduction unit is specifically configured to dynamically learn and predict coupled risks in a human-machine-environment system based on time-series graph neural network technology.

[0015] The physical risk assessment value is a quantitative indicator reflecting a degree to which a system approaches a failure boundary of a major safety accident.

[0016] Further, the data trust deficit quantification unit is specifically configured to:

[0017] establish a historical normal baseline model based on project historical data or a similar engineering experience database through statistical learning to define a data pattern under normal construction conditions;

[0018] compare a deviation degree of the standardized time-series data stream from the historical normal baseline model;

[0019] when data abnormality is perfectly monitored or a submission time of a manual self-checking table is highly regularized, it is determined that the data is seriously inconsistent with the historical normal baseline model, so as to generate the data trust deficit indicator.

[0020] Further, the counteractive incentive modeling unit is specifically configured to:

[0021] receive and analyze business instructions containing aggressive financial incentive policies;

[0022] analyze how the aggressive financial incentive policy introduces a zero-sum game and incentivizes the frontline staff to prioritize progress over compliance to generate the adversarial incentive pressure signal.

[0023] Further, the data trust deficit quantification unit is further configured to:

[0024] receive the adversarial incentive pressure signal from the adversarial incentive modeling unit;

[0025] use the adversarial incentive pressure signal as a control parameter to dynamically adjust the deviation degree comparison logic;

[0026] automatically reduce the tolerance for perfect data as the incentive pressure increases to make the data trust deficit indicator more sensitive.

[0027] Further, the coupling decision arbitration unit is specifically configured to:

[0028] determine that the data source is substantially trustworthy when the data trust deficit indicator is below a preset blind warning threshold;

[0029] in the state of determining that the data source is substantially trustworthy, use the physical risk assessment value output by the time-series risk inference unit as the main decision basis for regular risk management;

[0030] in the state, also use the data trust deficit indicator as a meta-weight to adjust the credibility of the physical risk assessment value.

[0031] Further, the coupling decision arbitration unit is further configured to:

[0032] when the data trust deficit indicator accumulates and reaches or exceeds the blind warning threshold, execute a trust veto logic;

[0033] the blind warning threshold is generated based on statistical analysis and inverse calibration of cases in historical data where data distortion led to assessment failure;

[0034] when executing the trust veto logic, determine that the physical risk assessment value is untrustworthy regardless of whether the physical risk assessment value is safe;

[0035] after determining that the physical risk assessment value is untrustworthy, automatically switch to a data authenticity reconstruction mode;

[0036] after switching to the data authenticity reconstruction mode, output the high-priority intervention instruction.

[0037] Further, the anti-fragile strategy execution unit is specifically configured to:

[0038] Immediately after receiving the high-priority intervention instruction, a preset anti-fragile action is performed;

[0039] The anti-fragile action includes assigning tasks to an independent flight inspection team, randomly and dynamically conducting a surprise on-site inspection on a subcontractor with the highest data suspicion, or forcibly installing a third-party encryption sensor that is difficult to tamper with;

[0040] The on-site real data obtained through the surprise on-site inspection is fed back to the data trust deficit quantification unit.

[0041] A tunnel blasting construction safety evaluation method, comprising the following steps:

[0042] Step 1: receiving and standardizing processing multi-source heterogeneous data through a multi-source heterogeneous data fusion unit to generate standardized time series data stream;

[0043] Step 2: receiving the standardized time series data stream through a time series risk deduction unit to output a physical risk assessment value;

[0044] Step 3: receiving the standardized time series data stream in parallel through a data trust deficit quantification unit, and combining an adversarial incentive pressure signal from an adversarial incentive modeling unit to output a data trust deficit index;

[0045] Step 4: receiving the physical risk assessment value and the data trust deficit index through a coupled decision arbitration unit, making a decision based on a risk-trust double evaluation and anti-fragile decision closed-loop mechanism, and outputting a high-priority intervention instruction;

[0046] Step 5: receiving the high-priority intervention instruction through an anti-fragile strategy execution unit, performing an anti-fragile action to obtain on-site real data, and feeding the on-site real data back to the data trust deficit quantification unit for calibration and resetting the data trust deficit index.

[0047] Further, when the data trust deficit index is lower than a preset blind warning threshold, regular risk management is performed based on the physical risk assessment value as the main decision basis;

[0048] When the data trust deficit index accumulates and reaches or exceeds the blind warning threshold, a trust veto logic is executed;

[0049] After executing the trust veto logic, it is determined that the physical risk assessment value is not trustworthy;

[0050] After determining that the physical risk assessment value is not trustworthy, automatically switch to a data authenticity reconstruction mode;

[0051] After switching to the data authenticity reconstruction mode, output the high-priority intervention instruction.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. This invention pioneers a dual assessment mechanism for physical risk and data trust. It not only dynamically extrapolates coupled risks within the human-machine-environment system but also innovatively quantifies the trust deficit in construction data in parallel. This design overcomes the shortcomings of traditional safety assessment systems that over-rely on the apparent authenticity of data. Through risk-trust coupled decision-making, it significantly enhances the ability to identify hidden risks and the reliability of decisions in complex construction environments.

[0054] 2. This invention can intelligently identify and quantify data distortion caused by management pressure. Through an adversarial incentive modeling unit, the system can analyze how aggressive financial incentives and other directives create a zero-sum game, causing data compliance to give way to project progress. This model can dynamically adjust the sensitivity of the trust deficit algorithm; when management pressure increases, the system automatically reduces its tolerance for perfect data, penetrating the data fog at its source.

[0055] 3. This invention possesses the ability to deeply understand abnormally perfect data. By establishing a historical baseline model, the system treats occasional anomalies as the norm, while remaining vigilant towards manually generated data that is abnormally perfect or whose submission times are highly regular. When such perfect data deviating from the baseline is detected, the system can determine its low credibility, effectively preventing the evaluation system from being deceived by false or embellished data.

[0056] 4. This invention constructs an anti-fragile decision-making closed loop from trust rejection to data reconstruction. When the data trust deficit index exceeds the blindness warning threshold, the system will reject the current physical risk assessment value and automatically switch to the data authenticity reconstruction mode. By assigning anti-fragile actions such as surprise inspections by flight inspection teams to obtain real on-site data, the system is forced to return to a trusted state, ensuring security redundancy and control resilience in the event of data distortion. Attached Figure Description

[0057] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0058] Figure 1 This is a system structure diagram of the present invention.

[0059] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0061] Example 1

[0062] Please see Figure 1A safety evaluation system for tunnel blasting construction, comprising:

[0063] The multi-source heterogeneous data fusion unit is used to receive and standardize multi-source heterogeneous data from all work sites to generate a standardized time-series data stream.

[0064] The time-series risk simulation unit is used to receive standardized time-series data streams, dynamically simulate the transmission chain of implicit risks, and output physical risk assessment values.

[0065] The data trust deficit quantification unit is used to receive standardized time-series data streams in parallel, quantify the trust deficit of construction data, and output the data trust deficit index.

[0066] The adversarial incentive modeling unit is used to receive and analyze business instructions from the project management team, identify the root causes of the data trust deficit, and output adversarial incentive pressure signals to the data trust deficit quantification unit.

[0067] The coupled decision arbitration unit is used to receive physical risk assessment values ​​and data trust deficit indicators, make decisions based on risk-trust dual assessment and antifragile decision-making closed-loop mechanism, and output high-priority intervention instructions.

[0068] The antifragile strategy execution unit is used to receive high-priority intervention instructions, execute antifragile actions to obtain real-world data, and feed the real-world data back to the data trust deficit quantification unit for calibration and resetting the data trust deficit indicator.

[0069] This system demonstrates a complete risk-trust dual assessment and decision-making closed loop;

[0070] The multi-source heterogeneous data fusion unit constitutes the system's data entry point; its core purpose is to solve the problems of inconsistent data standards, chaotic formats, and delays caused by multiple subcontractors. This unit is configured to receive data from all work sites in real time, including high-dimensional unstructured construction videos, microseismic sensor data, personnel positioning data, and manually filled safety self-inspection forms. To achieve this, the unit integrates a data cleaning and standardization engine, which transforms these raw data into standardized multimodal data streams through methods such as timestamp alignment, format conversion (e.g., converting unstructured video streams into feature vectors), OCR-based manual reports, and data completion algorithms.

[0071] This standardized data stream serves as a common foundation for subsequent analyses and is distributed to subsequent units. For example, this unit generates a dynamic graph structure data stream for the time series graph neural network for the time series risk extrapolation unit, such as nodes being personnel / equipment and attributes being sensor readings. At the same time, it generates a feature vector data stream for the statistical model for the data trust deficit quantification unit, such as statistical features including report submission time intervals and the mean and variance of microseismic events, to initiate parallel dual assessments.

[0072] The time-series risk simulation unit is specifically designed to receive this standardized data stream. Its purpose is to simulate the implicit risk transmission chain in the human-machine-environment system. In this embodiment, it can dynamically learn and predict coupled risks based on, but not limited to, time-series graph neural network (TGNs) technology. This unit focuses on analyzing the physical state described by the data and assessing the degree to which the system, such as the tunnel face, is close to a major safety accident or failure boundary. This unit ultimately outputs a physical risk assessment value and sends it to the coupled decision arbitration unit.

[0073] While the risk simulation unit is operating, the data trust deficit quantification unit receives the same data stream in parallel. The innovation of this unit is that its purpose is not to analyze the content of the data, but to analyze the credibility of the data itself. Its operating mechanism is to monitor abnormally perfect or highly regular data by comparing real-time data with a pre-established historical normal baseline model. These data are seriously inconsistent with the historical normal baseline model, thereby actively quantifying the systematic distortion risk of the data.

[0074] To ensure the accuracy and foresight of trust quantification, the system is further configured with an adversarial incentive modeling unit. Its unique purpose is to identify the motivations and root causes that lead to data fraud. When this unit analyzes management instructions using techniques such as Natural Language Processing (NLP) and identifies high-pressure policies that introduce zero-sum games, such as large reward and punishment plans for rushing milestones, it outputs an adversarial incentive pressure signal. This signal is then sent to the data trust deficit quantification unit and used as a key dynamic adjustment control parameter. The underlying logic is that when the pressure signal increases, i.e., the motivation to commit fraud increases, the quantification unit anticipates this and automatically reduces its tolerance for perfect data.

[0075] Based on this adjustment, the data trust deficit quantification unit outputs a data trust deficit index and sends it to the coupled decision arbitration unit.

[0076] The coupled decision-making arbitration unit constitutes the core of the invention's decision-making process; it simultaneously receives physical risk assessment values ​​and data trust deficit indicators; the arbitration mechanism of this unit operates based on a risk-trust dual assessment and anti-fragile decision-making closed-loop mechanism: when the data trust deficit indicator is low, i.e. below the blindness warning threshold, it uses the physical risk assessment value as the primary basis for routine risk management; once the data trust deficit indicator accumulates to exceed the threshold, the unit will immediately execute the trust rejection logic—that is, no matter how safe the physical risk assessment value appears, the arbitration unit will determine that the value is untrustworthy because the data source it relies on has been determined to be systematically distorted; at this time, the arbitration unit automatically switches to the data authenticity reconstruction mode and outputs a high-priority intervention instruction;

[0077] The high-priority intervention instruction is sent to the antifragile strategy execution unit. The purpose of this unit is to execute strategies that are more costly and cumbersome under normal circumstances, but are more effective in restoring the authenticity of data. Upon receiving the instruction, the unit immediately executes preset antifragile actions, such as automatically assigning a task to an independent, expensive flight inspection team to conduct random, dynamic surprise on-site inspections of the subcontractor whose data they deem most perfect and therefore most suspicious; or triggering a procurement process that mandates the installation of expensive but tamper-proof third-party encrypted sensors in the work area. A key closed loop of this invention is that the real-world data obtained by the unit in executing the actions is forcibly fed back to the data trust deficit quantification unit for calibration and resetting of the data trust deficit index.

[0078] Through the collaborative work of the above six units, this system achieves a complete closed loop of deficit detection -> decision intervention -> verification of authenticity -> rebuilding trust. It surpasses traditional AI systems that only assess physical risks, innovatively quantifying and managing the risks inherent in data trust itself. This is fundamentally different from existing technologies, i.e., the background technology, which passively rely on trusted data and cannot cope with motivated adversarial interference. The anti-data vulnerability decision-making closed loop of this invention can effectively prevent the system from becoming factually blind due to the manipulation of input data, proactively avoid major security incidents caused by data fraud that AI cannot predict, thereby improving the risk protection of high-risk engineering projects.

[0079] Example 2

[0080] The specific configuration of the temporal risk simulation unit is as follows: based on temporal graph neural network technology, it dynamically learns and predicts the coupling risks in the human-machine-environment system;

[0081] Physical risk assessment value is a quantitative indicator reflecting the degree to which a system approaches the failure boundary of a major safety accident;

[0082] In this embodiment, its purpose is to simulate the hidden risk transmission chain that is difficult to detect in high-dimensional data in real time; it receives standardized time-series data streams from multi-source heterogeneous data fusion units; to achieve its function, this unit can adopt the temporal graph neural network (TGNs) technology; TGNs technology is particularly good at dynamically learning and predicting coupling risks in human-machine-environment systems. For example, it can model the dynamic temporal correlation between the fatigue state of construction workers, the abnormal vibration of the tunneling machine, and the microseismic reading ring at the working face.

[0083] The output of this unit, namely the physical risk assessment value, is a quantitative indicator used to characterize the degree of proximity between the current state of a system, such as a tunnel face, and a pre-defined catastrophic accident, i.e., the failure boundary. Here, the failure boundary is what is commonly referred to in the field as a major safety accident as defined by the project or regulatory agency, such as any accident that results in 3 or more deaths or direct economic losses exceeding 50 million. In specific implementation, the physical risk assessment value is divided into multiple qualitative levels, such as safe, concern, warning, and high risk, to reflect the degree to which the system is close to the failure boundary of a major safety accident, so as to facilitate management decision-making.

[0084] By employing Time-Series Graph Neural Networks (TGNs) technology, the Time-Series Risk Inference Unit can more deeply explore and predict the dynamic coupling risks in the human-machine-environment system. Compared with traditional static or low-dimensional models, it can identify implicit risk transmission chains earlier. This makes the physical risk assessment value output by the system more accurate and forward-looking, thereby improving the system's ability to assess physical risks under the assumption that the data is true.

[0085] Example 3

[0086] The data trust deficit quantification unit is specifically used for:

[0087] Establish a historical normal baseline model. The historical normal baseline model is established through statistical learning based on historical project data or experience databases of similar projects. It is used to define the data pattern under normal construction conditions.

[0088] The deviation of the standardized time-series data stream from the historical baseline model is compared.

[0089] When abnormally perfect data or highly regular submission times of manual self-inspection forms are detected, it is determined that they are seriously inconsistent with the historical baseline model, so as to generate a data trust deficit indicator.

[0090] The internal working logic of the data trust deficit quantification unit is designed to proactively quantify the construction data trust deficit.

[0091] The operation of this unit relies on the establishment of a historical normal baseline model. This historical normal baseline model is a model built using statistical learning methods, with data sources including the project's own historical data or experience databases of similar projects. The establishment process may include: collecting, for example, construction data from the past 6 months that has been verified as real and safe, including sensor readings, report submission time distribution, etc., and learning normal data distribution and fluctuation ranges through methods such as Gaussian Mixture Model (GMM) or anomaly detection algorithms such as IsolationForest, thereby defining the data pattern under normal construction conditions. In this definition, a core common knowledge is that normal construction inevitably includes occasional anomalies, such as occasional sensor jumps and slight random fluctuations in report submission times; that is, occasional anomalies are the norm.

[0092] During operation, this unit receives standardized time-series data streams from the multi-source heterogeneous data fusion unit in real time and compares them with the aforementioned historical normal baseline model in real time to determine their deviation. This deviation can be calculated using, for example, Mahalanobis distance or KL divergence to quantify the degree to which real-time data points fall outside the normal distribution. For example, when using KL divergence, this deviation can be expressed as... ,in It is the current data distribution extracted from a standardized time-series data stream. It is the normal data distribution defined by the historical normal baseline model;

[0093] The innovation of this unit lies in its reverse judgment logic: when it detects that real-time data becomes abnormally perfect, for example, the microseismic sensor readings remain on an overly smooth safety straight line for a long time, or the submission time of the manual self-inspection form shows a high degree of regularity, for example, it is submitted on time at 17:00:00 every day for several consecutive weeks, then the system will determine that: this perfection is seriously inconsistent with the historical normal baseline model; this inconsistency is quantified to generate a data trust deficit index; the data trust deficit index can be a dynamically accumulated score normalized to [0,1], where 0 represents complete trust and 1 represents reaching the blindness warning threshold;

[0094] By employing this baseline model comparison and reverse judgment mechanism, this invention solves the problem of the difficulty in quantifying data trust deficit. It enables the data trust deficit quantification unit to proactively identify fabricated and false data by analyzing the data patterns themselves, rather than relying on the data content. This design allows the system to effectively distinguish between real occasional anomalies and false, perfect data, providing reliable and quantifiable input for subsequent trust rejection logic.

[0095] Example 4

[0096] The adversarial incentive modeling unit is specifically used for:

[0097] Receive and analyze business instructions that include aggressive financial incentive policies;

[0098] This study analyzes how aggressive financial incentive policies introduce a zero-sum game and incentivize frontline staff to prioritize progress over compliance, thereby generating adversarial incentive pressure signals.

[0099] The data trust deficit quantification unit is also used for:

[0100] Receive adversarial incentive pressure signals from the adversarial incentive modeling unit;

[0101] The antagonistic excitation pressure signal is used as the control parameter for dynamically adjusting the deviation comparison logic;

[0102] When incentive pressure increases, the tolerance for perfect data is automatically reduced to make the data trust deficit indicator more sensitive.

[0103] An innovative coupling mechanism between management incentives and data trust quantification; the purpose of this mechanism is to incorporate the root cause of the data trust deficit—namely, the impact of management on data authenticity—into the scope of AI model considerations.

[0104] In this configuration, the adversarial incentive modeling unit is used as a management stress sensor; it is configured to receive and analyze business instructions issued by project management; it pays particular attention to aggressive financial incentive policies, such as large reward and punishment schemes for rushing milestones; the unit further analyzes how these policies introduce a zero-sum game among subcontractors, and how this game systematically incentivizes frontline staff to prioritize schedule over compliance.

[0105] Based on this analysis, the unit generates an adversarial excitation pressure signal; the calibration of this signal can be achieved based on a set of preset mapping rule engines. For example, if the NLP only analyzes the rush order, then the signal coefficients... The calibration is set to 1.5; if NLP analysis identifies rush work and includes zero-sum game keywords such as huge rewards and penalties, then the signal coefficient is... The calibrated value is 3.0; the adversarial incentive pressure signal is an adjustment signal used to characterize the strength of the motivation of front-line construction teams to falsify data due to the incentive policies of management. For example, it can be a qualitative signal with three levels: low, medium and high, or an adjustment coefficient in the range of [1.0, 3.0].

[0106] The data trust deficit quantification unit receives adversarial incentive pressure signals from the adversarial incentive modeling unit. Its core innovation lies in using this pressure signal as a dynamically adjusted control parameter for its internal deviation comparison logic. For example, if the pressure signal has a low coefficient of 1.0, the trust deficit index is calculated normally; if the pressure signal has a high coefficient of 3.0, the data trust deficit quantification unit will multiply the calculated deviation score by 3.0, or automatically tighten its statistical threshold for judging perfect data, for example, reducing the p-value for judging anomalies from 0.05 to 0.01, thereby automatically reducing its tolerance for perfect data.

[0107] This dynamic adjustment makes the data trust deficit indicator more sensitive to signs of data distortion under high management pressure;

[0108] This system implements a motivation-behavior coupling analysis; it no longer passively waits for perfect data to become a given fact, but actively predicts the motivation for data fraud through adversarial incentive modeling units; when the pressure of the motivation to commit fraud increases, the system actively raises its level of suspicion; this design greatly enhances the system's ability to deal with motivated adversarial interference, making the assessment of the data trust deficit index more forward-looking, proactive and sensitive.

[0109] Example 5

[0110] The coupled decision arbitration unit is specifically used for:

[0111] When the data trust deficit index is lower than the preset blindness warning threshold, the data source is determined to be basically trustworthy.

[0112] Under the condition that the data source is basically reliable, routine risk management is based primarily on the physical risk assessment value output by the time-series risk simulation unit.

[0113] In this state, the data trust deficit indicator will also be used as a meta-weight to adjust the credibility of the physical risk assessment value.

[0114] The coupled decision arbitration unit is also used for:

[0115] When the data trust deficit indicator accumulates and reaches or exceeds the blindness warning threshold, the trust denial logic is executed.

[0116] The blindness warning threshold is generated based on statistical analysis and reverse calibration of historical data cases where data distortion led to assessment failure;

[0117] When executing the trust veto logic, the physical risk assessment value is deemed unreliable regardless of whether it is safe or not.

[0118] After determining that the physical risk assessment value is unreliable, it automatically switches to the data authenticity reconstruction mode;

[0119] After switching to data authenticity reconstruction mode, output high-priority intervention instructions;

[0120] The coupled decision arbitration unit—which implements a risk-trust dual assessment and antifragile decision-making closed-loop mechanism—simultaneously receives physical risk assessment values ​​from the time-series risk deduction unit and data trust deficit indicators from the data trust deficit quantification unit. The decision logic of this unit is configured with two mutually exclusive states:

[0121] State 1: Normal State

[0122] Its trigger condition is when the data trust deficit index, i.e. the degree of suspicion of data distortion, such as a score of [0,1], is lower than a preset blindness warning threshold;

[0123] The blindness warning threshold is a key, configurable parameter. It should be understood that the specific value of this threshold is not fixed, for example, set to 0.85. Instead, it can be generated through statistical analysis and reverse calibration based on the type of project being processed and cases in historical data where data distortion ultimately led to the failure of AI assessment, i.e., failure to predict accidents. Its purpose is to ensure that the system intervenes in time before complete blindness.

[0124] In this state, the arbitration unit determines that the data source is basically reliable; its decision-making logic is based primarily on the physical risk assessment value, and it conducts routine risk management, such as issuing an early warning if the physical risk is high.

[0125] As an optimized configuration, under normal circumstances, this lower data trust deficit index can also be used as a meta-weight to make a slight adjustment to the credibility of the physical risk assessment value. For example, slightly lower the credibility of a seemingly safe physical risk value, but the physical risk assessment value still dominates.

[0126] State 2: Rejection

[0127] The trigger condition is when the data trust deficit indicator accumulates dynamically and reaches or exceeds the aforementioned blindness warning threshold;

[0128] At this point, the arbitration unit will execute a trust rejection logic; this is the core innovation of this invention that distinguishes it from the prior art: when executing this logic, the arbitration unit's judgment changes: no matter how perfect or safe the physical risk assessment value appears, the arbitration unit will determine that the physical risk assessment value is unreliable.

[0129] The underlying logic is that the input data source on which the physical risk assessment value depends has now been determined to be systematically distorted, therefore the assessment value is factually blind and has no decision-making reference value.

[0130] After determining that the physical risk assessment value is unreliable, the strategic goal of the arbitration unit is no longer to optimize explicit efficiency such as project progress speed, but to automatically switch to the data authenticity reconstruction mode.

[0131] After switching to the data authenticity reconstruction mode, the only action of this unit is to output a high-priority intervention command to the antifragile strategy execution unit.

[0132] This invention establishes a robust circuit breaker mechanism within the AI ​​security system. It makes a clear and quantifiable strategic choice between pursuing short-term project progress based on physical risk assessment values ​​and maintaining long-term data trust based on data trust deficit indicators. This trust veto logic has anti-data vulnerability, ensuring that the AI ​​system will not be deceived by fabricated data due to data trust deficit when facing organized adversarial interference, thus effectively avoiding catastrophic accidents.

[0133] Example 6

[0134] The antifragile strategy execution unit is specifically used for:

[0135] Upon receiving a high-priority intervention instruction, immediately execute the preset antifragile action;

[0136] Antifragile actions include assigning tasks to independent flight inspection teams, conducting random and dynamic surprise site inspections of subcontractors with the most suspicious data, or mandating the installation of third-party encrypted sensors that are difficult to tamper with.

[0137] The real data obtained through surprise on-site inspections will be fed back to the data trust deficit quantification unit.

[0138] The specific configuration and actions of the antifragile strategy execution unit, which is the final executor of the data authenticity reconstruction model;

[0139] The purpose of this unit is to implement suboptimal security strategies that are more costly and cumbersome under normal circumstances, but are more effective at restoring the authenticity of data.

[0140] This unit is configured to receive high-priority intervention instructions from the coupled decision arbitration unit; the instructions can be configured as a structured data signal containing the intervention type, target subcontractor identifier, and coordinates of the area with the highest data suspicion; the antifragile strategy execution unit integrates a task scheduling engine, which automatically parses the instruction content upon receiving the instruction and calls the project management system's API interface to generate work orders for the flight inspection team, or sends a request to the procurement system to force the installation of third-party sensors;

[0141] Upon receiving this instruction, the system has already determined that the data is untrustworthy, and the unit immediately executes the preset antifragile action;

[0142] Antifragile actions are specific, high-intensity interventions, examples of which include, but are not limited to:

[0143] Action Example 1 Unit automatically assigns a task to a separate, expensive flight inspection team; the team is authorized to conduct random, dynamic surprise on-site inspections of the subcontractor whose data is the most perfect and who is the most suspicious.

[0144] Action Example 2 unit triggers a procurement or installation process that mandates the installation of costly but tamper-proof third-party encrypted sensors in the work area where data distortion is most suspected.

[0145] The ultimate goal of these actions is to obtain real-time data from the scene. The key closed loop of this invention is to forcibly feed back the real-time data obtained through surprise on-site inspections or third-party sensors to the data trust deficit quantification unit.

[0146] This feedback data will be used to calibrate and reset the data trust deficit index; for example, if the verification finds that the data is indeed falsified, the trust deficit index will remain high until the new sensor data stabilizes; if the verification finds that it is a misjudgment, the real data can be used to correct the baseline model and immediately reset the trust deficit index to restore the system to a normal state.

[0147] The configuration of the antifragile strategy execution unit provides a powerful execution loop for the trust veto logic. By executing highly deterministic antifragile actions such as surprise inspections, it breaks the cycle of subcontractor data manipulation. By feeding back the acquired real-world data to the data trust deficit quantification unit, it completes a full closed loop of deficit discovery -> decision intervention -> verification of authenticity -> rebuilding trust. This ensures that the system can self-repair and calibrate after experiencing a trust crisis, restoring its evaluation capabilities.

[0148] Example 7

[0149] Please see Figure 2 A method for safety evaluation of tunnel blasting construction, comprising:

[0150] Step 1: Receive and standardize multi-source heterogeneous data through the multi-source heterogeneous data fusion unit to generate a standardized time-series data stream;

[0151] Step 2: Receive standardized time-series data streams through the time-series risk simulation unit and output physical risk assessment values;

[0152] Step 3: The data trust deficit quantification unit receives standardized time-series data streams in parallel and combines them with adversarial incentive pressure signals from the adversarial incentive modeling unit to output the data trust deficit index.

[0153] Step 4: Receive physical risk assessment values ​​and data trust deficit indicators through the coupled decision arbitration unit, make decisions based on the risk-trust dual assessment and anti-fragile decision-making closed-loop mechanism, and output high-priority intervention instructions;

[0154] Step 5: Receive high-priority intervention instructions through the antifragile strategy execution unit, execute antifragile actions to obtain real-world data, and feed the real-world data back to the data trust deficit quantification unit for calibration and resetting of the data trust deficit indicator;

[0155] Step 4 specifically includes:

[0156] When the data trust deficit index is lower than the preset blindness warning threshold, routine risk management is carried out based primarily on the physical risk assessment value.

[0157] When the data trust deficit indicator accumulates and approaches or exceeds the blindness warning threshold, the trust denial logic is executed.

[0158] After executing the trust rejection logic, the physical risk assessment value is determined to be unreliable;

[0159] After determining that the physical risk assessment value is unreliable, it automatically switches to the data authenticity reconstruction mode;

[0160] After switching to data authenticity reconstruction mode, output high-priority intervention instructions;

[0161] This method achieves all the functions of the aforementioned system by performing a series of steps;

[0162] The method includes:

[0163] Step 1: Data Fusion

[0164] This step corresponds to the function of the multi-source heterogeneous data fusion unit; when the method starts, it receives and standardizes multi-source heterogeneous data from all work sites, such as construction videos, microseismic data, and manual self-inspection forms, to generate a standardized time-series data stream.

[0165] Step 2: Physical Risk Simulation

[0166] This step corresponds to the function of the time-series risk deduction unit; the method receives a standardized time-series data stream, analyzes the data content through techniques such as Time-Series Graph Neural Networks (TGNs), deduces the human-machine-environment coupling risk, and outputs the physical risk assessment value to step 4;

[0167] Step 3: Quantifying the Trust Deficit

[0168] This step corresponds to the collaborative function of the data trust deficit quantification unit and the adversarial incentive modeling unit; this step receives standardized time-series data streams in parallel and compares their deviation with the historical normal baseline model; at the same time, this step incorporates adversarial incentive pressure signals from the adversarial incentive modeling unit to dynamically adjust the sensitivity of its comparison logic; when abnormally perfect data is detected or a high pressure signal is received, this step outputs the data trust deficit index to step 4.

[0169] Step 4: Coupled Decision Arbitration

[0170] This step is the core of the method, corresponding to the function of the coupled decision arbitration unit. This step receives the physical risk assessment value from step 2 and the data trust deficit indicator from step 3, and performs a risk-trust dual assessment:

[0171] The conventional logic is that when the data trust deficit index is lower than the preset blindness warning threshold, the method determines that the data is trustworthy and uses the physical risk assessment value as the main basis for decision-making and conducts conventional risk management.

[0172] When the data trust deficit indicator accumulates and reaches or exceeds the blindness warning threshold, the method executes the trust rejection logic.

[0173] After the trust veto logic is executed, the method determines that the physical risk assessment value is unreliable.

[0174] After determining that the physical risk assessment value is unreliable, the method automatically switches to the data authenticity reconstruction mode.

[0175] After switching to the data authenticity reconstruction mode, the final output of this step is to output a high-priority intervention instruction to step 5.

[0176] Step 5: Antifragile Execution and Closure:

[0177] This step corresponds to the function of the antifragile strategy execution unit; the method receives high-priority intervention instructions and executes antifragile actions such as assigning a flight inspection team to conduct surprise inspections in order to obtain real-time data on site.

[0178] This step feeds back real-world data to the data trust deficit quantification unit (Step 3) to calibrate and reset the data trust deficit indicator.

[0179] This method provides a complete, operable, and closed-loop evaluation process. It places the physical risk assessment step 2 and the data trust assessment step 3 in parallel, and achieves proactive management of data trust deficits by coupling the core step of decision arbitration step 4. The antifragile execution and closed loop in step 5 ensure that when faced with motivated data pollution, the method can abandon the ineffective analysis of distorted data and instead perform the action of reconstructing the authenticity of the data, thereby ensuring the long-term effectiveness and reliability of the security assessment.

[0180] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A safety evaluation system for tunnel blasting construction, characterized in that, include: The multi-source heterogeneous data fusion unit is used to receive and standardize multi-source heterogeneous data from all work sites to generate a standardized time-series data stream. The time-series risk simulation unit is used to receive the standardized time-series data stream, dynamically simulate the implicit risk transmission chain, and output the physical risk assessment value. The time-series risk simulation unit is specifically configured to: dynamically learn and predict the coupling risks in the human-machine-environment system based on time-series graph neural network technology, and the physical risk assessment value is a quantitative indicator reflecting the degree to which the system approaches the failure boundary of a major safety accident. The data trust deficit quantification unit is used to receive the standardized time-series data stream in parallel, quantify the trust deficit of the construction data, and output the data trust deficit index. Specifically, the data trust deficit quantification unit is used to: establish a historical normal baseline model, which is established through statistical learning based on historical project data or a database of experience from similar projects, and is used to define the data pattern under normal construction conditions; compare the deviation between the standardized time-series data stream and the historical normal baseline model; when abnormally perfect data or highly regular submission times of manual self-inspection forms are detected, it is determined that the data is seriously inconsistent with the historical normal baseline model, thereby generating the data trust deficit index. The adversarial incentive modeling unit is used to receive and analyze business instructions from the project management, identify the root causes of the data trust deficit, and output an adversarial incentive pressure signal to the data trust deficit quantification unit. The adversarial incentive pressure signal is an adjustment signal used to characterize the strength of the data falsification motivation of the front-line construction team due to the management's incentive policies. The coupled decision arbitration unit is used to receive the physical risk assessment value and the data trust deficit index, make decisions based on the risk-trust dual assessment and antifragile decision closed-loop mechanism, and output high-priority intervention instructions. The antifragile strategy execution unit is used to receive the high-priority intervention instructions, execute antifragile actions to obtain real-world data, including assigning tasks to independent flight inspection teams, conducting random and dynamic surprise on-site inspections of subcontractors with the highest data suspicion, or forcibly installing tamper-proof third-party encrypted sensors; and feeding back the real-world data obtained through the surprise on-site inspections to the data trust deficit quantification unit for calibrating and resetting the data trust deficit index. The data trust deficit quantification unit is also used to: receive the adversarial incentive pressure signal from the adversarial incentive modeling unit; The antagonistic excitation pressure signal is used as a control parameter to dynamically adjust the deviation comparison logic; When the incentive pressure increases, the tolerance for perfect data is automatically reduced to make the data trust deficit indicator more sensitive. The coupled decision arbitration unit is specifically used to: determine that the data source is basically trustworthy when the data trust deficit index is lower than the preset blindness warning threshold; If the data source is deemed to be basically reliable, routine risk management is carried out based primarily on the physical risk assessment value output by the time-series risk simulation unit. The coupled decision arbitration unit is also used to: execute trust rejection logic when the data trust deficit index accumulates and reaches or exceeds the blindness warning threshold; The blindness warning threshold is generated based on statistical analysis and reverse calibration of historical data cases where data distortion caused assessment failure; When executing the trust rejection logic, regardless of whether the physical risk assessment value is safe or not, the physical risk assessment value is determined to be untrustworthy. After determining that the physical risk assessment value is unreliable, the system automatically switches to data authenticity reconstruction mode. After switching to the data authenticity reconstruction mode, the high-priority intervention command is output.

2. The tunnel blasting construction safety evaluation system as described in claim 1, characterized in that, The adversarial incentive modeling unit is specifically used to: receive and analyze business instructions containing aggressive financial incentive policies; The analysis examines how the aggressive financial incentive policy introduces a zero-sum game and motivates frontline staff to prioritize progress over compliance, thereby generating the adversarial incentive pressure signal.

3. A method for evaluating the safety of tunnel blasting construction, applied to a tunnel blasting construction safety evaluation system as described in any one of claims 1 to 2, characterized in that, The process includes the following steps: Step 1: Receive and standardize multi-source heterogeneous data through a multi-source heterogeneous data fusion unit to generate a standardized time-series data stream; Step 2: Receive the standardized time-series data stream through the time-series risk simulation unit and output the physical risk assessment value; Step 3: The standardized time-series data stream is received in parallel by the data trust deficit quantification unit, and combined with the adversarial incentive pressure signal from the adversarial incentive modeling unit, the data trust deficit index is output. Step 4: Receive the physical risk assessment value and the data trust deficit index through the coupled decision arbitration unit, make a decision based on the risk-trust dual assessment and antifragile decision closed-loop mechanism, and output a high-priority intervention instruction; Step 5: Receive the high-priority intervention instruction through the antifragile strategy execution unit, execute antifragile actions to obtain real-world data, and feed the real-world data back to the data trust deficit quantification unit for calibration and resetting the data trust deficit indicator.

4. The method for safety evaluation of tunnel blasting construction as described in claim 3, characterized in that, Step 4 specifically includes: when the data trust deficit index is lower than the preset blindness warning threshold, routine risk management is carried out based on the physical risk assessment value as the main decision-making basis; When the data trust deficit indicator accumulates and reaches or exceeds the blindness warning threshold, the trust denial logic is executed. After executing the trust rejection logic, the physical risk assessment value is determined to be unreliable; After determining that the physical risk assessment value is unreliable, the system automatically switches to data authenticity reconstruction mode. After switching to the data authenticity reconstruction mode, the high-priority intervention command is output.

Images