A method for identifying hidden risks, tracing causes and predicting evolution of pier cracks

Abstract

The application relates to the technical field of bridge structure health monitoring, and provides a bridge pier crack implicit risk identification, cause tracing and evolution prediction method, which comprises the following steps: obtaining multi-source monitoring data of a bridge pier, determining a dynamic fusion weight based on historical quality indexes of each data source, and fusing to obtain a multi-source feature set; determining the contribution probability of each implicit risk factor to the crack based on the multi-source feature set and a causal inference model to obtain a cause analysis result; taking the physical degradation law of the bridge pier as a constraint condition and integrating the constraint condition into a data-driven model, and predicting the evolution trend of the crack based on the multi-source feature set and the constrained data-driven model; and generating a risk assessment result and early warning information based on the evolution trend and the cause analysis result. The application realizes early identification of the implicit risk of the bridge pier crack, accurate diagnosis of the crack cause and physically consistent prediction of the crack evolution trend, and simultaneously improves the early warning preposition and attribution.

Description

Technical Field

[0001] This application relates to the field of bridge structural health monitoring technology, specifically to a method for identifying hidden risks, tracing causes, and predicting the evolution of cracks in bridge piers. Background Technology

[0002] In the field of bridge structural health monitoring, bridge piers, as key load-bearing components, are often the result of the coupled effects of multiple factors such as concrete hydration heat, settlement and creep, load fatigue, and environmental erosion. Traditional crack detection methods mostly rely on sensors to detect cracks only after they have developed to a visible stage, resulting in a significant detection lag and making it difficult to detect latent damage risks inside the component that have not yet been exposed in the early stages.

[0003] Existing technologies typically rely on single sensor threshold judgments or periodic manual inspections to identify cracks. This makes it difficult to leverage long-term, multi-source monitoring data for in-depth analysis of the complex causes of crack formation, leading to misjudgments or missed detections in scenarios with multiple coupled factors. Furthermore, while some technologies have begun to incorporate big data and artificial intelligence algorithms for crack trend prediction, existing purely data-driven models are prone to divergence or producing predictions that violate physical principles in long-term, small-sample predictions, such as propagation rates exceeding material limits. The lack of effective physical mechanism constraints results in insufficient reliability of predictions. This makes it difficult to predict risks and diagnose causes when cracks are in their latent stages, hindering early intervention by maintenance personnel and potentially causing latent risks to worsen into visible damage.

[0004] Therefore, it is necessary to provide a method for identifying the hidden risks, tracing the causes, and predicting the evolution of cracks in bridge piers to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the difficulties in identifying latent risks and tracing the causes of cracks in bridge piers in existing technologies, and the lack of physical constraints in pure data-driven prediction models leading to poor early warning and attribution, this application proposes a method for identifying latent risks, tracing the causes, and predicting the evolution of cracks in bridge piers. This method enables early identification of latent risks of cracks in bridge piers, accurate diagnosis of crack causes, and physical consistency prediction of crack evolution trends.

[0006] To achieve the above objectives, this application adopts the following technical solution: This application provides a method for identifying latent risks, tracing causes, and predicting the evolution of cracks in bridge piers. The method includes: acquiring multi-source monitoring data of the bridge piers; determining dynamic fusion weights based on historical quality indicators of each data source; fusing the multi-source monitoring data based on the dynamic fusion weights to obtain a multi-source feature set; determining the contribution probability of each latent risk factor to the cracks in the bridge pier based on the multi-source feature set and a causal inference model to obtain the causal analysis results of the cracks; incorporating the physical degradation law of the bridge piers as a constraint into a data-driven model; predicting the evolution trend of the cracks based on the multi-source feature set and the constrained data-driven model; and generating risk assessment results and early warning information for the bridge piers based on the evolution trend and the causal analysis results.

[0007] Preferably, historical quality indicators of each data source within a historical time period are obtained, and the confidence level of each data source is calculated based on the historical quality indicators. The confidence level of each data source is normalized, and the normalized confidence level is determined as the dynamic fusion weight. The time-series features, image features, and environmental features in the multi-source monitoring data are weighted and fused based on the dynamic fusion weight to obtain the multi-source feature set. A causal graph containing multi-source parameter nodes and crack nodes is constructed. The conditional probability of the causal graph is updated based on the multi-source feature set, and the contribution probability of each latent risk factor to the crack is calculated based on the updated causal graph to obtain the causal analysis result of the crack.

[0008] Preferably, the causal graph is constructed using a Bayesian network or decision tree algorithm, wherein the multi-source parameter nodes include load nodes, environment nodes, and structural response nodes, and the crack nodes include crack sub-nodes of different types; based on the multi-source feature set, the conditional probability table between the multi-source parameter nodes and the crack nodes is calculated using the maximum likelihood estimation method; the multi-source feature set at the current time is input into the causal graph, and the posterior probability of each crack sub-node is calculated through Bayesian inference, the posterior probability is used as the contribution probability, and the causal analysis result of the crack is output, the causal analysis result including the probability distribution corresponding to each causal type, and the event that triggered the crack.

[0009] Preferably, the physical degradation law is transformed into a physical residual penalty term, which is added to the loss function of the data-driven model. The parameters of the data-driven model are updated by minimizing the loss function containing the physical residual penalty term, resulting in a soft-constrained data-driven model. Alternatively, the physical degradation law is transformed into a physical boundary truncation condition. After the data-driven model outputs an initial prediction result, the initial prediction result is truncated and corrected based on the physical boundary truncation condition, resulting in a hard-constrained data-driven model. Based on the multi-source feature set, the evolution trend of the crack is predicted using the soft-constrained data-driven model and / or the hard-constrained data-driven model.

[0010] Preferably, the material fatigue damage equation and / or structural creep model of the bridge pier are obtained as the physical degradation law; a time-series neural network is constructed as the data-driven model, the time-series neural network including a long short-term memory network layer and a fully connected output layer; the physical evolution residual of the crack between adjacent time steps is calculated based on the physical degradation law, and the sum of squares of the physical evolution residuals is used as the physical residual penalty term; a joint loss function including a data prediction error term and the physical residual penalty term is constructed, the gradient of the joint loss function is calculated based on the historical monitoring dataset, and the weight parameters of the time-series neural network are updated using the backpropagation algorithm to obtain the soft-constrained data-driven model.

[0011] Preferably, the current severity index of the crack, the predicted growth rate index determined based on the evolution trend, and the structural anomaly index determined based on the multi-source feature set are obtained; a risk index is obtained by weighting and summing the current severity index, the predicted growth rate index, and the structural anomaly index according to their respective preset weights; the risk index is compared with multiple preset risk thresholds to determine the risk level of the bridge pier; and the warning information, which includes a warning level and recommended maintenance strategy, is generated based on the risk level, the causal analysis results, and the evolution trend.

[0012] Preferably, the subsequent maintenance records and actual evolution data of the bridge piers are obtained, and the subsequent maintenance records and actual evolution data are added to the training sample library; when the deviation between the prediction result of the data-driven model and the actual evolution data exceeds a predetermined error threshold, incremental training is performed on the causal inference model and the data-driven model based on the updated training sample library, and the parameters of the causal inference model and the data-driven model are updated; based on the statistical distribution of the multi-source monitoring data in the most recent preset time period, the dynamic fusion weights and the preset risk threshold are adaptively adjusted.

[0013] Furthermore, this application also provides a system for identifying latent risks, tracing causes, and predicting the evolution of cracks in bridge piers, comprising: a data fusion module for acquiring multi-source monitoring data of bridge piers, determining dynamic fusion weights based on historical quality indicators of each data source, and fusing the multi-source monitoring data based on the dynamic fusion weights to obtain a multi-source feature set; a cause tracing module for determining the contribution probability of each latent risk factor to the cracks in the bridge pier based on the multi-source feature set and a causal reasoning model, and obtaining the cause analysis results of the cracks; an evolution prediction module for incorporating the physical degradation law of the bridge pier as a constraint into a data-driven model, and predicting the evolution trend of the cracks based on the multi-source feature set and the constrained data-driven model; and an early warning generation module for generating risk assessment results and early warning information for the bridge pier based on the evolution trend and the cause analysis results.

[0014] In addition, this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method for identifying latent risks, tracing causes, and predicting the evolution of bridge pier cracks as described above.

[0015] In addition, this application also provides a computer-readable storage medium storing computer instructions for causing a computer to execute the method for identifying latent risks, tracing causes, and predicting the evolution of bridge pier cracks as described above.

[0016] Beneficial effects:

[0017] The technical solution provided in this application transforms the physical degradation laws of bridge piers into physical residual penalty terms or physical boundary truncation conditions, deeply integrating them into the training or inference process of data-driven models, thus constructing a dual-driven evolution prediction mechanism based on both physics and data. This mechanism utilizes physical laws as prior knowledge to constrain the solution space of the model, effectively suppressing overfitting and physical violations caused by data noise or sample scarcity in long-term predictions of purely data-driven models. This significantly improves the physical rationality and long-term reliability of crack evolution trend prediction. Traditional crack detection methods mostly rely on sensors to complete detection only after cracks have developed to the visible stage, resulting in significant detection lag. This application can predict risks in advance when cracks are still in the latent stage and simultaneously conduct causal diagnosis and analysis, facilitating early intervention by maintenance personnel and preventing latent risks from continuously deteriorating into visible damage.

[0018] Meanwhile, this application dynamically calculates the fusion weights based on historical quality indicators from various data sources and constructs a causal relationship map between multi-source parameters and crack states using a causal inference model, achieving accurate causal tracing in multi-factor coupled scenarios. Compared to traditional fixed-weight fusion or simple correlation analysis, this scheme can adaptively shield against interference from low-confidence data and quantify the contribution probability of each latent risk factor to crack development, effectively overcoming the technical problem of misjudging causes under complex working conditions, and achieving early identification of latent risks of bridge pier cracks, accurate diagnosis of crack causes, and physical consistency prediction of crack evolution trends. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for identifying, tracing the causes of, and predicting the evolution of bridge pier cracks according to an embodiment of this application. Figure 2 This is a flowchart of the detection, early warning, and feedback closed loop of bridge pier cracks according to an embodiment of this application; Figure 3 This is a system block diagram of a system for identifying hidden risks, tracing causes, and predicting the evolution of cracks in bridge piers, according to an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0022] Example 1:

[0023] like Figure 1 and Figure 2 As shown, this embodiment provides a method for identifying hidden risks, tracing causes, and predicting the evolution of bridge pier cracks. This method achieves accurate perception and early warning of the state of bridge pier cracks through the coordinated execution of multi-source data dynamic fusion, causal relationship tracing, and physical constraint prediction. It should be understood that, although... Figure 1 The flowchart shows a linear process, but in actual engineering deployment, each step can be processed in parallel or scheduled in a pipeline manner according to computing resources and real-time requirements, as long as the logical dependencies of data flow are met.

[0024] In step S1, multi-source monitoring data of the bridge piers are acquired. Dynamic fusion weights are determined based on historical quality indicators of each data source, and the multi-source monitoring data is fused based on these dynamic fusion weights to obtain a multi-source feature set. Specifically, the multi-source monitoring data includes time-series signals reflecting the mechanical behavior of the bridge piers (such as strain and vibration), image data characterizing apparent damage, and environmental parameters describing external environmental loads. Unlike existing technologies that use fixed weights or simple averaging fusion, this embodiment introduces historical quality indicators as a dynamic benchmark for weight calculation. Historical quality indicators refer to performance stability measures of each sensor or data acquisition channel over a preset time window, such as signal-to-noise variance, data missing rate, drift amplitude, or signal-to-noise ratio. The current reliability of each data source is evaluated in real time based on these indicators, thereby generating dynamic fusion weights. For example, when a strain sensor experiences a significant increase in recent noise variance due to aging, its corresponding fusion weight automatically decreases, while the weight of a stable environmental sensor increases accordingly. This mechanism ensures that the fused multi-source feature set can adaptively suppress interference from low-quality data and maintain a high-confidence state representation.

[0025] In step S2, based on the multi-source feature set and causal inference model, the contribution probability of each latent risk factor to the cracks in the bridge pier is determined, yielding the causal analysis results. Specifically, the causal inference model is a structured model capable of analyzing the causal dependencies between variables. Its input is a multi-source feature set, and its output is not a simple crack category label, but rather a quantified contribution probability of each latent risk factor (such as vehicle load fatigue, temperature gradient effect, uneven foundation settlement, concrete shrinkage and creep, etc.) to the current crack state. For example, the model might output a probability distribution with "load fatigue" contributing 70%, "temperature effect" contributing 20%, and "others" contributing 10%. This probability-based causal expression effectively solves the common problem in bridge engineering where the coupling of multiple factors is difficult to decouple, avoiding the shortcomings of traditional correlation analysis which easily misjudges accompanying phenomena as the root cause.

[0026] In step S3, the physical degradation law of the bridge pier is incorporated as a constraint into the data-driven model. Based on the multi-source feature set and the constrained data-driven model, the evolution trend of cracks is predicted. The physical degradation law refers to the deterministic mechanism equation describing the deterioration of bridge pier material properties or the accumulation of structural damage, such as the concrete fatigue damage evolution equation, fracture mechanics propagation criteria, or creep constitutive model. The data-driven model is a black-box or gray-box model based on mining time-series mapping relationships from historical monitoring data. Specific implementation paths for incorporating the physical degradation law of the bridge pier as a constraint into the data-driven model include, but are not limited to, the following two methods or combinations thereof: First, during the model training phase, the physical degradation law is transformed into a penalty or regularization term in the loss function, forcing the model to satisfy physical conservation or evolution laws while learning data mapping relationships; second, during the model inference phase, physical boundary conditions (such as non-negative crack width and propagation rate not exceeding material limits) are used to truncate or correct the initial prediction output of the model. This deep fusion mechanism effectively overcomes problems such as overfitting, divergence, or violation of physical common sense (such as predicting negative crack width) that are prone to occur in pure data models in small sample or long-term predictions, and significantly improves the physical rationality of evolution trend prediction results.

[0027] In step S4, risk assessment results and early warning information for the bridge piers are generated based on the evolution trend and causal analysis results. Specifically, the risk assessment results are not determined solely by a single static indicator such as the current crack width, but are generated by combining the future evolution trend output in step S3 (such as the expected growth rate and the time to reach the critical value) with the causal type output in step S2. For example, even if the current crack width is small, a high-level risk warning will still be generated if the prediction shows that it will expand rapidly under load fatigue; conversely, if the crack is mainly caused by seasonal temperature changes and the prediction tends to be stable, the risk level will be lowered accordingly. The early warning information not only includes a risk level identifier but can also be associated with recommended maintenance strategies, such as recommending load limits or reinforcement for load fatigue-type cracks and anti-corrosion coating repair for environmental erosion-type cracks.

[0028] Through the coordinated execution of steps S1 to S4, this embodiment ensures the quality of input data through dynamic weight fusion and solves the problems of misjudgment of causes and prediction distortion under complex working conditions by synergistic enhancement of causal reasoning and physical constraints, thus ensuring the refined maintenance of bridge piers. This application can predict risks in advance when cracks are still in the latent stage and carry out cause diagnosis analysis simultaneously, which facilitates early intervention by operation and maintenance personnel and avoids the continuous deterioration of latent risks into manifest defects. This enables early identification of latent risks of bridge pier cracks, accurate diagnosis of crack causes, and physical consistency prediction of crack evolution trends. It overcomes the problem that traditional crack detection methods mostly rely on sensors to complete detection only after the crack has developed to the manifest stage, resulting in obvious detection lag.

[0029] Example 2:

[0030] Building upon Example 1, this example obtains historical quality indicators for each data source over a historical time period, calculates the confidence level of each data source based on these historical quality indicators, normalizes the confidence levels of each data source, and determines the normalized confidence levels as dynamic fusion weights. Specifically, historical quality indicators are statistical measures that quantify the reliability of sensor data, and their specific forms may include at least one of signal-to-noise variance, data missing rate, drift amplitude, or signal-to-noise ratio. For example, for a strain sensor, the noise variance can be calculated using the residual sequence from the past 24 hours. For image acquisition devices, the percentage of valid frames can be used as a completeness indicator. Confidence score is used to map heterogeneous quality indicators to a unified confidence score; for example, an inverse variance weighted method, i.e., confidence score, can be used. ,in To prevent smoothing factors with zero denominators, the confidence scores of all data sources are normalized so that the sum of the dynamic fusion weights of each data source is 1, thus ensuring the stability of the fused multi-source feature set on a numerical scale. This dynamic weighting mechanism based on real-time quality assessment, compared to fixed-weight schemes, can adaptively suppress the contamination of the overall analysis results by low-quality data caused by sensor aging, environmental interference, or communication failures, ensuring that the feature vectors input to subsequent models always have high fidelity.

[0031] Based on dynamic fusion weights, time-series features, image features, and environmental features from multi-source monitoring data are weighted and fused to obtain a multi-source feature set. Specifically, time-series features may include frequency domain energy of strain time-history curves, rate of change of natural vibration frequencies, etc.; image features may include morphological parameters such as crack width, length, tortuosity, and distribution density; environmental features may include external load parameters such as temperature gradient, cumulative humidity, and extreme wind speed. The weighted fusion process is not a simple numerical concatenation, but rather a linear or nonlinear combination of each modal feature vector according to dynamic fusion weights to form a high-dimensional feature representation that comprehensively characterizes the current state of the bridge pier. It should be understood that although this embodiment lists three types of features—time-series, image, and environmental—in other embodiments, other types of features such as acoustic emission and electrochemical corrosion can be introduced according to the actual monitoring configuration, as long as the fusion requirements of multi-source heterogeneous data are met.

[0032] A causal graph containing multi-source parameter nodes and crack nodes is constructed. The conditional probabilities of the causal graph are updated based on the multi-source feature set, and the contribution probability of each latent risk factor to the crack is calculated based on the updated causal graph to obtain the causal analysis results. Specifically, the causal graph is a directed acyclic graph structure, where nodes represent variables and directed edges represent direct causal dependencies between variables. In this embodiment, the nodes of the causal graph are explicitly divided into two categories: one category is multi-source parameter nodes, which serve as causal variables, specifically including load nodes (such as vehicle axle load, traffic flow), environmental nodes (such as temperature difference, freeze-thaw cycle count), and structural response nodes (such as settlement, tilt angle, stress level); the other category is crack nodes, which serve as outcome variables, including different types of crack sub-nodes, such as longitudinal cracks, transverse cracks, diagonal cracks, or network cracks. Edges express the physical action path from causal nodes to outcome nodes, for example, "load node" points to "fatigue crack sub-node," and "environmental node" points to "shrinkage crack sub-node," thereby explicitly embedding the a priori mechanical mechanisms of bridge engineering into the data analysis model.

[0033] As a preferred implementation, a causal graph is constructed using Bayesian networks or decision tree algorithms. The multi-source parameter nodes include load nodes, environmental nodes, and structural response nodes, while the crack nodes include different types of crack child nodes. Bayesian networks are preferred due to their ability to handle uncertainty and fuse multi-source information; however, it should be understood that decision trees, structural equation models, or other graphical models supporting causal inference are also applicable to this application, as long as they can achieve a mapping from observed data to causal probabilities. Based on the multi-source feature set, a conditional probability table between the multi-source parameter nodes and crack nodes is calculated using the maximum likelihood estimation method. The conditional probability table quantifies the probability distribution of child nodes (results) occurring given a parent node (cause). Maximum likelihood estimation approximates the true conditional probability using the joint observation frequency in a historical monitoring sample database, for example, by statistically analyzing the historical occurrence frequency of "lateral crack propagation" under conditions of "heavy traffic" and "high temperature difference." With the continuous ingestion of new data, the conditional probability table can be incrementally updated through an online learning mechanism, enabling the model to adapt to the time-varying characteristics of bridge service performance.

[0034] The multi-source feature set at the current moment is input into the causal graph. Bayesian inference is used to calculate the posterior probability of each crack sub-node. The posterior probability is used as the contribution probability, and the causal analysis results are output. The causal analysis results include the probability distribution corresponding to each cause type, as well as the events that triggered the crack. Specifically, when the measured features at the current moment (such as "abnormal increase in strain" and "sudden drop in temperature") are obtained as evidence variables and input into the causal graph, the Bayesian inference engine will inversely deduce the posterior probability of each potential cause node being in an active state based on the conditional probability table and graph structure. For example, the inference results may show that "load fatigue" causes the current crack state with a probability of 85%, "temperature stress" with a probability of 10%, and "foundation settlement" with a probability of 5%. At the same time, the most likely combination of triggering events can also be identified, such as "recent frequent passage of overloaded vehicles combined with cooling weather".

[0035] This embodiment employs a causal graph instead of traditional Pearson correlation coefficients or simple regression analysis to address the common problems of spurious correlations and multi-factor coupling in bridge health monitoring. In actual engineering, many variables exhibit highly synchronous changes over time (e.g., temperature increases and crack width increases occur simultaneously), but this does not necessarily indicate a direct causal relationship. They may be driven by common hidden variables (e.g., seasonal variations) or simply statistical coincidence. Simple correlation analysis cannot distinguish this accompanying relationship from the true mechanical causal chain, easily leading to misdiagnosis. The causal graph, by explicitly defining the directed dependency structure between variables and combining it with conditional independence tests, can extract the true causal effect from statistical associations. Furthermore, pier cracks are often the result of the interplay of multiple factors. The causal graph supports joint reasoning with multiple parent nodes, enabling the quantification and analysis of the independent contributions and interaction effects of each factor under specific working conditions. This provides diagnostic results with greater interpretability and engineering guidance value than single threshold judgments or black-box classification models, significantly enhancing the credibility of the proposed technical solution in complex and unstructured environments.

[0036] Example 3:

[0037] Building upon Example 1, this example transforms physical degradation patterns into physical residual penalty terms, incorporates these penalty terms into the loss function of the data-driven model, and updates the model's parameters by minimizing the loss function containing the physical residual penalty terms, resulting in a soft-constrained data-driven model. Alternatively, it transforms physical degradation patterns into physical boundary truncation conditions, and after the data-driven model outputs its initial prediction results, it truncates and corrects these results based on the physical boundary truncation conditions, resulting in a hard-constrained data-driven model. Based on a multi-source feature set, the evolution trend of cracks is predicted using both the soft-constrained and hard-constrained data-driven models. This example, through a collaborative path of soft and hard constraints, deeply embeds prior mechanical knowledge of bridge engineering into the data-driven prediction process, solving the technical problem that pure data models easily violate physical common sense in long-term predictions. It achieves physically consistent prediction of the evolution trend of bridge pier cracks, while also improving early warning effectiveness and attribution.

[0038] Specifically, for soft-constraint paths, the material fatigue damage equation and / or structural creep model of the bridge piers are obtained as physical degradation laws. Physical degradation laws are deterministic mathematical expressions describing the deterioration process of bridge pier material properties over time, under load, and environmental influences. For example, for prestressed concrete bridge piers subjected to repeated vehicle loads, fatigue damage evolution equations based on continuum damage mechanics can be used; for bridge piers subjected to long-term constant loads, formulas considering temperature and humidity effects on the concrete creep coefficient can be used. These equations provide theoretical benchmarks for the evolution of crack state variables (such as crack width and damage factor) over time, are independent of the statistical distribution of monitoring data, and possess universal physical truth.

[0039] A time-series neural network is constructed as the data-driven model. This network comprises a Long Short-Term Memory (LSTM) layer and a fully connected output layer. The LSM layer captures long-range temporal dependencies in multi-source monitoring data, while the fully connected output layer maps high-dimensional hidden states to specific predicted values ​​of crack evolution indicators. It should be understood that although this embodiment uses an LSM network as an example, other implementations may employ gated recurrent units, temporal convolutional networks, or Transformer architectures—any neural network capable of time-series modeling—as long as it can process sequential multi-source feature sets and output evolution trends.

[0040] The physical evolution residuals of cracks between adjacent time steps are calculated based on the laws of physical degradation, and the sum of squares of these residuals is used as a penalty term. Specifically, the physical evolution residual refers to the deviation between the predicted state change by the data-driven model and the theoretical change specified by the laws of physical degradation. For example, if the laws of physical degradation describe the theoretical increment of crack width at a specific stress level, and the increment predicted by the model deviates from this theoretical value, a non-zero physical evolution residual is generated. Summing the squares of the physical evolution residuals across all time steps within a time window constitutes a scalar index quantifying the degree to which the model violates physical laws.

[0041] A joint loss function is constructed, comprising a data prediction error term and a physical residual penalty term. The gradient of the joint loss function is calculated based on historical monitoring datasets, and the weight parameters of the time-series neural network are updated using the backpropagation algorithm, resulting in a soft-constrained data-driven model. The mathematical expression of the joint loss function is typically as follows: ,in This represents the mean square error or absolute error between the model's predicted values ​​and the actual monitored values, and is used to ensure the model's accuracy in fitting the observed data. This represents the physical residual penalty term, used to constrain the solution space of the model to conform to the laws of mechanical evolution; These are the physical constraint weights, used to balance the relationship between data fitting and physical consistency. During training, the backpropagation algorithm adjusts parameters not only based on data errors to approximate the observed values, but also based on physical residuals to eliminate predictive tendencies that violate physical laws. When monitoring data contains noise or is missing, It can prevent the model from overfitting to outlier data points and guide the model to learn evolution patterns that conform to the physical nature of the data.

[0042] For hard-constraint paths, the physical degradation law is transformed into physical boundary truncation conditions. After the data-driven model outputs the initial prediction results, these results are truncated and corrected based on the physical boundary truncation conditions. These physical boundary truncation conditions are insurmountable state boundaries determined by engineering common sense and material limits. For example, the crack width must satisfy a non-negativity constraint (…). The crack propagation rate cannot exceed the theoretical upper limit determined by the material's fracture energy release rate. (This means that the initial predictions from the data-driven model must not exceed these physical boundaries, or the cumulative damage factor cannot exceed 1.0.) When the initial predictions from the data-driven model exceed these physical boundaries, they are automatically projected onto the nearest legal boundary. For example, if the model predicts a crack width of -0.02 mm at the next moment, it is forcibly corrected to 0 mm; if the predicted propagation rate is 5 mm / h but the material limit is 2 mm / h, it is corrected to 2 mm / h. This hard constraint mechanism ensures that no matter how the model performs its internal calculations, the final output warning information is always physically reasonable and interpretable, preventing predictions that violate physical common sense, such as "negative cracks" or "superluminal propagation."

[0043] To verify the effectiveness of the technical solution in this embodiment, the prediction performance of the pure data-driven model and the dual-driven model of physics and data proposed in this application were compared in actual engineering verification under scenarios of sparse data and strong noise interference. Test results show that when sensor failure resulted in three consecutive days of missing data, the prediction curve of the pure data-driven model exhibited violent fluctuations and produced multiple non-physical predictions with negative crack widths, resulting in an average absolute percentage error exceeding 35%. In contrast, the model employing the soft and hard constraint collaborative mechanism of this embodiment maintained a smooth and monotonically increasing prediction curve, without any anomalies that violated physical principles, and the average absolute percentage error was controlled within 8%. This significant difference fully demonstrates that this application, through a deep fusion mechanism, truly makes physical laws the intrinsic constraint of the data-driven model, thereby achieving a level of reliability far exceeding traditional methods in complex and ever-changing actual monitoring environments.

[0044] Example 4:

[0045] Building upon Examples 1 to 3, this example acquires a current severity index for cracks, a predicted growth rate index determined based on evolution trends, and a structural anomaly index determined based on a multi-source feature set. Specifically, the current severity index directly quantifies the current state of cracks in bridge piers, typically defined as the ratio of the measured crack width to the width allowed by design specifications, or the ratio of crack depth to the thickness of the protective layer, used to characterize the immediate urgency of the damage. The predicted growth rate index, output by a physically constrained data-driven model, reflects the rate of crack expansion within a preset time window or the reciprocal of the time to reach a critical state, embodying the dynamic development trend of the risk. Based on structural response parameters other than the cracks themselves from the multi-source feature set, structural anomaly indices are extracted, such as the frequency of strain gauge stress exceeding limits, the abnormal rotation amplitude of inclinometers, or significant shifts in vibration frequency, used to capture potential structural hazards that may induce crack aggravation. This approach overcomes the limitations of traditional methods that rely solely on a single crack width threshold for alarms, enabling comprehensive identification of complex conditions such as "minor current state but rapid development" or "serious current state but stabilizing," significantly improving the comprehensiveness and accuracy of risk assessment.

[0046] The risk index is obtained by weighting and summing the current severity index, the predicted growth rate index, and the structural anomaly index according to their respective preset weights. The preset weights can be pre-set based on bridge type, service life, or expert experience, or dynamically determined through statistical analysis using a historical accident case database. For example, for prestressed concrete bridge piers in their early operational phase, the predicted growth rate index may have a higher weight to focus on early cracking trends; while for older bridges, the current severity index may have a more dominant weight. The formula for calculating the risk index can be expressed as follows: Where R is the risk index. These are the normalized current severity index, the predicted growth rate index, and the structural anomaly index, respectively. The corresponding preset weight and satisfying It should be understood that although this embodiment uses linear weighted summation as an example, in other implementations, nonlinear mapping, fuzzy comprehensive evaluation, or regression models based on machine learning can also be used to calculate the risk index, as long as the mapping from multidimensional indicators to a unified risk scalar can be achieved.

[0047] The risk index is compared with multiple preset risk thresholds to determine the risk level of the bridge piers. These preset risk thresholds are typically divided into multiple levels, such as normal, attention, warning, and danger, each corresponding to a different risk index range. Based on the risk level, causal analysis results, and evolution trends, early warning information is generated, including warning levels and recommended maintenance strategies. Specifically, the early warning information is not merely a simple audible and visual alarm signal, but rather contains structured decision support content with rich semantics. For example, when the system determines the risk level to be "warning" and the causal analysis results show that the main contributing factor is "load fatigue," the generated early warning information will not only display a "yellow warning," but will also automatically generate targeted recommended maintenance strategies such as "restricting heavy-load vehicle traffic" or "inspecting the supports." If the cause is "environmental erosion" and the predicted growth rate is moderate, "surface sealing treatment" or "anti-corrosion coating repair" may be recommended.

[0048] To enable the system to self-evolve throughout its entire lifecycle, this embodiment also introduces a closed-loop adaptive feedback mechanism. Subsequent maintenance records and actual evolution data of the bridge piers are acquired and added to the training sample library. Subsequent maintenance records include manual inspection reports, repair and reinforcement construction records, and material replacement information. Actual evolution data refers to the real crack state data continuously collected by sensors within a certain period after the warning is issued. When the deviation between the prediction result of the data-driven model and the actual evolution data exceeds a predetermined error threshold, incremental training is performed on the causal inference model and the data-driven model based on the updated training sample library, updating the parameters of both models. The predetermined error threshold is the criterion for triggering model updates; for example, it can be set as the root mean square error between the predicted crack width and the actual crack width exceeding 0.05 mm for three consecutive days. Incremental training refers to fine-tuning the model parameters using new samples while retaining the original model knowledge, rather than retraining from scratch. This ensures the continuity of learning and significantly reduces computational overhead. In this way, the model can continuously absorb new disease evolution patterns and maintenance intervention effects, gradually improving the prediction accuracy of the bridge's individual behavior.

[0049] Based on the statistical distribution of multi-source monitoring data within a recent preset time period, the system adaptively adjusts the dynamic fusion weights and preset risk thresholds. This mechanism is independent of incremental training of model parameters and focuses on online calibration of system hyperparameters. Specifically, the system uses a sliding window to statistically analyze the noise level, missing rate, and outlier distribution of each sensor's data in real time. For example, if the variance of a strain sensor's data increases significantly within the past week, indicating a decrease in its signal-to-noise ratio or performance degradation, the system will automatically reduce the dynamic fusion weight of that sensor in step S1 to prevent inferior data from contaminating the fusion results. Similarly, if a recent overall increase in ambient temperature causes widespread thermal expansion and contraction shifts in crack width, the system can automatically raise the warning threshold based on the quantile distribution of the current data to avoid frequent false alarms caused by seasonal changes. This adaptive adjustment strategy based on data statistical characteristics allows the system to adapt to long-term disturbances such as sensor aging, environmental changes, and gradual changes in structural performance without frequent manual intervention, ensuring the reliability of the monitoring and early warning system over its decades-long service life.

[0050] Example 5:

[0051] like Figure 3As shown, this embodiment provides a system for identifying hidden risks, tracing causes, and predicting the evolution of cracks in bridge piers. This system implements the functions described in the aforementioned method embodiments through a modular logical architecture, ensuring the standardization and scalability of the data processing flow. It should be understood that the system described in this embodiment is a functional virtual device, and its modules can be implemented by one or more combinations of software programs, firmware, or application-specific integrated circuits, and are not limited to a specific physical hardware carrier. The system includes: The data fusion module acquires multi-source monitoring data of the bridge piers, determines dynamic fusion weights based on historical quality indicators of each data source, and fuses the multi-source monitoring data based on these weights to obtain a multi-source feature set. Specifically, the data fusion module serves as the system's data entry point, integrating a data cleaning unit, a quality assessment unit, and a weighted fusion unit. The data cleaning unit removes outliers and aligns the raw acquired signals spatiotemporally; the quality assessment unit calculates historical quality indicators such as noise variance and missing rate for each channel in real time according to a preset time window and maps them to normalized confidence weights; the weighted fusion unit adaptively combines time-series, image, and environmental features based on these weights. This modular design decouples the data preprocessing logic from the subsequent analysis logic. When adding a new sensor type or replacing the data acquisition equipment, only the adaptation interface and quality assessment parameters within the data fusion module need to be updated, without reconstructing the entire system architecture, significantly improving the system's engineering adaptability and maintenance efficiency.

[0052] The causal analysis module, based on a multi-source feature set and a causal inference model, determines the contribution probability of each latent risk factor to the cracks in the bridge piers, thus obtaining the causal analysis results. Specifically, the module incorporates a causal graph construction engine and a probabilistic inference engine. The causal graph construction engine maintains a directed acyclic graph structure containing nodes for load, environment, structural response, and crack type, along with its conditional probability table. The probabilistic inference engine receives the multi-source feature set from the data fusion module as evidence input and calculates the posterior probability distribution of each causal node through Bayesian inference or other causal inference algorithms. The module's output is not merely a single label classification but a structured probability vector, clearly quantifying the contribution of each factor. By encapsulating the causal inference logic as an independent module, the system supports offline updates of the causal graph structure or retraining of the conditional probability table without interrupting online monitoring services, achieving continuous iteration and hot loading of the knowledge model.

[0053] The evolution prediction module incorporates the physical degradation patterns of bridge piers as constraints into the data-driven model. Based on a multi-source feature set and the constrained data-driven model, it predicts the evolution trend of cracks. This is the most unique core component in the system architecture. Specifically, the evolution prediction module does not simply contain a general time series predictor, but rather deeply integrates a physical constraint sub-unit. This physical constraint sub-unit has a tight logical coupling with the data-driven predictor, rather than a simple sequential connection. In one implementation, the physical constraint sub-unit is embedded in the model training loop as a loss function generator, used to transform the material fatigue damage equation or creep model into a physical residual penalty term, participating in the gradient backpropagation process, thereby achieving soft constraints on the data-driven model. In another implementation, the physical constraint sub-unit is deployed at the model inference output as a post-processing validator, correcting the validity of the initial predicted values ​​based on physical boundary truncation conditions, achieving hard constraints. In actual system deployment, the physical constraint sub-unit can exist as an independent microservice, dynamic link library, or embedded operator, as long as it can apply physical constraints during the training or inference process of the data-driven model. This architectural design ensures that the evolution trend of the system output always conforms to the basic principles of mechanics, fundamentally avoiding the non-physical understanding that may arise from a pure black box model.

[0054] The early warning generation module generates risk assessment results and early warning information for bridge piers based on evolutionary trends and causal analysis. Specifically, it receives probability distribution vectors from the causal tracing module and future state curves from the evolution prediction module, and combines these with preset risk index calculation rules and multi-level threshold strategies to comprehensively determine the current risk level. More importantly, this module incorporates a decision knowledge graph, enabling it to map abstract risk levels to specific causal types into actionable recommended maintenance strategies. For example, when identifying load fatigue-type cracks with high growth rates, it automatically generates load-limiting suggestions and reinforcement schemes; when identifying environmental erosion-type cracks, it recommends anti-corrosion repair measures. The output interface of the early warning generation module supports multiple protocol adaptations, allowing structured early warning information to be pushed to monitoring screens, mobile terminals, or maintenance management systems, achieving a closed loop from data analysis to business action.

[0055] Example 6:

[0056] This embodiment provides an electronic device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform the methods for identifying latent risks, tracing causes, and predicting the evolution of bridge pier cracks described in any of the foregoing embodiments. Specifically, this electronic device constitutes the physical hardware foundation for implementing the technical solution of this application. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor, etc., or a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. The memory can include volatile memory, such as random access memory (RAM); or non-volatile memory, such as flash memory, hard disk, or solid-state drive (SSD). In actual engineering deployments, the specific form of this electronic device can be flexibly selected according to the application scenario. For example, in an edge computing scenario at a bridge site, the electronic device can be an industrial-grade control computer or an embedded AI box embedded in a monitoring cabinet, directly connecting to sensor acquisition cards for real-time data fusion and local early warning; in a centralized cloud analysis scenario, the electronic device can be a high-performance server cluster in a data center, used to process massive amounts of historical data and perform complex incremental model training.

[0057] This embodiment also provides a computer-readable storage medium storing computer instructions for causing a computer to execute the methods for identifying latent risks, tracing causes, and predicting the evolution of bridge pier cracks described in any of the foregoing embodiments. Specifically, a computer-readable storage medium refers to any tangible medium capable of containing, storing, communicating, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device. Specific forms include, but are not limited to, read-only memory, random access memory, CD-ROM, DVD, magnetic tape, disk storage devices, optical storage devices, and semiconductor memories. It should be understood that the storage medium protected in this application is non-transient and does not include carrier signals or transient electrical signals themselves. By embedding the computer instructions for implementing the aforementioned methods in this storage medium, the technical solution of this application can be distributed, sold, or deployed as a software product. When maintenance personnel insert this storage medium into the computing device of the bridge health monitoring system and load the instructions therein, the originally general-purpose computing device can be transformed into a dedicated device with intelligent diagnostic capabilities for bridge pier cracks. This implementation method not only facilitates the promotion and upgrading of the technology, but also provides a convenient carrier for subsequent optimization of physical constraint models or adjustment of causal inference parameters through software updates, ensuring the maintainability and scalability of the technical solution throughout its entire life cycle.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. For example, the causal inference model, time series neural network architecture, and specific mathematical expressions of physical degradation laws mentioned in the foregoing embodiments are merely illustrative examples. In practical applications, equivalent substitutions or combinations can be made according to the bridge type, monitoring conditions, and computing resources. As long as it can achieve the functions of dynamic fusion of multi-source data, causal probability tracing, and physical constraint prediction, it falls within the substantive protection scope of this application.

Claims

1. A method for identifying hidden risks, tracing causes, and predicting the evolution of cracks in bridge piers, characterized in that, include: Multi-source monitoring data of bridge piers are acquired, dynamic fusion weights are determined based on historical quality indicators of each data source, and the multi-source monitoring data are fused based on the dynamic fusion weights to obtain a multi-source feature set. Based on the multi-source feature set and causal reasoning model, the contribution probability of each hidden risk factor to the cracks in the bridge pier is determined, and the causal analysis results of the cracks are obtained. The physical degradation law of the bridge pier is incorporated as a constraint into the data-driven model. Based on the multi-source feature set and the constrained data-driven model, the evolution trend of the crack is predicted. Based on the evolution trend and the causal analysis results, risk assessment results and early warning information for the bridge piers are generated.

2. The method according to claim 1, characterized in that, The dynamic fusion weights are determined based on historical quality indicators from each data source, and the multi-source monitoring data are fused based on these dynamic fusion weights to obtain a multi-source feature set, including: Obtain historical quality indicators for each data source within a historical time period, calculate the confidence level of each data source based on the historical quality indicators, normalize the confidence level of each data source, and determine the normalized confidence level as the dynamic fusion weight. Based on the dynamic fusion weights, the temporal features, image features, and environmental features in the multi-source monitoring data are weighted and fused to obtain the multi-source feature set. Based on the multi-source feature set and causal inference model, the contribution probability of each latent risk factor to the cracks in the bridge pier is determined, and the causal analysis results of the cracks are obtained, including: A causal graph containing multi-source parameter nodes and crack nodes is constructed. The conditional probability of the causal graph is updated based on the multi-source feature set. The contribution probability of each latent risk factor to the crack is calculated based on the updated causal graph, and the causal analysis result of the crack is obtained.

3. The method according to claim 2, characterized in that, The process of constructing a causal graph containing multi-source parameter nodes and crack nodes, updating the conditional probability of the causal graph based on the multi-source feature set, and calculating the contribution probability of each latent risk factor to the crack based on the updated causal graph to obtain the causal analysis results of the crack includes: The causal graph is constructed using a Bayesian network or decision tree algorithm, wherein the multi-source parameter nodes include load nodes, environment nodes, and structural response nodes, and the crack nodes include crack sub-nodes of different types. Based on the multi-source feature set, the conditional probability table between the multi-source parameter nodes and the crack nodes is calculated using the maximum likelihood estimation method; The multi-source feature set at the current moment is input into the causal graph. The posterior probability of each crack sub-node is calculated through Bayesian inference. The posterior probability is used as the contribution probability, and the causal analysis result of the crack is output. The causal analysis result includes the probability distribution corresponding to each causal type and the event that triggered the crack.

4. The method according to claim 1, characterized in that, The step of incorporating the physical degradation law of the bridge pier as a constraint into the data-driven model, and predicting the evolution trend of the crack based on the multi-source feature set and the constrained data-driven model, includes: The physical degradation law is transformed into a physical residual penalty term, which is then added to the loss function of the data-driven model. The parameters of the data-driven model are updated by minimizing the loss function containing the physical residual penalty term, resulting in a soft-constrained data-driven model; and / or, The physical degradation law is transformed into a physical boundary truncation condition. After the data-driven model outputs the initial prediction result, the initial prediction result is truncated and corrected based on the physical boundary truncation condition to obtain the data-driven model with hard constraints. Based on the multi-source feature set, the evolution trend of the crack is predicted by the data-driven model under soft constraints and / or the data-driven model under hard constraints.

5. The method according to claim 4, characterized in that, The step of transforming the physical degradation law into a physical residual penalty term, adding the physical residual penalty term to the loss function of the data-driven model, and updating the parameters of the data-driven model by minimizing the loss function containing the physical residual penalty term to obtain the soft-constrained data-driven model includes: The material fatigue damage equation and / or structural creep model of the bridge pier are obtained as the physical degradation law; A time-series neural network is constructed as the data-driven model, and the time-series neural network includes a long short-term memory network layer and a fully connected output layer; The physical evolution residual of the crack between adjacent time steps is calculated based on the physical degradation law, and the sum of squares of the physical evolution residuals is used as the physical residual penalty term. A joint loss function is constructed, which includes a data prediction error term and a physical residual penalty term. The gradient of the joint loss function is calculated based on the historical monitoring dataset. The weight parameters of the time series neural network are updated using the backpropagation algorithm to obtain the data-driven model with soft constraints.

6. The method according to claim 1, characterized in that, Based on the evolutionary trend and the causal analysis results, the risk assessment results and early warning information for the bridge piers are generated, including: Obtain the current severity index of the crack, the predicted growth rate index determined based on the evolution trend, and the structural anomaly index determined based on the multi-source feature set; The risk index is obtained by weighting and summing the current severity index, the predicted growth rate index, and the structural anomaly index according to their respective preset weights. The risk index is compared with multiple preset risk thresholds to determine the risk level of the bridge pier; Based on the risk level, the causal analysis results, and the evolution trend, the early warning information, which includes the early warning level and recommended maintenance strategies, is generated.

7. The method according to claim 6, characterized in that, The method further includes: Obtain the subsequent maintenance records and actual evolution data of the bridge piers, and add the subsequent maintenance records and actual evolution data to the training sample library; When the deviation between the prediction result of the data-driven model and the actual evolution data exceeds a predetermined error threshold, incremental training is performed on the causal inference model and the data-driven model based on the updated training sample library, and the parameters of the causal inference model and the data-driven model are updated. Based on the statistical distribution of the multi-source monitoring data within the most recent preset time period, the dynamic fusion weight and the preset risk threshold are adaptively adjusted.

8. A system for identifying hidden risks, tracing causes, and predicting the evolution of cracks in bridge piers, characterized in that, include: The data fusion module is used to acquire multi-source monitoring data of bridge piers, determine dynamic fusion weights based on historical quality indicators of each data source, and fuse the multi-source monitoring data based on the dynamic fusion weights to obtain a multi-source feature set. The cause tracing module is used to determine the contribution probability of each hidden risk factor to the cracks in the bridge pier based on the multi-source feature set and causal reasoning model, and to obtain the cause analysis results of the cracks. An evolution prediction module is used to incorporate the physical degradation law of the bridge pier as a constraint into a data-driven model, and predict the evolution trend of the crack based on the multi-source feature set and the constrained data-driven model. The early warning generation module is used to generate risk assessment results and early warning information for the bridge piers based on the evolution trend and the causal analysis results.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method for identifying latent risks, tracing causes, and predicting the evolution of bridge pier cracks as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the method for identifying the latent risks, tracing the causes, and predicting the evolution of bridge pier cracks as described in any one of claims 1 to 7.

Images