Method for predicting technical condition of tunnel civil engineering structure

Abstract

The invention provides a technical condition prediction method for a tunnel civil engineering structure, and relates to the technical field of traffic control monitoring systems.A biological acoustic composite sensing system comprising a distributed optical fiber acoustic sensing system and microorganism sample collection and analysis is constructed, and acoustic depth mode characteristics are mined by applying a deep learning model; the microbial dynamic biomarker is analyzed and identified by using bioinformatics; multi-modal information, a domain knowledge graph and a causal inference algorithm are fused, and a dynamic causal network model for revealing an internal driving relation in the degradation process is constructed; further, training and applying an adaptive neural network prediction model based on causal driving and fusing physical and biochemical mechanism constraints, and carrying out probabilistic prediction on future technical conditions of the tunnel; and executing anti-fact reasoning by using the causal network and the prediction model, and generating and optimizing an intervention strategy of cooperation of physical measures and biochemical measures. According to the method, the accuracy and advance of tunnel structure state prediction can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic control monitoring systems, and in particular to a method for predicting the technical condition of tunnel civil engineering structures. Background Art

[0002] There are already various technical means in the field of tunnel structure health monitoring and condition prediction. Traditional methods of regular manual inspections and non-destructive testing, such as crack observation, rebound method for strength measurement, ultrasonic testing, etc., although intuitive, have disadvantages such as strong subjectivity, high labor intensity, low efficiency, difficulty in comprehensive coverage, usually only being able to detect relatively significant damages, and limited prediction ability. Monitoring technologies based on point sensors, such as installing strain gauges, displacement gauges, crack gauges, piezometers, etc., can obtain physical quantity information at key points of the structure, but the number of sensor deployments is limited, the spatial coverage rate is low, it is difficult to capture the overall behavior of the structure and the distributed characteristics of damages, and the sensors themselves are vulnerable to environmental impacts and fail. Summary of the Invention

[0003] Technical Problems to be Solved In view of the deficiencies of the prior art, the present invention provides a method for predicting the technical condition of tunnel civil engineering structures, which solves the problems of the prior art.

[0004] Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: A method for predicting the technical condition of tunnel civil engineering structures, the prediction method comprising the following steps: Sp1: Construct a composite perception system of biological information and acoustic information and collect data; deploy a distributed fiber optic acoustic sensing system inside or on the surface of the tunnel structure, and set biological sample collection points for collecting seepage water, surface samples of the structure or internal fillers; use the distributed fiber optic acoustic sensing system to continuously collect acoustic or vibration wave field data distributed along the optical fiber; regularly collect biological samples, and through high-throughput sequencing or metabolite analysis technology, obtain biological information data characterizing the microbial community structure, functional gene composition or metabolic activity; at the same time, collect tunnel environmental parameters and conventional structure monitoring data; Sp2: Mine deep acoustic patterns and microbial dynamic biomarkers; apply a deep learning model to analyze the acoustic or vibration wave field data, and extract deep acoustic pattern features for reflecting structural micro-damage, stress changes or seepage states; perform bioinformatics analysis on the biological information data to identify dynamic change indicators of the microbial community structure or key functional genes and metabolite biomarkers related to tunnel environmental changes or structural deterioration states; Sp3: Construct a causal network by integrating multi-modal information; build a causal network model; integrate the deep acoustic pattern features, microbial dynamic biomarkers, environmental parameters, conventional monitoring data extracted by Sp2, and the pre-constructed knowledge graph of tunnel deterioration field; use causal inference algorithms to mine and construct a dynamic causal network model that reveals the interaction relationships between environmental factors, biochemical processes, physical and mechanical behaviors, and the evolution of the technical condition of the tunnel structure; Sp4: Train and apply an adaptive neural network prediction model; establish an adaptive neural network prediction model that at least includes: a) a physical information module constrained by partial differential equations of physical laws such as structural mechanics, fluid mechanics, or material damage; b) a biochemical information module constrained by a key biochemical reaction kinetics model, which is used to simulate processes such as microbial growth, biocorrosion, or chemical reaction equilibrium; c) a causal driving mechanism whose structure or information flow is guided by the causal network model constructed by Sp3; d) a mechanism for online parameter or structure adaptive adjustment according to new input data; use historical and real-time data to train the adaptive neural network prediction model; use the trained model to predict the future technical condition indicators and their probability distributions of the tunnel structure; Sp5: Generate an intervention strategy that coordinates physical and biochemical measures based on counterfactual reasoning; use the causal network model and the adaptive neural network prediction model to perform counterfactual reasoning, and simulate the impacts of different intervention scenarios, including physical repair measures, environmental regulation measures, or biological regulation measures, on the future structural technical condition and microbial environment; according to the results of counterfactual reasoning, combined with cost-benefit and risk assessment, generate and recommend an optimized intervention strategy that coordinates physical and biochemical measures.

[0005] Preferably, the high-throughput sequencing technology in Sp1 includes one or more of metagenomic sequencing, metatranscriptomic sequencing, or amplicon sequencing based on internal transcribed spacer or ribosomal ribonucleic acid genes.

[0006] Preferably, the deep learning model in Sp2 is selected from convolutional neural networks, recurrent neural networks, long short-term memory networks, transformer networks, graph neural networks, or their combinations, and is used to process acoustic wave field data or biological information sequence data.

[0007] Preferably, the causal inference algorithm in Sp3 is used to process mixed-type time series data and combine the prior causal relationships in the knowledge graph.

[0008] Preferably, the adaptive adjustment mechanism of the adaptive neural network prediction model in Sp4 adopts online learning, continuous learning, or meta-learning strategies.

[0009] Preferably, the biochemical reaction kinetics model included in the biochemical information module in Sp4 includes at least one of the following processes: Sp4.1, the process of microbially induced calcium carbonate precipitation; Sp4.2, the corrosion process caused by sulfate-reducing bacteria; Sp4.3, the influence of nitrification or denitrification processes on acidity; Sp4.4, the erosion process of organic acids on concrete.

[0010] Preferably, the environmental control measures or biological control measures in Sp5 include changing the chemical composition of seepage water, adjusting the temperature, humidity or ventilation conditions in the tunnel, or introducing beneficial microorganisms for bioremediation or bio-blocking.

[0011] Preferably, the prediction method further includes using the uncertainty information predicted by Sp4 to guide the sampling frequency, location of biological samples in Sp1 or the key monitoring areas of the distributed fiber optic acoustic sensing system.

[0012] Preferably, a tunnel civil engineering structure technical condition prediction system corresponding to the prediction method is characterized by including: A composite perception system of biological information and acoustic information; A data processing and biological information analysis unit; A central computing unit configured to execute the steps of the prediction method, including running a deep learning model, a causal inference algorithm, training and running the aforementioned adaptive neural network prediction model containing physical and biochemical information, performing counterfactual reasoning and generating collaborative intervention strategies; A user interface; A computer-readable storage medium storing a computer program, and the computer program is executed by a processor.

[0013] Beneficial effects: The present invention provides a method for predicting the technical condition of a tunnel civil engineering structure. It has the following beneficial effects: 1. By integrating the depth pattern recognition ability of the distributed fiber optic acoustic sensing system and the dynamic monitoring ability of microbial metagenomic dynamics, the present invention constructs a bio-acoustic composite perception system. This enables the system to not only capture the subtle acoustic features generated by the cumulative micro-damage, stress changes or seepage state evolution at the physical level of the structure, but also keenly sense the early biochemical fingerprint information of the material degradation process driven by biochemical actions. Compared with the technologies relying on traditional physical quantity monitoring or macroscopic inspection, this method can identify potential degradation risks earlier and more accurately.

[0014] 2. The present invention utilizes a causal inference algorithm, integrates multi-modal perception data, historical information, and a knowledge graph in the field of tunnel deterioration, and constructs a dynamic causal network model that reveals the interaction relationships among environmental factors, biochemical processes, and physical and mechanical behaviors. On this basis, an adaptive neural network prediction model is constructed, which further enforces the constraints of physical laws and key biochemical reaction kinetic models, ensuring the physical rationality and biochemical logic of the prediction results. It significantly improves the accuracy, interpretability, and robustness of the prediction model in the face of complex and variable working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a system architecture diagram of the present invention; Figure 2 is a system composition diagram of the present invention; Figure 3 is a system flow chart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Specific Embodiment 1: As Figures 1 - 3 shown, a method for predicting the technical condition of a tunnel civil structure, the prediction method includes the following steps: Sp1: Construct a composite perception system for biological and acoustic information and collect data; deploy a distributed fiber optic acoustic sensing system inside or on the surface of the tunnel structure, and set up biological sample collection points for collecting seepage water, structural surface samples or internal fillers; use the distributed fiber optic acoustic sensing system to continuously collect acoustic or vibration wave field data distributed along the fiber; regularly collect biological samples, and obtain biological information data characterizing the microbial community structure, functional gene composition or metabolic activity through high-throughput sequencing or metabolite analysis techniques; at the same time, collect tunnel environmental parameters and conventional structural monitoring data. The specific provisions for the collected environmental parameters and conventional structural monitoring data are as follows: Environmental parameters include but are not limited to: air temperature at a specific cross-section in the tunnel, air relative humidity, air pressure, carbon dioxide concentration, oxygen concentration, hydrogen sulfide concentration, chloride aerosol concentration, pH value of seepage water or structural surface condensate; conductivity of seepage water, dissolved oxygen content in seepage water, sulfate ion concentration in seepage water, chloride ion concentration in seepage water, water level elevation of groundwater around the tunnel structure, and temperatures inside and outside the lining structure. Structural monitoring data includes but is not limited to: three-dimensional absolute displacement monitoring values at characteristic points arranged on key sections of the tunnel lining (must include the crown, spandrel, side wall, and invert), obtained through an automated total station or a connected-tube static level system; clearance convergence value of the tunnel cross-section, obtained through a convergence meter or a laser profiler; vertical settlement value of the lining crown, strain values at embedded measuring points on the surface and inside the concrete structure in a specific section, obtained through a resistance strain gauge or a fiber Bragg grating strain sensor; stress or strain values of the main load-bearing steel bars, opening and offset amounts of the tunnel expansion joints or construction joints, lengths, maximum widths, average widths and three-dimensional spatial distribution forms of known main cracks, obtained through a high-precision crack width observation instrument or three-dimensional laser scanning technology; for shield tunnels, it is also necessary to collect the opening amount and shear displacement of the segment joints, as well as the pre-tightening force or stress of the segment connection bolts. The systematic collection and accurate recording of these parameters and data are the basis for subsequent multi-physical field coupling analysis, deterioration mechanism identification and prediction model training. High-throughput sequencing technologies include one or more of metagenomic sequencing, metatranscriptomic sequencing, or amplicon sequencing based on internal transcribed spacer or ribosomal ribonucleic acid genes.

[0018] Constructing a composite perception system for biological and acoustic information and collecting data lies in constructing a comprehensive perception network that can synchronously and accurately capture the dynamic changes of the internal micro-biological and chemical environment and the macroscopic and mesoscopic physical and mechanical responses of the tunnel structure, and ensuring the data quality and spatio-temporal correlation.

[0019] Sensing principle and technology selection: mainly using coherent optical time domain reflectometry (Φ-OTDR) or coherent optical frequency domain reflectometry (C-OFDR) technology based on Rayleigh scattering. Φ-OTDR senses vibration or dynamic strain along the optical fiber by detecting the interference phase change between the return light from different scattering points in the same pulse, and is suitable for long-distance, dynamic event monitoring. C-OFDR provides static or quasi-static strain measurement with extremely high spatial resolution (up to millimeter level) through swept frequency interferometry. Select appropriate demodulation technology and equipment according to the tunnel length and monitoring requirements (dynamic events vs. slowly changing strain).

[0020] Optical fiber layout and coupling: Select single-mode communication optical fiber (using G.652 or G.657) with good mechanical properties and transmission characteristics, or use special sensing optical fiber with optimized coating and enhanced backscattering as needed. The layout method must ensure good strain transfer coupling between the optical fiber and the tunnel structure, and use special adhesives such as epoxy resin to stick the optical fiber to the key path of the lining surface (using the longitudinal and transverse measurement lines of the vault, and both sides of the joints), or groove and bury the optical fiber with low modulus grouting material, or directly pre-bury it at a specific depth during the concrete pouring process. Special attention should be paid to joint protection, turning radius control, and optical fiber protection during construction to avoid damage and signal attenuation.

[0021] System performance and data acquisition: High-performance demodulators should be able to provide strain resolution at the nanostrain (nε) level. The spatial resolution can be selected from the centimeter level to the ten-meter level according to the technology type and configuration (usually there is a trade-off between distance and resolution). The frequency response range of dynamic strain measurement can cover from sub-hertz to tens of kilohertz. The demodulator needs to continuously or at a set high frequency (using a sampling rate of seconds or minutes) collect phase, frequency or time delay information on the all-fiber path. The amount of raw data is huge, and preliminary signal processing and noise reduction are usually required on-site or at the edge.

[0022] Detailed explanation of biological sample collection, processing and information acquisition: Sampling strategies and methods: Based on the tunnel's geological and hydrological conditions, disease distribution characteristics and monitoring objectives, a detailed sampling plan is developed to determine the sampling point locations, types (seepage water, surface swabs, rock and soil core samples) and frequencies (intensive sampling during the baseline period to establish the background, and then adjusted to monthly, quarterly or event-triggered sampling based on seasonal changes, rainfall events or structural state changes). Seepage water should be collected in sterile containers to avoid cross contamination; surface swabs should use standardized swabs and buffers; and coring should minimize disturbance to the in situ microbial community.

[0023] Sample Preservation and Nucleic Acid / Metabolite Extraction: After sample collection, the samples should be immediately placed at low temperature (-80°C), or added with specific nucleic acid protectants (RNADefender) or metabolite quenching agents to maintain the original state of biomolecules to the greatest extent. Mature kits (soil / water microbial DNA / RNA extraction kits) should be used for nucleic acid extraction in the laboratory, which includes cell lysis steps (mechanical bead milling, chemical lysis) and purification steps. For metabolite extraction, appropriate solvent systems (methanol / water, acetonitrile) and extraction methods should be selected according to the polarity of the target molecules.

[0024] Details of Sequencing and Metabolomics Analysis: High-throughput sequencing (using the Illumina NovaSeq platform) should ensure sufficient sequencing depth (reaching the Giga-base level for metagenomics, and tens of thousands of effective sequences for each sample in amplicon sequencing) to ensure coverage of low-abundance species and genes. Universal and highly discriminatory primer pairs (16S V3-V4 region) should be selected for 16S / ITS amplicon sequencing. For metabolomics analysis (using UHPLC-QTOF-MS), column selection, gradient elution program optimization should be carried out, and a standard library should be established or public databases (METLIN, HMDB) should be used for metabolite identification and quantification. Strict quality control (removing low-quality reads, chimeric sequences, and evaluating sequencing saturation) should be carried out after all bioinformatics data are generated.

[0025] Spatio-temporal Alignment and Management of Multi-source Data: Establish a unified tunnel health monitoring database (a relational database supporting spatio-temporal indexing or a dedicated time series database such as InfluxDB, TimescaleDB can be used). All data, including waveform data or feature data output by distributed fiber optic acoustic sensing demodulators (accompanied by fiber optic mileage and timestamps), biological sample analysis results (accompanied by the three-dimensional coordinates of the sampling point and sampling time), environmental sensor readings, and conventional monitoring instrument readings (displacement meter readings accompanied by the measurement point number and time), must be accurately aligned in terms of time and spatial reference coordinates and stored in a structured manner. Data cleaning, interpolation (handling missing values), filtering (removing noise), and normalization processes need to be designed; Sp2: Mining Acoustic Depth Patterns and Microbial Dynamic Biomarkers; Apply deep learning models to analyze the acoustic or vibration wave field data, extract depth acoustic pattern features for reflecting structural micro-damage, stress changes, or seepage states, and apply deep learning models to process the acoustic or vibration wave field data collected by distributed fiber optic acoustic sensing systems. The specific technical paths for the selection, processing, and label making of model training data are specified as follows: Sources of Model Training Data: 1. Raw waveform data records of distributed fiber optic acoustic sensing monitoring continuously collected from the target tunnel and similar tunnels with similar geological structures, surrounding rock characteristics, support types, and service environments at different operation stages (covering the structural health period, the period of known micro-damage gestation and development, the period of specific environmental disturbances such as heavy rain or adjacent blasting construction period, and after maintenance measures are taken); 2. Acoustic and acoustic emission signals synchronously and high-fidelity recorded during uniaxial or triaxial loading failure tests, fatigue loading tests, freeze-thaw cycle tests, acoustic emission monitoring tests, or chemical erosion tests on concrete, rock and other material specimens taken from the target tunnel or similar projects under standard laboratory conditions; 3. Using multi-physics field coupling numerical simulation software, the finite element method is combined with the fracture mechanics model to simulate the initiation, propagation, and penetration process of microcracks under different stress states.

[0026] Data processing includes the following steps: 2.1. Preprocess the raw acoustic waveform data. Use a digital bandpass filter (filter out the slow-varying signals below 1 Hz and the ultrasonic band noise above 20 kHz according to the expected frequency range of the target signal) to filter out the out-of-band noise, and apply the adaptive noise cancellation or Wiener filtering algorithm to suppress the in-band interference; 2.2. Segment the continuous acoustic data stream according to a fixed time window length (this length is set according to the specific monitoring target and can range from 0.1 second to several minutes), or use an event-triggered mechanism based on the mutation of statistics such as short-time energy, zero-crossing rate, and kurtosis for intelligent segmentation; 2.3. Perform feature transformation on each data segment, including calculating its short-time Fourier transform to obtain a time-frequency spectrogram.

[0027] Label making adopts the following hierarchical strategy according to the available information sources and accuracy: The first layer, strong label making. When there is high-precision ground truth information, through an independent and calibrated acoustic emission source localization system or synchronous endoscope observation and high-speed photography, it is confirmed that a microcrack event occurs at a specific position (accurate to the centimeter level) at a specific moment. Then, the acoustic data segment within the corresponding spatial range at that moment is assigned an accurate event category label ("Type A tensile microcrack", "Type B shear microcrack") and the occurrence time and spatial coordinates; The second layer, weak label making. When there is only damage information at the regional level or for a long time period, the engineering log records that new water seepage was observed in a certain tunnel section in the past week. Then, all acoustic data segments in this section during that week are initially assigned the weak label of "suspected period of leakage activity", and are refined later through methods such as multi-instance learning; The third layer is based on unsupervised learning for label-assisted generation. Models such as variational autoencoders or generative adversarial networks are used to train on a large amount of acoustical data with confirmed structural health to learn the distribution of normal patterns. For newly input data, if its reconstruction error is large or the probability of being recognized as a "fake" sample by the discriminator is high, it is labeled as "acoustical anomaly". Such abnormal data will be combined with other sensor information and handed over to domain experts for analysis, and preliminary event types or status labels will be assigned; The fourth layer is active learning and iterative labeling. The model selects samples with the most information value or the highest uncertainty for the current model's classification or regression task from the unlabeled data pool and submits them to the expert annotation platform. The expert annotates them based on their professional knowledge and referring to all available auxiliary information. The labeled data is added to the training set, and the model performs incremental training and iterates this process until the model performance reaches the preset metrics. Through the above rigorous data selection, processing, and multi-strategy label-making process, a high-quality training dataset is constructed to train a deep learning model so that it can reliably extract and quantify the deep acoustical pattern features that characterize structural microdamage, stress redistribution, or seepage path evolution from the input acoustical data; After that, bioinformatics analysis is performed on the biological information data to identify microbial community structure dynamic change indicators or key functional genes and metabolite biomarkers related to tunnel environmental changes or structural degradation states. The deep learning model is selected from convolutional neural networks, recurrent neural networks, long short-term memory networks, transformer networks, graph neural networks, or combinations thereof, and is used to process acoustical wavefield data or biological information sequence data; The core of mining deep acoustical patterns and microbial dynamic biomarkers lies in using advanced computational intelligence techniques to extract low-dimensional features with clear physical or biological significance that are closely related to the evolution of the structural health state from massive, high-dimensional, and heterogeneous raw data.

[0028] Acoustical deep pattern feature extraction includes: Model architecture and training: For the characteristics of distributed acoustical sensing data, it can be designed as follows: 1) One-dimensional convolutional neural networks directly process the time-domain waveform or frequency-domain representation along a certain point on the optical fiber; 2) Two-dimensional convolutional neural networks process the acoustical activity map or time-frequency spectrum map composed of time and space (optical fiber mileage) information; 3) Recurrent neural networks (especially long short-term memory networks or gated recurrent units) capture the time-evolution patterns of acoustical signals at a single location or multiple locations; 4) Transformer models use their self-attention mechanism to capture the long-range dependencies of acoustical events in time and space, which is especially suitable for analyzing global vibration mode changes or large event propagations; 5) The spatio-temporal graph convolutional network treats the optical fiber path as a graph structure and explicitly models the spatio-temporal interactions between adjacent or functionally related (located in the same structural unit) sensing points. During training, a suitable loss function (reconstruction error for anomaly detection, cross-entropy for event classification), a powerful optimizer (AdamW), and data augmentation techniques (adding noise, time-domain stretching, frequency-domain masking) should be used to improve the generalization ability of the model.

[0029] Feature instances and interpretations: The extracted deep acoustic mode features include: a low-dimensional embedding vector obtained from the bottleneck layer of the autoencoder, which can represent the main change patterns of the acoustic signal; the fractal dimension and Lyapunov exponent, which characterize the complexity or chaos degree of the signal; the energy proportion and its change rate in specific frequency bands (using the acoustic emission frequency band related to cracks and the low-frequency noise frequency band related to seepage); the two-dimensional or three-dimensional distribution map of the internal medium wave velocity of the structure (related to density, water content, and damage degree) obtained by wave velocity tomography inversion; and specific event patterns identified by the model (using micro-fracture characteristic waveforms and continuous resonance at specific frequencies).

[0030] Details of microbial dynamic biomarker identification: Details of the bioinformatics analysis process: After quality control of the original sequencing data, taxonomic unit operations (generating amplicon sequence variants ASVs or aligning to the reference genome for species identification) and functional annotation (mapping gene sequences to KEGG Orthology or Gene Ontology entries) are performed. Metabolite data are then detected, aligned, normalized, and identified. The core lies in finding biological features that are significantly correlated with time, space, environmental gradients, or known structural state hierarchies (by checking and evaluating).

[0031] Marker discovery and verification: Statistical methods such as DESeq2, edgeR, or non-parametric tests (Wilcoxon rank-sum test) are used to identify differentially abundant species, genes, or metabolites. Machine learning algorithms such as LEfSe (linear discriminant analysis effect size) or sparse learning-based methods (LASSO) are used to screen marker combinations with the ability to distinguish different states. Findings: As leakage intensifies, the abundance of sulfate-reducing bacteria involved in anaerobic respiration (sulfate reduction) increases significantly; in the concrete carbonation area, the abundance of basophilic bacteria decreases while that of neutrophilic bacteria increases; when steel bars start to corrode, the abundance of iron redox-related genes or the concentration of specific siderophore metabolites increases. These markers need to be verified in combination with literature knowledge and experiments to confirm their relevance to specific deterioration processes; Verification of the association between biomarkers and specific deterioration processes, with the combined literature knowledge and experimental verification specified as follows: The combination of literature knowledge involves a comprehensive and systematic search and analysis of scientific and technological literature databases, including but not limited to the Web of Science Core Collection database, Scopus citation database, PubMed biomedical literature database, as well as the American Chemical Society (ACS) journal database, Elsevier's ScienceDirect database, SpringerLink database, etc.; professional journals cover but are not limited to "Cement and Concrete Research", "Materials and Structures", "Corrosion Science", "Water Research", "Applied and Environmental Microbiology", "Environmental Science & Technology", "Geomicrobiology Journal", "Science of The Total Environment", etc. The core content of the search is the material deterioration processes (sulfate salt crystallization and chemical erosion, microbial-induced acid corrosion of concrete, microbial-influenced corrosion of reinforcing steel MIC, changes in material surface properties due to biofilm formation, nitrate erosion) driven or significantly accelerated by the metabolic activities of specific microbial groups (sulfate-reducing bacteria, sulfur-oxidizing bacteria, iron bacteria, nitrifying and denitrifying bacteria, acid-producing fungi and bacteria) in tunnel and similar underground engineering (mines, nuclear waste repositories) under specific geochemical environments (high sulfate, high chloride, acidic, rich in organic matter, strongly oxidizing or strongly reducing conditions).

[0032] Through bibliometric analysis, review reading, and intensive reading of key original research papers, microbial taxonomic units (accurate to the genus or species level), core functional genes (dsrAB gene encoding sulfate reductase, sox series genes encoding sulfur oxidation-related enzymes, genes encoding key enzymes involved in the organic acid synthesis pathway, genes encoding enzymes related to extracellular polymer synthesis), or characteristic metabolites (sulfides of different valence states, specific volatile or non-volatile organic acids, biosurfactants, isotope fractionation characteristics of specific elements) that have been confirmed by experimental or field research and have a clear indicative relationship with the occurrence, development stage, or severity of the above specific deterioration processes are extracted and solidified.

[0033] Experimental verification includes the following specific experimental types and operating procedures: Accelerated degradation simulation experiments under laboratory-controlled conditions, constructing a microreaction system, placing tunnel common concrete specimens (with clear mix proportions and material sources), steel bar specimens or geotechnical samples that have undergone standard curing in a reactor simulating the specific degradation environment of the target tunnel. This environment can precisely control temperature, humidity, gas composition, and the concentrations of key chemical substances (sulfate ions, chloride ions, organic carbon sources, specific pH buffers), and selectively inoculate pure cultures or mixed microbial communities of target microorganisms isolated and screened from the tunnel site or purchased from a standard microbial strain library.

[0034] During a preset experimental period (which can last from several weeks to several years), aseptically collect water samples, gas samples, biofilm samples on the material surface, and material bulk samples from the reaction system at regular intervals for high-throughput sequencing (metagenomics, metatranscriptomics, amplicon sequencing) to analyze the dynamic succession of microbial communities, the abundances and expression changes of key functional genes, and use techniques such as high-performance liquid chromatography-mass spectrometry to analyze the changes in metabolite profiles; simultaneously, conduct physical and mechanical property tests (compressive strength, flexural strength, elastic modulus), chemical property tests (pore solution ion concentration, carbonation depth), electrochemical tests (for steel bar specimens, such as corrosion potential, polarization resistance, alternating current impedance spectroscopy), and microstructure and phase analysis (by means of scanning electron microscopy combined with X-ray energy dispersive spectroscopy, X-ray diffraction, Fourier transform infrared spectroscopy, etc.) on the material specimens.

[0035] Through strict statistical correlation analysis, regression modeling, and time series analysis of changes in microbial indicators and material degradation indicators (mass loss rate, strength decline rate, corrosion current density, crack propagation length), establish a quantitative response relationship between the two.

[0036] In-situ verification and long-term positioning monitoring at the tunnel site. In the tunnel sections and their reference sections (without such biomarkers or with significantly different degradation degrees) that have been initially screened and identified for potential biomarkers, deploy a multi-parameter sensor array (including micro pH electrodes, redox potential electrodes, specific ion-selective electrodes, micro temperature and humidity sensors, gas sensors, etc.) that can conduct long-term in-situ monitoring and is matched with the laboratory research, and conduct periodic and standardized biological sample collection and laboratory analysis. At the same time, use advanced non-destructive testing techniques (such as ultrasonic tomography, electrical resistivity imaging, infrared thermography, laser speckle interferometry) in this area to finely image and quantify the internal damage of the structure, and supplement it with small-diameter core sampling for microscopic verification.

[0037] By conducting collaborative monitoring and data analysis on the dynamics of in-situ biomarkers, environmental factors, and the evolution of structural damage for several complete hydrological years or longer, further confirm the effectiveness, stability, and sensitivity of the correlations discovered in the laboratory in a real engineering environment, and finally determine the biomarker combination for early warning and its response threshold. Third, the application of advanced verification techniques such as stable isotope probing (SIP), under laboratory or controlled in-situ conditions, introduce specific substrates (such as specific organic acids, sulfates, nitrates, water) labeled with stable isotopes (¹³C, ¹ 5 N, ³ 4 S, ¹ 8 O), and by analyzing which members of the microbial community have their deoxyribonucleic acid, ribonucleic acid, or specific biomarkers (such as phospholipid fatty acid PLFA) isotopically labeled, directly identify which microorganisms are truly actively involved in the transformation of the target substrate and related deterioration processes under field conditions, providing the most direct functional-level evidence for biomarker selection. Through the in-depth excavation of the above-mentioned literature knowledge and the multi-level, multi-scale experimental verification system, ensure that the selected biomarker combination has high specificity, sensitivity, and reliability for the target deterioration process.

[0038] Sp3: Construct a causal network by integrating multi-modal information; construct a causal network model; integrate the deep acoustic pattern features, microbial dynamic biomarkers, environmental parameters, conventional monitoring data, and the pre-constructed knowledge graph of tunnel deterioration extracted by Sp2; use a causal inference algorithm to mine and construct a dynamic causal network model that reveals the interaction relationships between environmental factors, biochemical processes, physical and mechanical behaviors, and the evolution of the technical condition of the tunnel structure. The causal inference algorithm is used to process mixed-type time series data and combine the prior causal relationships in the knowledge graph; The fusion of multi-source input data, the output representation of the causal network model, the concretization of environmental factor indicators, and the construction of the knowledge graph of tunnel deterioration are as follows: The fusion processing flow of multi-source input data for causal inference is as follows: 3.1.1. Align data time and unify sampling rate. For the deep acoustic mode features extracted by Sp2 (the "frequency of occurrence of microcrack type A signals" and "average energy spectral density of seepage type B noise" output once per minute), microbial dynamic biomarkers (the "relative abundance of sulfate-reducing bacteria genus C" and "concentration of metabolite D" updated monthly or quarterly), environmental parameters collected by Sp1 (air temperature and seepage water pH value recorded hourly or daily), and conventional structural monitoring data (crack width and convergence deformation value recorded daily or weekly), all are unified to a preset analysis time scale through interpolation (for low-frequency data, linear interpolation or holding the previous value interpolation is used) or aggregation (for high-frequency data, the mean, maximum value, change rate, or specific statistic within a time window is used). This time scale is usually daily or weekly.

[0039] 3.1.2. Data normalization. For all aligned variables, the Z-standardization method, which is to divide by the standard deviation after subtracting the mean, or the min-max normalization method is used to linearly map them to the interval from 0 to 1, so as to eliminate the influence of different variable dimensions and value ranges on the causal inference algorithm.

[0040] 3.1.3. Feature matrix construction. All normalized variables (which all appear as time series at this time) are constructed into a feature matrix. Each row of this matrix represents an observation sample at a unified time step, and each column represents a specific feature variable. This feature matrix is the direct input data for the subsequent causal inference algorithm. The output representation of the causal network model is a structured graphical model, specifically a weighted directed graph. Each node of the weighted directed graph represents an input feature variable, such as "sulfate ion concentration in seepage water", "average temperature on the lining surface", "relative abundance of sulfate-reducing bacteria species X", "occurrence probability of acoustic mode Y", and "weekly expansion rate of crack at Z". The directed edges between nodes represent the directly inferred causal relationships between variables, and the direction of the arrow indicates the direction of causal influence.

[0041] Each directed edge is attached with the following quantitative attributes: 3.2.1. Causal effect strength. This strength is quantified by the standardized path coefficient in path analysis or the estimated parameter in structural equation modeling, indicating the average degree of change in the result variable caused by a unit change in the cause variable; 3.2.2. Confidence level or statistical significance of the causal effect. This level is represented by the p-value of the hypothesis test or the confidence interval obtained by the bootstrap method, indicating the reliability of the existence of this causal connection; 3.2.3. Time delay of the causal effect. This delay represents the average time span required for the result variable to start responding after the cause variable changes, and is measured in units of the analysis time scale.

[0042] The environmental factor indicators considered in the causal network model specifically include: air temperature, relative air humidity, carbon dioxide concentration in the air, oxygen concentration in the air, hydrogen sulfide concentration in the air at specific measuring points of the lining structure inside the tunnel; the pH value, redox potential, conductivity, dissolved oxygen concentration, total dissolved solids content, and concentrations of key cations and anions of seepage water or groundwater near the structure. The key cations include calcium ions, magnesium ions, sodium ions, potassium ions, and ferrous ions, and the key anions include sulfate ions, chloride ions, bicarbonate ions, and nitrate ions; it also includes the annual average rainfall, frequency of extreme rainfall events, and depth of the groundwater table around the tunnel structure where the tunnel structure is located.

[0043] The method for obtaining the knowledge graph of tunnel deterioration follows the steps below: 3.3.1. Core ontology construction: A multi-disciplinary expert team consisting of experts in tunnel engineering, structural engineering, geotechnical engineering, materials science, corrosion science, environmental microbiology, geochemistry, etc. collaborates. Based on international standards such as ISO 15926 and mature ontologies in related fields, they jointly define the core concept set of the tunnel deterioration field, namely entity types ("tunnel lining", "steel bar", "concrete", "sulfate", "pyrite", "sulfate-reducing bacteria", "crack", "corrosion product"), entity attributes ("concrete strength", "steel bar diameter", "crack width"), and relationship types between entities ("component of", "physical contact with", "exposed to", "chemically reacts with", "metabolically produces", "causes to occur", "accelerates the process", "inhibits the occurrence"), forming the top-level design of the domain ontology.

[0044] 3.3.2. Knowledge acquisition and integration: Structured and unstructured knowledge is systematically extracted from the following sources: first, authoritative textbooks, monographs, engineering design specifications, construction and acceptance standards, and operation and maintenance manuals; second, high-impact academic journal papers, doctoral dissertations, and proceedings of important international conferences; third, tunnel engineering accident investigation reports, long-term performance observation reports, and design and completion data of large-scale engineering projects. Natural language processing techniques (including named entity recognition, relation extraction, and event extraction) are used for preliminary information extraction from text resources, and data extraction and transformation tools are used to obtain knowledge from structured databases (such as material property databases, chemical reaction databases, and microbial gene databases). All extracted knowledge fragments need to be manually reviewed, verified, and disambiguated by domain experts.

[0045] 3.3.3. Knowledge Representation and Storage: The extracted knowledge is represented in triples (subject - predicate - object) or other forms of structured representation using the Resource Description Framework (RDF) or the Web Ontology Language (OWL), and stored in a dedicated graph database (Neo4j, GraphDB) to ensure the logical consistency, queryability, and inferability of the knowledge. The construction of the knowledge graph is a continuously evolving and dynamically updated process, which is continuously expanded and improved with the emergence of new knowledge and the deepening of cognition.

[0046] The purpose of constructing a causal network by integrating multi - modal information is to go beyond the phenomenon association. Using causal inference theories and algorithms, combined with domain knowledge, a structured causal model is constructed that can reflect the internal driving relationships and feedback loops among various factors in the tunnel deterioration system.

[0047] Deep Application of the Knowledge Graph: The knowledge graph not only provides prior constraints (using physically impossible causal relationships), but also enables: 1) Variable Selection and Representation: Guiding the selection of which monitoring variables, environmental factors, and biological indicators to include in the causal model and defining the potential relationship types between them; 2) Result Interpretation: Comparing the causal relationships discovered by data - driven methods with the known mechanisms in the knowledge graph to verify the rationality of the discoveries or discover new and not - yet - widely - recognized interactions; 3) Model Completion: In the case of sparse data or latent variables, using the deterministic knowledge in the knowledge graph to supplement or correct causal connections.

[0048] Causal Inference Algorithm Selection and Implementation Considerations: Considering the characteristics of the tunnel deterioration system (multi - variable, time - series, non - linear, potential feedback loops, existence of unobserved common causes such as geological tectonic activities), advanced causal inference techniques need to be selected. Dynamic Bayesian network learning algorithms combined with time - delay information, or unified structural equation models (uSEM) or dynamic causal models (DCM) based on state - space models that can handle non - linearity and feedback can be used. For dealing with latent variable problems, the FCI (Fast Causal Inference) algorithm and its variants can be considered. During implementation, it is necessary to note: Selecting appropriate conditional independence tests (using non - linear independence tests based on kernel methods) or scoring functions (using the Bayesian Information Criterion (BIC) that considers model complexity); dealing with multiple testing problems; and testing the robustness of the discovered causal relationships (using the Bootstrap method).

[0049] Quantification and Fusion of Causal Relationships: Not only the existence of causal connections needs to be discovered, but also the magnitude of the causal effect needs to be quantified (using regression coefficients of linear models or more complex non-linear models such as CausalForest to estimate the average treatment effect). Bayesian fusion of the data-driven discovered causal relationships (with confidence or probability) and the prior knowledge in the knowledge graph (represented as logical rules or probabilistic priors) is carried out to obtain a more credible posterior causal network model that is both based on data evidence and conforms to domain common sense.

[0050] Sp4: Train and apply an adaptive neural network prediction model; establish an adaptive neural network prediction model that at least includes: a) a physical information module constrained by partial differential equations of physical laws such as structural mechanics, fluid mechanics, or material damage; b) a biochemical information module constrained by a key biochemical reaction kinetics model, which is used to simulate processes such as microbial growth, biocorrosion, or chemical reaction equilibrium; c) a causal driving mechanism whose structure or information flow is guided by the causal network model constructed in Sp3; d) a mechanism for online parameter or structure adaptive adjustment according to new input data; train the adaptive neural network prediction model using historical and real-time data; use the trained model to predict the future technical condition indicators and their probability distributions of the tunnel structure. The adaptive adjustment mechanism of the adaptive neural network prediction model adopts online learning, continuous learning, or meta-learning strategies. The biochemical reaction kinetics model included in the biochemical information module includes at least one of the following processes: Sp4.1: Microbially induced calcium carbonate precipitation process; Sp4.2: Corrosion process caused by sulfate-reducing bacteria; Sp4.3: Influence of nitrification or denitrification processes on acidity; Sp4.4: Erosion process of organic acids on concrete; Data sources for model training: The core training data set comes from the target tunnel and a reference tunnel group with similar geological environments, structural forms, and deterioration patterns. Since its operation or since the deployment of this monitoring system, it is a multi-variable, long-time series fusion feature data set formed by continuously collecting through the Sp1 step and processing through the Sp2 and Sp3 steps.

[0051] The data set must cover the state records of the tunnel under different seasons, different environmental loads (such as traffic flow changes, extreme climate events), and different operation years, especially including those typical data segments that can clearly reflect the complete process of the structural performance from healthy to showing early signs of deterioration and then to the gradual development of damage.

[0052] If the historical observation data for a specific degradation mode (such as a rare type of chemical erosion) is insufficient, the data generated through the laboratory accelerated degradation simulation experiments or high-fidelity multi-physics numerical simulations mentioned in Sp1, after being aligned with the field data feature space, is used as supplementary training data to enhance the model's cognitive ability for low-probability high-risk events.

[0053] Data quality control is a prerequisite for selection. All data included in the training set must undergo strict quality assessment to eliminate data segments with obvious anomalies, distortions, or long-term missing data caused by sensor failures, data transmission errors, human interference, or force majeure factors.

[0054] Data processing aims to meet the model input requirements and improve the training efficiency and effect. The specific steps include: 1. Time series sliding window processing: The continuous multi-variable fusion feature time series is segmented according to the preset input window length (denoted as N_in time units) and prediction window length (denoted as N_out time units) to form a large number of "input sequence - target sequence" sample pairs. The input sequence contains all the fusion feature values in the past N_in time units, and the target sequence contains the actual observed values of one or more structural state indicators to be predicted in the future N_out time units. The lengths of N_in and N_out are determined through cross-validation based on the period (short-term, medium-term, long-term) of the prediction task and the autocorrelation characteristics of the data.

[0055] 2. Feature scaling and normalization: For each input feature variable and target prediction variable, within their respective training dataset ranges, independent Z-standardization (subtracting the mean and dividing by the standard deviation) or min-max scaling (linearly mapping to a specific interval such as -1 to 1 or 0 to 1) is performed to avoid potential adverse effects of different feature dimensions and numerical ranges on the model weight learning and to accelerate model convergence.

[0056] 3. Handling of imbalanced data: If the prediction target involves rare events (such as a specific type of sudden structural damage or reaching a certain extreme degradation threshold), the number of samples of this category in the training set will be much less than that of normal state samples. At this time, data balancing techniques must be adopted, including oversampling of minority class samples (SMOTE algorithm, i.e., synthetic minority over-sampling technique, or its variant ADASYN algorithm), undersampling of majority class samples (random undersampling or clustering-based undersampling), or introducing class weights in the loss function of model training (i.e., cost-sensitive learning) to ensure that the model has sufficient learning ability and prediction sensitivity for rare but critical events.

[0057] 4. Dataset division: Randomly divide all the prepared sample pairs into a training set, a validation set, and a test set according to a certain ratio (usually 70%-80% for training, 10%-15% for validation, and 10%-15% for final testing), ensuring that the original ratio of samples in each category is maintained during the division process (stratified sampling), and there is no overlap between the three datasets to ensure the objectivity and reliability of model evaluation.

[0058] The goal of training and applying the adaptive neural network prediction model is to build a computational brain with high predictive ability and adaptive ability that can simulate the complex behavior of the tunnel structure under the multi-field coupling of biology-chemistry-physics.

[0059] Model architecture and implementation details: The core of this model (formerly known as CA-BCPINN) is a hybrid, modular deep neural network. Its technical implementation includes: Input layer: Receive the multi-modal feature vectors (acoustic mode, biomarkers, environmental parameters, causal relationship features, etc.) processed by Sp2 and Sp3, as well as time and space coordinates. The adaptive neural network prediction model adopts a hybrid deep learning architecture integrating graph neural network and sequence learning module. The processing of spatial features is achieved through the graph neural network: 1. Discretize the tunnel structure in space into a set of nodes, which can represent predefined tunnel sections (such as one section every 10 meters), key structural components (such as specific segment rings, lining blocks), or the exact layout positions of sensors.

[0060] The connection relationships between nodes (i.e., the edges of the graph) are defined according to one or more of the following methods: Physical adjacency relationship, that is, establish connections between directly adjacent nodes in space; Mechanical influence relationship, that is, according to structural mechanics analysis, establish connections between nodes with strong correlations in force transmission or co-deformation; Causal relationship, that is, the causal connections between variables with significant spatial correlations (such as the change in the chemical composition of seepage water at point A leading to accelerated corrosion at point B) identified in the causal network model constructed by Sp3.

[0061] The attributes of each node at a specific moment are composed of the multi-modal feature vectors extracted by Sp2 (including acoustic mode features, biomarker abundances, environmental parameter values, and conventional structural monitoring data at this spatial position).

[0062] The graph neural network operates on this spatial graph structure at each time step, aggregates the information within the neighborhood of each node through the message passing mechanism, and thus learns and encodes the spatial dependence features and state distribution patterns of the structure at a certain moment.

[0063] The time coordinate can be input as one of the static attributes of the nodes at this stage. The processing of time features is achieved through a sequence learning module: the sequence of spatial feature vectors extracted by the graph neural network at each time step, or the time series of multimodal features of each node itself, is input into a dedicated sequence learning module, which adopts architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), or Transformer. To capture the dynamic laws, long-term dependencies, and periodic patterns of the evolution of each node (or the entire spatial state) over time.

[0064] The spatial coordinate can be used in the sequence learning module to adjust the learning weights of time series at different spatial positions or as position encoding information. The collaborative processing and fusion of spatio-temporal features are achieved as follows: first, spatial feature extraction is performed, and then time series learning is carried out. That is, at each time step, the graph neural network is applied to extract the spatial state representation of the entire structure, and then the time series of these representations is input into a recurrent neural network or a Transformer for time series prediction. Through this hierarchical or integrated architecture design, combined with the input time and spatial coordinate information (used for node identification, distance calculation, position encoding, etc.), the model can effectively learn from complex multimodal data and represent the spatio-temporal evolution law of the technical condition of the tunnel structure.

[0065] Encoder / Feature Extractor: It contains parallel sub-networks, which use Long Short-Term Memory to process time series features and use graph neural networks (if a graph structure is constructed based on tunnel segments or sensor networks) to process spatial correlation features.

[0066] Fusion Module: It adopts an attention mechanism (self-attention, cross-attention) to dynamically weigh the importance of information from different sources and at different time points for the current prediction.

[0067] Core Prediction / State Evolution Module: It is a recursive structure (adopting Neural ODE inspired by ordinary differential equation solvers) or a special network layer inspired by physical / biochemical equations (adopting convolutional layers that incorporate the idea of finite differences / finite elements into the network structure).

[0068] Implementation of Physical / Biochemical Constraints (Deepening the PINN Idea): During the training process, in addition to the loss term for fitting the observed data (using mean squared error), the key is to add loss terms for physical and biochemical constraints. This is achieved by randomly sampling a large number of configuration points (interior points, boundary points, initial points) in the spatio-temporal domain, using automatic differentiation to calculate the residuals of the network output at these points for relevant partial differential equations (using elasticity equations, diffusion equations) or ordinary differential equations (using reaction kinetics equations), and adding the norms of these residuals (using L2 norm) to the total loss function for minimization. In this way, the function learned by the network is forced to not only match the data but also satisfy the underlying governing equations.

[0069] Implementation of the causality-driven mechanism: The causality network implementation directly affects the model's structure or learning process. Structurally, the implementation designs a sparse connection pattern according to the causal graph or adopts a graph neural network architecture, whose message passing mechanism naturally follows the graph connection relationship. During the learning process, the implementation uses the causal effect strength as a regularization term to encourage the relationship weights learned by the model to be consistent with the causal strength; or suppresses the information flow that violates the causal direction during the gradient-based optimization process.

[0070] Implementation of the adaptive mechanism: Use online Bayesian inference methods (adopt online versions based on variational inference or Markov chain Monte Carlo) to update the probability distribution of the model parameters, enabling the model to continuously absorb new data and quantify prediction uncertainty. Alternatively, use continuous learning algorithms (adopt Elastic Weight Consolidation EWC, Gradient Episodic Memory GEM) to prevent catastrophic forgetting and protect important old knowledge when training on new data. Meta-learning then trains the model to "learn to learn" so that it can quickly adapt to the new environment using a small amount of new data.

[0071] Output and interpretation: The output should be multi-dimensional, including predicted values of future monitoring indicators (physical quantities, biological quantities), probability distributions (providing uncertainty ranges), probabilities or times to reach warning thresholds, risk level assessment maps, identification of potential dominant deterioration modes and their contribution degrees, etc. Combined with the causality network, the model can also provide a certain degree of interpretability, explaining which changes in input factors are the main driving forces for the prediction results. The prediction indicators directly output by the model include the following categories, and their specific contents are predefined according to the monitoring plan and analysis requirements: The predicted value of the physical quantity refers to the structural response or material state parameters at a specific future time point (or the average value of a time period), including: The three-dimensional spatial displacement vector at a specified monitoring point (such as point A at the crown of tunnel mileage K2+345); The predicted maximum width and length of a specified crack (such as crack number L007); The predicted average compressive strength of concrete in a specific section (such as lining ring number R105), in megapascals, or the predicted elastic modulus; The predicted annual average corrosion current density or predicted cross-sectional loss rate of key steel bars (such as steel bar S02).

[0072] The predicted value of the biological quantity refers to the microbial-related indicators at a specific future time point (or the average value of a time period), including: The predicted cell density of the target microbial group (total number of sulfate-reducing bacteria or specific corrosion-causing bacterial species) or its predicted relative abundance in the microbial community at a specified sampling point (such as leakage point W01); The predicted copy number of the key functional gene (dsrAB gene) at the sampling point; Predicted concentration of specific indicative metabolites (hydrogen sulfide or specific organic acids) at sampling points.

[0073] The output of the probability distribution refers to the directly predicted physical or biological quantity for each of the above. The model not only outputs its expected value (point prediction), but also outputs the complete parameter set of the probability distribution that the predicted value follows.

[0074] The method for obtaining the probability or time to reach the warning threshold is as follows: For any monitoring index with a defined warning threshold (crack width greater than 0.3 mm, or sulfate-reducing bacteria abundance exceeding 10^5 cells per milliliter), use the probability distribution function of this index at each future consecutive time step (such as daily, weekly) output by the model; calculate the cumulative probability that the index value exceeds its preset warning threshold at each future time step, so as to obtain a sequence of "at the t-th future time unit, the probability that index X exceeds the threshold is P%"; by setting an acceptable probability level (such as 95% confidence level), find the first time point at which the upper confidence limit of the index prediction value exceeds the warning threshold, and this time point is the predicted "time to reach the warning threshold"; in this case, the training label is the time interval when each index first reaches its threshold in the historical data. The method for identifying the potential dominant deterioration mode and obtaining its contribution degree is as follows: 1. By setting parallel multi-task learning heads in the output layer of the adaptive neural network prediction model, one task head is responsible for predicting the above physical and biological quantities, and another independent classification task head directly outputs the classification discrimination of the most likely dominant deterioration mode for the current or future specific section of the tunnel. The category is selected from one or more of the predefined deterioration mode libraries (including "sulfate chemical erosion", "microbially induced concrete acid corrosion", "chloride-induced corrosion of steel bars", "freeze-thaw damage of lining", "alkali-aggregate reaction"), and gives the activation probability or confidence score of each mode.

[0075] 2. Use posterior model interpretation techniques, such as the SHAP value analysis method (SHapley Additive exPlanations), the LIME analysis method (Local Interpretable Model-agnostic Explanations), or the integrated gradient method, to analyze the trained adaptive neural network prediction model. These techniques can calculate the contribution value or importance score of each input feature (including specific acoustic pattern features extracted by Sp2, microbial dynamic biomarkers, and environmental parameters collected by Sp1) to a specific structural state index deterioration (the increased value of the crack propagation rate, the predicted value of the corrosion depth) in the final prediction of the model. If the analysis results show that the input feature combination closely related to the sulfate erosion mechanism (high sulfate ion concentration, high abundance of sulfate-reducing bacteria biomarkers, and specific acoustic patterns indicating internal expansion stress) has the highest cumulative contribution value to the predicted structural damage, then "sulfate erosion" is determined to be the current or future dominant deterioration mode, and its contribution degree is quantified by the normalized cumulative contribution value of these features or the total SHAP value. At the same time, in the causal network model constructed by Sp3, the cumulative effect intensity of each causal path from a specific initial cause (such as "high-concentration chloride salt environment") to the final structural damage effect (such as "steel bar cross-section loss") also provides a direct mechanistic basis for judging the dominant deterioration mode and its relative contribution.

[0076] Sp5: Generate an intervention strategy that coordinates physical and biochemical measures based on counterfactual reasoning; use the causal network model and the adaptive neural network prediction model to perform counterfactual reasoning, and simulate the impact of different intervention scenarios, including physical repair measures, environmental control measures, or biological control measures, on the future structural technical condition and microbial environment; according to the counterfactual reasoning results, combined with cost-benefit and risk assessment, generate and recommend an optimized intervention strategy that coordinates physical and biochemical measures. Environmental control measures or biological control measures include changing the chemical composition of seepage water, adjusting the temperature, humidity, or ventilation conditions in the tunnel, or introducing beneficial microorganisms for bioremediation or bio-blocking; use the uncertainty information predicted by Sp4 to guide the sampling frequency, location of biological samples in Sp1, or the key monitoring areas of the distributed fiber optic acoustic sensing system.

[0077] Generating an intervention strategy that coordinates physical and biochemical measures based on counterfactual reasoning is a key step in transforming the predictive insights of the model into an executable and optimizable maintenance action plan.

[0078] Counterfactual Reasoning Engine: Based on the theory of structural causal models (using Pearl's causal ladder), it performs counterfactual inference using a trained causal network and an adaptive prediction model. The specific operation is as follows: Given an observed current state, first perform "abduction", that is, infer the probability distribution of the external causes (exogenous variables) of the system that led to the current state; then execute "action", that is, modify the graph structure or variable values according to the intervention measures to be simulated in the causal graph model; finally, perform "prediction", that is, in the modified model, use the forward propagation of the adaptive neural network to calculate the predicted distribution of the target variable after the intervention (using the structural state index at a future moment).

[0079] Collaborative Intervention Strategy Simulation and Optimization: The system can simulate a variety of complex intervention scenarios and evaluate the effects of implementing physical reinforcement alone, environmental regulation alone (by changing the tunnel ventilation mode to reduce humidity), biological intervention alone (by spraying specific antibacterial agents), and various collaborative intervention schemes of pairwise combinations or three-way combinations. For each scheme, simulate its multi-dimensional impacts on the long-term performance of the structure (using remaining life, safety), maintenance costs (including initial investment and long-term effects), and even environmental impacts (using chemical agent usage). Then, use multi-objective optimization techniques (using NSGA-II, MOEA / D) or methods combined with the decision-maker's preferences (using weighted summation, constraint method) to balance multiple objectives such as cost, risk, and benefit, and search for the Pareto-optimal set of intervention strategies to provide a quantitative basis for the final decision. Specific Embodiment 2: Adopt Figures 1 - 3 As shown, a tunnel civil engineering structure technical condition prediction system corresponding to the prediction method includes: Compound perception system of biological information and acoustic information; Data processing and biological information analysis unit; Central computing unit, configured to execute the steps of the prediction method, including running a deep learning model, a causal inference algorithm, training and running the aforementioned adaptive neural network prediction model containing physical and biochemical information, performing counterfactual reasoning, and generating collaborative intervention strategies; User interface; Computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program is executed by a processor. Specific Embodiment 3: Based on the technical solutions of Specific Embodiments 1 and 2, a further application case is given: In a case of a highway tunnel passing through coal measures strata and facing a serious risk of sulfate erosion, the complete application of this prediction method aims to accurately evaluate and predict sulfate erosion induced by pyrite oxidation and structural degradation under the synergistic action of sulfate-reducing bacteria (SRB), and formulate targeted prevention and control strategies. In the Sp1 stage, the construction of the perception system and data collection focus on capturing the key driving factors of sulfate erosion and structural responses. In addition to deploying a conventional distributed fiber optic acoustic sensing system along the longitudinal direction of the tunnel and at key cross-sections, fiber optics are particularly densely deployed in the contact sections adjacent to coal seams or pyrite-rich surrounding rocks, known water leakage points, and structural stress concentration areas, and coherent optical time domain reflectometry is used to capture high-frequency acoustic emission signals of microcrack propagation and strain accumulation caused by sulfate crystal expansion or corrosion. Biological sample collection points are preferentially set at the water leakage outlets and wet lining surfaces in these high-risk areas, and the sampling frequency is dynamically adjusted according to rainfall and groundwater level changes (increased frequency after the rainy season). Samples are analyzed by metagenomic sequencing to focus on the dynamic changes in the abundances of sulfur cycle-related microorganisms such as sulfate-reducing bacteria (using Desulfovibrio) and sulfur-oxidizing bacteria and their functional genes (especially key genes in the sulfate reduction pathway such as dsrAB); metabolomics focuses on metabolites such as sulfides and short-chain fatty acids; at the same time, high-frequency hydrochemical analysis (sulfate ion, ferrous ion, pH value, redox potential) and chemical composition analysis of surrounding rock samples are important components of the perception system in this case. In the Sp2 stage, the mining of acoustic depth patterns and microbial dynamic biomarkers focuses on the characteristic signals of sulfate erosion. A deep learning model (using a two-dimensional convolutional neural network based on time-frequency spectrograms) is trained to identify unique high-frequency, short-duration acoustic emission signal clusters in acoustic data caused by sulfate crystal expansion stress, and patterns of slight decreases or increased attenuation in the acoustic wave propagation speed due to internal damage accumulation. Bioinformatics analysis aims to identify activity indicators of sulfate-reducing bacteria (using the relative abundance change rate of the dsrAB gene) and trends in the concentration changes of specific metabolites that are significantly correlated with sulfate concentration gradients or structural damage degrees through differential abundance calculation and biomarker screening algorithms (using LEfSe), which will serve as early warning signals for the biochemical process of sulfate erosion. In the Sp3 stage, the core of constructing a causal network by integrating multi-modal information is to clarify the dominant degradation path in this specific scenario.Integrate acoustic damage characteristics, sulfate-reducing bacteria activity indicators, hydrochemical data (sulfate concentration is the key variable), environmental temperature and humidity, and structural strain monitoring data, and combine with a domain knowledge graph containing the kinetic of pyrite oxidation reaction, the chemical reaction mechanism of sulfate erosion of concrete, and the metabolic conditions of sulfate-reducing bacteria. Use the dynamic causal inference algorithm (adopt dynamic Bayesian network learning considering time delay) to construct and quantify the core causal chain of "surrounding rock seepage (carrying oxygen) → pyrite oxidation → increase in sulfate concentration → create anaerobic environment + provide sulfate → active proliferation of sulfate-reducing bacteria → generation of sulfide / acidic substances → accelerate steel corrosion / concrete deterioration → abnormal acoustic signals → macroscopic structural damage", and evaluate the regulatory effects of each link by environmental factors (using temperature, pH). In the Sp4 stage, the goal of training and applying the adaptive neural network prediction model is to establish a prediction engine that can simulate the coupled process of sulfate diffusion-reaction-corrosion. This model enforces the partial differential equation for the transport of sulfate ions in concrete pores (adopt Fick's second law considering adsorption and reaction terms) and the relevant constraints for the calculation of sulfate crystallization expansion stress in the physical information module; the biochemical information module embeds the kinetic equations describing the growth of sulfate-reducing bacteria (affected by the concentration of sulfate, organic matter, inhibitor) and the sulfide generation rate. The causal drive mechanism will particularly strengthen the influence of the identified key driving factors (using sulfate concentration, sulfate-reducing bacteria abundance) on the prediction of structural damage indicators (using steel corrosion rate, concrete strength loss rate, crack propagation rate). The adaptive ability of the model is reflected in its ability to quickly adjust internal parameters according to the real-time monitored changes in the groundwater level or the sudden infiltration of high-concentration sulfate wastewater, and update the prediction of the future corrosion rate and the remaining life of the structure. In the Sp5 stage, the intervention strategy that generates the synergy of physical and biochemical measures based on counterfactual reasoning is committed to finding the most efficient prevention and control plan. Using the trained causal network and prediction model, conduct a counterfactual simulation of "How will the average erosion depth and maintenance cost change after five years if high impermeable concrete replacement is carried out on specific sections of the tunnel and sulfate-reducing bacteria inhibitors are regularly put into the seepage water source area?" By systematically evaluating the long-term effects and cost-effectiveness of single measures (using optimized drainage, surface coating protection, electrochemical repair) and combined measures (using source plugging combined with process inhibition), finally recommend the comprehensive synergy strategy of "implementing curtain grouting to cut off the sulfur-rich water source for the main leakage sources + carrying out structural repair and applying high-performance anti-corrosion coatings to the eroded sections + monitoring and timely intervening in the activity of sulfate-reducing bacteria". Specific Embodiment Four: Based on the technical solutions of Specific Embodiments One and Two, further give an application case: For an existing highway tunnel located in mountainous areas, experiencing significant seasonal freeze-thaw cycles and having complex leakage problems, it mainly faces the risks of material fatigue damage, crack propagation, and increased leakage caused by freeze-thaw cycles. The application goal of this prediction method here is to deeply understand the process of freeze-thaw damage accumulation and the dynamics of leakage, predict the decline of structural performance in advance, and guide the implementation of effective waterproofing, drainage, and repair measures. In the Sp1 stage, the construction of the perception system and data collection need to comprehensively capture the temperature field, water activity, and structural response. The distributed fiber optic sensing system is not only used for acoustic monitoring but also can integrate the distributed temperature sensing (DTS) function, which is deployed longitudinally, transversely, and at different depths along the tunnel to obtain the refined three-dimensional temperature field inside and around the lining and its changes with seasons and day and night. Acoustic monitoring (DAS) focuses on the portal section, water-bearing areas or tectonic influence zones behind the lining, and near known leakage points to capture the acoustic emissions of microcracks caused by frost heave stress, the flow noise of meltwater, and the impact of freeze-thaw damage accumulation on the dynamic characteristics of the structure. The biological sample collection points are mainly set at stable leakage points and seasonal water outlet points at different elevations. By using metagenomic sequencing to analyze the microbial community structure, it can assist in judging the source of leakage water (using specific microbial community characteristics to distinguish between atmospheric precipitation recharge and deep groundwater recharge) and the flow path. At the same time, high-precision displacement gauges, crack gauges, and pore water pressure gauges are important conventional monitoring means in this case and need to be synchronously collected with environmental data such as temperature, humidity, and precipitation. In the Sp2 stage, the excavation of acoustic depth patterns and microbial dynamic biomarkers aims to identify the characteristics of freeze-thaw damage and leakage activities. The deep learning model (using a network combining convolutional and recurrent structures) needs to be trained to identify different types of acoustic emission signals generated by the formation, expansion, and melting of ice crystals in acoustic data (distinguishing tensile failure and shear failure), as well as the systematic drift of acoustic wave velocity, attenuation, or modal parameters (using natural frequency, damping ratio) caused by the change of the internal microstructure of the material due to repeated freeze-thaw. Bioinformatics analysis mainly serves for water source discrimination and water-rock interaction assessment. By comparing the similarities and differences of microbial communities at different leakage points (using beta diversity analysis) and combining hydrochemical data, it is possible to infer the source, mixing ratio, and residence time of leakage water. The abundance change of some psychrophilic microorganisms can also indirectly reflect the degree of freeze-thaw activity. In the Sp3 stage, constructing a causal network by fusing multi-modal information needs to clarify the complex coupling relationship between temperature - water - force - damage.Integrate temperature field data, distributed acoustic features (acoustic emission rate, wave velocity change, modal parameters), leakage water volume, water chemistry and microbial source information, and structural deformation data. Combine the knowledge graphs of freeze-thaw physics (water phase change, frost heave pressure), damage mechanics, and hydrogeology. Use causal inference methods that can handle strong nonlinearity and multi-physical field coupling (adopt state space models or process-based simulation models and data assimilation) to construct a dynamic causal network that describes "external temperature cycle → structural temperature field change → pore water freezing / melting → generation of frost heave / subsidence stress → initiation and propagation of microcracks (acoustic emission) → deterioration of material properties (change in wave velocity / modes) → increase in permeability / development of leakage paths → macroscopic deformation / stability decline", and incorporate the differences in the impacts of different water sources on leakage and structures. In the Sp4 stage, train and apply an adaptive neural network prediction model, which requires the model to have the ability to simulate the thermal-hydro-mechanical-damage coupling process. The physical information module needs to force the model output to satisfy a system of partial differential equations composed of the heat conduction equation, unsaturated / saturated seepage equation (using the Richards equation), and elastoplastic or damage mechanics constitutive equations considering phase change. An empirical or semi-empirical model describing the relationship between the number of freeze-thaw cycles and material damage accumulation needs to be introduced as a constraint. The causal driving mechanism should be able to adjust the sensitivity of the prediction of the damage accumulation rate according to the identified dominant factors (using the duration of negative temperature, water content). The self-adaptability of the model needs to enable it to accurately respond to the immediate impacts of extreme weather events (using sudden strong cold snaps, continuous heavy rains) on the temperature field, seepage field, and structural stress state, and update the long-term damage evolution prediction. In the Sp5 stage, generate an intervention strategy that coordinates physical and biochemical measures based on counterfactual reasoning, aiming to provide an optimized solution for freeze-thaw resistance and comprehensive water control. Use the model to simulate the effects of different intervention measures and ask the question: "If an insulation layer is applied to the outside of the lining of the tunnel portal section and the main leakage channels are drained, how much can the cumulative amount of freeze-thaw damage and maintenance costs of the structure be reduced in the next ten years compared to only repairing the surface cracks?" By comparing the simulation results of single or combined measures such as surface protection (hydrophobic agent, anti-freeze coating), structural insulation, crack treatment (different materials and processes), optimized drainage system (water interception, drainage, combination of blocking and drainage), and even changing the hydrogeological conditions of the surrounding rock (using curtain grouting), finally recommend a comprehensive treatment strategy of "applying external insulation to the easily frozen sections + installing a perfect internal drainage system + repairing the existing damaged cracks with low-temperature and high-elasticity materials".

[0083] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variation thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0084] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A prediction method for the technical condition of tunnel civil engineering structures, characterized in that: The prediction method includes the following steps: Sp1: Construct a composite perception system of biological information and acoustic information and collect data; deploy a distributed fiber optic acoustic sensing system inside or on the surface of the tunnel structure, and the distributed fiber optic acoustic sensing system continuously collects acoustic or vibration wave field data distributed along the fiber optic; regularly collect biological samples, collect tunnel environmental parameters and conventional structure monitoring data; Sp2: Mine acoustic depth patterns and microbial dynamic biomarkers; use a deep learning model to analyze the acoustic or vibration wave field data, and extract depth acoustic pattern features for reflecting structural micro-damage, stress changes or seepage states; Sp3: Integrate multi-modal information to construct a causal network and construct a causal network model; integrate the depth acoustic pattern features, microbial dynamic biomarkers, environmental parameters, conventional monitoring data extracted in Sp2, and a pre-constructed tunnel deterioration domain knowledge graph; Sp4: Train and apply an adaptive neural network prediction model; establish an adaptive neural network prediction model, and the adaptive neural network prediction model at least includes: a) A physical information module constrained by partial differential equations of physical laws such as structural mechanics, fluid mechanics or material damage; b) A biochemical information module constrained by a key biochemical reaction kinetic model; c) A causal driving mechanism in which the structure and information flow of the adaptive neural network prediction model are guided by the causal network model constructed in Sp3; d) A mechanism for online parameter or structure adaptive adjustment according to new input data; use historical and real-time data to train the adaptive neural network prediction model; use the trained model to predict the future technical condition indicators and their probability distributions of the tunnel structure; Sp5: Generate an intervention strategy that coordinates physical and biochemical measures based on counterfactual reasoning; use the causal network model and the adaptive neural network prediction model to perform counterfactual reasoning and generate an intervention strategy.

2. The prediction method for the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that, The high-throughput sequencing technology in Sp1 includes one or more of metagenomic sequencing, metatranscriptomic sequencing, or amplicon sequencing based on internal transcribed spacer or ribosomal ribonucleic acid genes.

3. A method for predicting the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that, The deep learning model in Sp2 is selected from one or more combinations of convolutional neural networks, recurrent neural networks, long short-term memory networks, transformer networks, and graph neural networks, and is used to process acoustic wave field data or biological information sequence data.

4. A method for predicting the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that, The causal inference algorithm in Sp3 is used to process mixed-type time series data and combine the prior causal relationships in the knowledge graph.

5. A prediction method for the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that, The adaptive adjustment mechanism of the adaptive neural network prediction model in Sp4 adopts one of online learning, continuous learning or meta-learning strategies.

6. The prediction method for the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that The biochemical reaction kinetic model included in the biochemical information module in Sp4 at least includes one of the following processes: Sp4.1: The process of microbial-induced calcium carbonate precipitation; Sp4.2: The corrosion process caused by sulfate-reducing bacteria; Sp4.3: The influence of nitrification or denitrification processes on the acidity; Sp4.4: The erosion process of organic acids on concrete.

7. A prediction method for the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that The environmental control measures or biological control measures in Sp5 include changing the chemical composition of seepage water, regulating the temperature, humidity or ventilation conditions in the tunnel, or introducing beneficial microorganisms for bioremediation or bio-blocking.

8. A method for predicting the technical condition of a tunnel civil engineering structure according to claim 1, characterized in that, The prediction method further includes using the uncertainty information predicted by Sp4 to guide the sampling frequency, location of biological samples in Sp1 or the key monitoring areas of the distributed fiber optic acoustic sensing system.

9. A prediction method for the technical condition of a tunnel civil engineering structure according to any one of claims 1-8, and a prediction system for the technical condition of a tunnel civil engineering structure corresponding to the prediction method, characterized in that, Including: A composite perception system of biological information and acoustic information; A data processing and biological information analysis unit; A central computing unit configured to execute the steps of the prediction method, including running a deep learning model, a causal inference algorithm, training and running the aforementioned adaptive neural network prediction model containing physical and biochemical information, performing counterfactual reasoning and generating a collaborative intervention strategy; A user interface; A computer-readable storage medium storing a computer program, which is executed by a processor.

Images