Bridge large sample safety evaluation method based on NLP and reinforcement learning

Abstract

The application discloses a bridge large sample safety evaluation method based on NLP and reinforcement learning, comprising the following steps: acquiring bridge multi-source heterogeneous historical data; constructing a deep learning model based on Bert-BiLSTM-CRF, adopting a BIO sequence labeling strategy to extract disease indexes from unstructured inspection text reports; filtering and processing dynamic weighing time series traffic load data, and extracting traffic load characteristics; aligning and fusing the disease indexes, traffic load characteristics and low-frequency regional meteorological geological data on a time axis to construct a multi-modal engineering feature tensor; inputting the multi-modal engineering feature tensor into a LightGBM model fused with a reinforcement learning agent, and outputting a technical condition evaluation grade of the bridge by maximizing an asymmetric reward function. The application applies exponential-level punishment to the missed reports through the asymmetric reward function, and effectively reduces the missed report rate of high-risk bridges.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of civil engineering structural health monitoring, intelligent bridge engineering detection, and artificial intelligence, specifically to a large-sample safety assessment method for bridges based on NLP and reinforcement learning. Background Technology

[0002] The total number of highway bridges in my country has exceeded one million. Among them, small and medium-span bridges built in the 1980s and 1990s are experiencing accelerated wear and tear on their structural fatigue life after long-term heavy traffic, material aging, and environmental erosion, gradually entering a period of rapid degradation and high accident incidence. Currently, bridge safety assessment mainly relies on on-site manual visual inspection by professional bridge inspection engineers, and strict scoring and rating according to the "Highway Bridge Technical Condition Assessment Standard". However, when facing large-scale bridge groups at the city or provincial level, existing manual inspection and traditional finite element analysis technology based on physical mechanics have revealed many technical bottlenecks.

[0003] Over the past decade, bridge maintenance platforms at all levels have accumulated a massive amount of unstructured routine inspection reports. The textual descriptions contained within these reports are of extremely high reference value for understanding the evolution of bridge conditions. However, due to the lack of a unified structured format, they cannot be directly used as feature variables in data-driven engineering calculations, resulting in a "rich in data, poor in information" dilemma. Furthermore, the raw sensor signals from the dynamic weighing systems typically equipped on modern bridges are often mixed with strong environmental white noise (such as structural micro-vibrations caused by wind loads and vehicle bounces caused by road surface unevenness). Traditional methods directly calculate the mean or extreme values ​​of the raw waveforms, masking the true high-frequency impact characteristics that lead to the accumulation of bridge fatigue damage, resulting in distorted underlying data. In addition, bridge degradation is a structural effect caused by both external factors (such as high-frequency temporal heavy load signals and low-frequency meteorological erosion) and internal factors (such as periodic textual damage records). Currently, the engineering community lacks an effective mathematical mechanism to perform cross-modal spatiotemporal resampling and deep fusion of discrete semantic text, high-frequency continuous sensor waveforms, and low-frequency environmental meteorological data, preventing assessment models from establishing a quantitative mapping relationship between load and damage. More critically, in real-world engineering sample databases, bridges in safe or good condition (Classes I, II, and III) often account for over 90%, while those in dangerous or near-dangerous condition (Classes IV and V) account for a very low percentage (typically between 3% and 6%), constituting a severe class imbalance problem. Traditional machine learning algorithms, in pursuit of the highest accuracy on the global dataset, tend to predict all marginal samples as safe, leading to fatal false negatives that misclassify dangerous bridges (Class V) as safe bridges (Class I). Existing oversampling or undersampling methods, by disrupting the true distribution patterns of engineering physics, cannot fundamentally solve this problem. Therefore, there is an urgent need for a large-sample bridge safety assessment method that can fully utilize multi-source heterogeneous data, effectively address class imbalance, and prioritize the detection rate of high-risk bridges. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a large-sample bridge safety assessment method based on NLP and reinforcement learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A large-sample security assessment method for bridges based on NLP and reinforcement learning includes the following steps: S1. Acquire multi-source heterogeneous historical data of bridges, including unstructured scheduled inspection reports, dynamic weighing time-series traffic load data, and low-frequency regional meteorological and geological data; S2. Construct a deep learning model based on Bert-BiLSTM-CRF, and use the BIO sequence labeling strategy to perform named entity recognition on unstructured inspection text reports, extracting disease indicators that include bridge disease types, disease locations and quantitative values. S3. Filter the dynamic weighing time series traffic load data to extract traffic load characteristics; S4. Align and fuse the extracted disease indicators with traffic load characteristics and low-frequency regional meteorological and geological data on the time axis to construct a multimodal engineering feature tensor; S5. Input the multimodal engineering feature tensor into the LightGBM model that incorporates a reinforcement learning agent, and output the technical condition rating of the bridge by maximizing the asymmetric reward function.

[0006] Furthermore, in S1, when obtaining unstructured inspection text reports, optical character recognition technology is used in advance to extract the text from the scanned image, and regular expressions are used to filter out interfering characters. At the same time, information involving personnel privacy and sensitive locations is automatically desensitized and cleaned.

[0007] Furthermore, S2 specifically includes: The BERT model in deep learning is used to perform deep semantic encoding on the fixed-time text report to generate high-dimensional word vectors. The bidirectional long short-term memory network in the deep learning model is used to extract forward and backward sequence features from high-dimensional word vectors; By using the conditional random field in the deep learning model to learn the state transition probability matrix, global Viterbi constraint decoding is performed on the sequence features to output the BIO labeled sequence. Based on the BIO annotation sequence, defect indicators containing the type, location, and quantitative value of bridge defects are extracted from the scheduled inspection text report.

[0008] Furthermore, in S3, when filtering the dynamic weighing time-series traffic load data, a Butterworth bandpass filter is used for zero-phase filtering to remove environmental white noise and extract the vehicle impact equivalent axle load.

[0009] Furthermore, S4 specifically includes: Using the time axis of the dynamic weighing time series traffic load data corresponding to the traffic load characteristics as a benchmark, a forward filling strategy is implemented on the disease indicators and low-frequency regional meteorological and geological data to achieve alignment and fusion on the time axis. The aligned and fused disease indicators, traffic load characteristics, and low-frequency regional meteorological and geological data were uniformly converted into float64 floating-point type and normalized using Z-score standardization to obtain multimodal engineering feature tensors.

[0010] Furthermore, in S5, the asymmetric reward function is set as follows: when the bridge safety assessment level predicted by the LightGBM model is lower than the actual risk level, a negative penalty that grows exponentially with the prediction error is applied; when the bridge safety assessment level predicted by the LightGBM model is higher than the actual risk level, a linear or constant penalty is applied.

[0011] Furthermore, by customizing the objective function of LightGBM, the first and second derivatives of the asymmetric reward function are calculated using the policy gradient algorithm, and the calculation results are fed back into the node splitting process of the decision tree in the LightGBM model to forcibly adjust the weight distribution of the leaf nodes in the decision tree.

[0012] Furthermore, before splitting the nodes of the decision tree, a Bayesian optimization algorithm is introduced to automatically search for at least one hyperparameter among the learning rate, tree depth, and number of leaf nodes in the LightGBM model, in order to optimize the optimization process of node splitting weights.

[0013] Furthermore, after assessing the technical condition of the bridge, the following also applies: The marginal contribution of each dimension of features in the multimodal engineering feature tensor is calculated using the SHAP game theory attribution algorithm, and the feature importance ranking is generated.

[0014] Furthermore, after assessing the technical condition of the bridge, the following also applies: The LightGBM model, which integrates reinforcement learning agents, is deployed in a microservice cluster based on an asynchronous non-blocking I / O architecture. It receives multi-source heterogeneous data from bridges collected in real time by sensors through an exposed RESTful API interface and performs real-time inference. When the inference indicates a high-risk state, an asynchronous non-blocking alarm task is triggered in the background.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs an asymmetric reward function. When the predicted bridge safety assessment level is lower than the actual risk level, a negative penalty that increases exponentially with the prediction error is applied; conversely, a linear or constant-level penalty is applied. This mechanism forces the LightGBM model, which incorporates a reinforcement learning agent, to apply larger gradient feedback in the direction of underestimating risk during training, thereby effectively suppressing the fatal error of misclassifying high-risk bridges as safe bridges.

[0016] 2. This invention constructs a deep learning model based on BERT-BiLSTM-CRF and employs a BIO sequence labeling strategy to perform named entity recognition on unstructured routine inspection text reports. Deep semantic encoding is performed using the BERT model, forward and backward sequence features are extracted through a bidirectional long short-term memory network, and global Viterbi-constrained decoding is performed through a conditional random field layer. This enables the accurate extraction of disease indicators containing bridge disease types, disease locations, and quantified values, while eliminating grammatical conflicts between quantified values ​​and their corresponding physical dimensions.

[0017] 3. This invention aligns and fuses extracted disease indicators, traffic load features extracted from dynamic weighing time-series traffic load data after filtering, and low-frequency regional meteorological and geological data on the time axis. Using the time axis of the dynamic weighing time-series traffic load data corresponding to the traffic load features as a reference, a forward-filling strategy is applied to the disease indicators and low-frequency regional meteorological and geological data to achieve alignment and fusion on the time axis. After being uniformly converted to float64 floating-point type, Z-score normalization is used to obtain a multimodal engineering feature tensor.

[0018] 4. After outputting the technical condition assessment level of the bridge, this invention uses the SHAP game theory attribution algorithm to calculate the marginal contribution of each dimension of the features in the multimodal engineering feature tensor, generating a ranking of feature importance. This mechanism can accurately quantify the contribution weight of each dimension of features to the determination of different risk levels, realizing the physical source of the assessment results.

[0019] 5. This invention deploys the LightGBM model, which integrates reinforcement learning agents, in a microservice cluster based on an asynchronous non-blocking I / O architecture. It receives multi-source heterogeneous bridge data collected in real-time by sensors through an exposed RESTful API interface for real-time inference. When the output indicates a high-risk state, a background asynchronous non-blocking alarm task is triggered. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 The diagram illustrates the process of using the BIO sequence labeling strategy to perform named entity recognition on unstructured periodic inspection text reports for deep learning models.

[0022] Figure 3 This is a schematic diagram of the feature importance ranking results in one embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] In the description of this invention, it should be noted that NLP stands for Natural Language Processing. In this invention, NLP is specifically used to identify and extract structured engineering information such as component location, defect type, and quantitative values ​​from unstructured bridge inspection texts. BERT stands for Bidirectional Encoder Representations from Transformers, a pre-trained language representation model based on the Transformer architecture; BiLSTM stands for Bidirectional Long Short-Term Memory, a variant structure of recurrent neural networks; and CRF stands for Conditional Random Field, a probabilistic graphical model used for sequence labeling tasks.

[0025] like Figure 1 As shown, this invention provides a large-sample bridge security assessment method based on NLP and reinforcement learning, comprising the following steps: S1. Acquire multi-source heterogeneous historical data of bridges, including unstructured scheduled inspection reports, dynamic weighing time-series traffic load data, and low-frequency regional meteorological and geological data.

[0026] The unstructured routine inspection reports are derived from Word or PDF format periodic inspection reports accumulated over the years by bridge maintenance platforms at all levels (such as the bridge historical asset management database), containing detailed natural language descriptions of bridge defects. Dynamic weighing time-series traffic load data are derived from piezoelectric WIM sensors (or dynamic weighing (WIM) systems) installed on the bridge deck and displacement sensors installed on the piers. WIM sensors can capture axle carrier-shaped signals from passing vehicles, with sampling frequencies ranging from 100Hz to 500Hz. Displacement sensors can measure the cumulative shear slip of the pier supports. Low-frequency regional meteorological and geological data (summarized monthly) are derived from the meteorological bureau's publicly available API and the bridge basic asset database, including information such as the region's average annual erosive rainfall (mm) and extreme freeze-thaw cycle days.

[0027] S2. Construct a deep learning model based on Bert-BiLSTM-CRF, and use the BIO sequence labeling strategy to perform named entity recognition on unstructured scheduled inspection text reports, extracting disease indicators that include bridge disease types, disease locations, and quantitative values.

[0028] The deep learning model based on BERT-BiLSTM-CRF consists of three cascaded core components: the input layer is a BERT (Deep Semantic Encoding) model, used to convert the original text into high-dimensional word vectors; the middle layer is a BiLSTM (Bidirectional Long Short-Term Memory) network, used to extract forward and backward contextual features from the word vector sequence; and the output layer is a CRF (Conditional Random Field), used to globally constrain and decode the output of the BiLSTM to generate the optimal BIO (Named Entity) labeled sequence. These three layers are connected sequentially to form an end-to-end named entity recognition model.

[0029] S3. Filter the dynamic weighing time-series traffic load data to extract traffic load characteristics.

[0030] S4. Align and integrate the extracted disease indicators with traffic load characteristics and low-frequency regional meteorological and geological data on the time axis to construct a multimodal engineering feature tensor.

[0031] S5. Input the multimodal engineering feature tensor into the LightGBM model that incorporates a reinforcement learning agent, and output the technical condition rating of the bridge by maximizing the asymmetric reward function.

[0032] The above steps S1 to S5 constitute the complete technical process of the bridge large-sample safety assessment method of this invention. Specifically, S2 and S3 can be executed in parallel, extracting defect indicators from unstructured routine inspection reports and traffic load features from dynamic weighing time-series traffic load data, respectively; S4 aligns and fuses the defect indicators extracted in S2, the traffic load features extracted in S3, and the low-frequency regional meteorological and geological data obtained in S1 to form a unified multimodal engineering feature tensor representing the bridge's health status; S5 uses a LightGBM model with a reinforcement learning agent to make decisions on this multimodal engineering feature tensor, maximizing the asymmetric reward function to output the bridge's technical condition rating.

[0033] In one embodiment of the present invention, when obtaining unstructured inspection text reports, optical character recognition (OCR) technology is used in advance to extract the text from the scanned image, and regular expressions are used to filter out interfering characters. At the same time, information involving personnel privacy and sensitive locations is automatically desensitized and cleaned.

[0034] Specifically, for inspection reports stored as scanned images, the OCR engine is first used to convert the images into an editable text format. Then, regular expressions are used to match and filter out non-character elements (such as headers and footers, table borders, and garbled characters). Finally, sensitive information that may be contained in the text, such as the names and contact information of the inspectors, and the specific latitude and longitude of the bridges, is automatically identified and anonymized to comply with data privacy protection regulations.

[0035] In one embodiment of the present invention, S2 specifically includes: S21. Use the BERT model in deep learning to perform deep semantic encoding on the scheduled text report to generate high-dimensional word vectors.

[0036] Traditional methods like Word2Vec cannot solve the problem of polysemy, while BERT can combine contextual information to map bridge defects into dense, high-dimensional word vectors. For example, it can accurately distinguish between the synonyms "crack" and "fissure," and can also identify the different semantic weights of "T-beam" and "box girder" in terms of structural stress.

[0037] S22 utilizes a bidirectional long short-term memory network in a deep learning model to extract forward and backward sequence features from high-dimensional word vectors.

[0038] Since Chinese descriptions of defects often involve long sentences, this invention employs a Bidirectional Long Short-Term Memory (BiLSTM) network to capture long-distance dependencies in the syntax. BiLSTM consists of two independent LSTM layers: a forward LSTM and a backward LSTM. These layers extract temporal features from both the forward and backward directions of the sequence, respectively, and then concatenate the features from both directions. This allows for the simultaneous capture of both the preceding and following contextual information for each element in the text sequence, effectively extracting both forward and backward features and clarifying the physical and spatial topological relationships between "defect types" (e.g., exposed reinforcement) and "specific components" (e.g., the cap beam of pier No. 3).

[0039] S23. Use the conditional random field in the deep learning model to learn the state transition probability matrix, perform global Viterbi constraint decoding on the sequence features, and output the BIO labeled sequence.

[0040] A Conditional Random Field (CRF) is introduced at the output of the BiLSTM layer, and a strict BIO (Begin, Inside, Outside) sequence labeling strategy is adopted. Figure 2As shown, the CRF layer, by learning the state transition probability matrix, can constrain the BIO annotation sequence globally. It should be noted that in the BIO annotation sequence, B represents the beginning of an entity (Begin), I represents the inside of an entity (Inside), and O represents a non-entity location (Outside); LOC represents the entity type of the component location, DIS represents the entity type of the defect type, and VAL represents the entity type of the quantified value. Therefore, B-LOC represents the beginning of the entity at the component location, I-LOC represents the inside of the entity at the component location, and so on.

[0041] BIO annotation sequence is a sequence of annotation results formed by annotating each character or word in the text report using the BIO sequence annotation strategy. For example, the CRF layer can force deep learning models to follow the syntax rule that "B-VAL (starting with a numerical value) can only be followed by I-VAL (inside the numerical value) and not B-LOC (starting with a position)," which can accurately segment natural language and assign corresponding labels (e.g., main span web (B-LOC) appears (O) longitudinal crack (B-DIS) width (O) 0.2mm (B-VAL)), thereby completely eliminating the serious syntax and dimensional conflict that traditional regular expression matching easily misattributes the numerical value "0.2mm" as crack length rather than crack width.

[0042] S24. Based on the BIO annotation sequence, extract the defect indicators, which include the defect type, defect location, and quantitative value of the bridge, from the scheduled inspection text report. Specifically, extract the text fragments corresponding to consecutive labels B-LOC and I-LOC in the BIO annotation sequence as component locations, extract the text fragments corresponding to consecutive labels B-DIS and I-DIS as defect types, and extract the text fragments corresponding to consecutive labels B-VAL and I-VAL as quantitative values, thereby forming structured defect indicators.

[0043] To verify the engineering effectiveness of this deep learning model, rigorous 10-fold cross-validation was performed on a self-built corpus containing over 100,000 real highway bridge inspection records. As shown in Table 1, the deep learning model performed excellently in F1 scores for various entity extraction types, especially achieving an accuracy of 98.2% in extracting the most critical quantitative numerical value (VAL) for engineering calculations, fully meeting the standards for high-precision engineering applications.

[0044] Table 1. Performance evaluation of deep learning models on a self-built bridge inspection corpus.

[0045] In one embodiment of the present invention, in S3, when filtering the dynamic weighing time-series traffic load data, a Butterworth bandpass filter is used for zero-phase filtering to remove environmental white noise and extract the vehicle impact equivalent axle load.

[0046] In complex field service environments, raw sensor signals are often overwhelmed by severe background noise, which comes from broadband disturbances from traffic flow, low-frequency disturbances from wind loads, and electromagnetic thermal noise from the sensor itself.

[0047] To address the aforementioned engineering challenge of spectral aliasing, this embodiment employs a Butterworth bandpass filter for filtering. According to digital signal processing theory, compared to Chebyshev or elliptic filters, the Butterworth filter exhibits a "maximally flat" amplitude response characteristic within the passband, meaning it can minimize amplitude distortion of the target characteristic waveform. In practice, the cutoff frequency of the bandpass filter is not hard-coded but rather adaptively and dynamically calibrated based on the finite element modal analysis results of the bridge.

[0048] Meanwhile, in order to completely eliminate the phase delay that the infinite impulse response filter inevitably brings to causal filtering, this invention adopts zero-phase filtering technology, that is, the signal passes through the filter once in the forward direction and once in the reverse direction, so that the phase distortion strictly cancels each other out.

[0049] The specific implementation is as follows: First, the bandpass filter parameters are dynamically set based on the fundamental frequency and higher-order modes of the specific bridge. It is assumed that finite element analysis determines the target damage-sensitive frequency characteristics to be concentrated in the 0.5Hz to 10.0Hz frequency band. After calculating the folding frequency according to the Nyquist theorem, the physical frequency is normalized to between 0 and 1 to adapt to the digital filter design. A fourth-order Butterworth bandpass filter is used for filtering. In the specific implementation, the cut-off frequencies of the bandpass filter are not hard-coded, but rather adaptively and dynamically calibrated based on the finite element modal analysis (FEM) results of the bridge.

[0050] Secondly, a zero-phase forward-backward filtering technique is adopted, in which the signal passes through the filter once in the forward direction and once in the reverse direction, so that the phase distortions strictly cancel each other out, ensuring that the precise timestamp of the damage in the time domain does not drift.

[0051] Finally, multidimensional statistical features were extracted from the filtered clean signal to represent the equivalent axle load of vehicle impact, and a structural health fingerprint dictionary was constructed. The extracted features include: peak amplitude, which reflects the instantaneous response to extreme impact loads such as heavy-duty trucks crossing bridges; root mean square value, which reflects the total energy level of the structure under continuous excitation vibration; crest factor, which is the ratio of peak value to effective value, and this index is extremely sensitive to early fatigue microcracks and local collisions; and signal variance, which reflects the severity of time-series fluctuations and the system stability margin.

[0052] In one embodiment of the present invention, data from different frequency sources (disease indicators extracted from unstructured routine inspection text reports, dynamic weighing time-series traffic load data, and low-frequency regional meteorological and geological data) need to be aligned and fused on the time axis.

[0053] In bridge integrated monitoring systems, sensor networks typically include various devices with vastly different sampling frequencies. For example, the sampling rate of a dynamic weighing system may be as high as 100Hz to 500Hz, while the update cycle for disease indicators is usually quarterly or annual, and the update cycle for low-frequency regional meteorological and geological data is usually monthly or annual. This significant "multi-source heterogeneity" and "asynchronous non-co-frequency" characteristic leads to serious data misalignment problems when constructing multi-dimensional feature matrices.

[0054] Therefore, this embodiment abandons the backfilling or linear interpolation strategies commonly used in traditional machine learning. In the physical world of bridge engineering, structural degradation, without human intervention and maintenance, is an irreversible, unidirectional physical process. Using linear interpolation introduces the illusion of a non-existent "smooth recovery," while using backfilling severely leaks future information. Therefore, this invention enforces a forward fill strategy that strictly adheres to temporal causality. Specifically, using the time axis of dynamic weighing time-series traffic load data as a benchmark, a forward fill strategy is applied to the defect indicators and low-frequency regional meteorological and geological data, extending the low-frequency data to the time axis of the high-frequency data, achieving resampling alignment on the time axis.

[0055] After timeline alignment, the aligned and fused disease indicators, traffic load characteristics, and low-frequency regional meteorological and geological data were uniformly converted to float64 floating-point type and normalized using Z-score standardization to obtain multimodal engineering feature tensors. During this process, to prevent gradient explosion due to differences in data magnitude, all multimodal features were uniformly converted to high-precision float64 floating-point type and normalized using Z-score standardization.

[0056] The feature matrix of the multimodal engineering feature tensor constructed in this embodiment is defined as shown in Table 2. This feature matrix covers multiple dimensions such as text semantics (e.g., sentiment score of defects in routine inspection texts, maximum longitudinal crack width of main beams), dynamic loads (e.g., WIM cumulative heavy load equivalent, cumulative shear slip of supports), meteorological environment (e.g., regional average annual erosive rainfall), and engineering priors (e.g., service age reduction and design load).

[0057] Table 2. Definition of the characteristic matrix of the multimodal engineering characteristic tensor

[0058] It should be noted that the service age reduction period and design load are prior engineering knowledge and can be added as optional feature dimensions to the multimodal engineering feature tensor to further enhance the accuracy of the evaluation model.

[0059] To address the "severe class imbalance" and the resulting "high-risk underreporting" defects faced by traditional machine learning in bridge assessment, this invention abandons the traditional symmetric cross-entropy loss function and instead defines the assessment of bridge technical condition level as a Markov decision process, and introduces an asymmetric penalty mechanism based on reinforcement learning.

[0060] Within the framework of Markov decision processes, this invention defines the following three core elements: State space: The multimodal engineering feature tensor after fusion and alignment. This tensor integrates disease indicators, traffic load characteristics, and low-frequency regional meteorological and geological data, representing a snapshot of the bridge's overall health at the current moment.

[0061] Action Space: The bridge technology status assessment level, which is the output of the LightGBM model with reinforcement learning agent, is a discrete set of actions corresponding to five risk levels from Class I to Class V.

[0062] Asymmetric reward function: During the training and exploration process of a reinforcement learning agent, the environment will be based on the true level labels annotated by experts. y and the action level predicted by the model a The feedback will include the following reward or punishment signals: .

[0063] In the formula: when a=y When the prediction is completely accurate, a positive reward is given. when a>y When this is the case, it indicates an overestimation of risk, and only a mild linear penalty is imposed. when a <y When this occurs, it indicates an underestimation of risk (i.e., underreporting), resulting in a severe exponential penalty.

[0064] a and β These are hyperparameters that control the steepness of the exponential and linear penalties, respectively. Preferably, in this invention… a= 1.2, β= 1.5.

[0065] The core feature of the aforementioned asymmetric reward function is that it applies asymmetric penalty strength to different types of prediction errors. Specifically, when the bridge safety assessment level predicted by the LightGBM model is lower than the actual risk level (i.e., underestimating or underreporting risk), a negative penalty is applied that grows exponentially with the prediction error; when the bridge safety assessment level predicted by the LightGBM model is higher than the actual risk level (i.e., overestimating risk), only a linear or constant penalty is applied; and when the bridge safety assessment level predicted by the LightGBM model is equal to the actual risk level, a positive reward is given.

[0066] The specific process of maximizing the asymmetric reward function to output the technical condition rating of a bridge in this invention is as follows: For the input multimodal engineering feature tensor, the LightGBM model, which incorporates a reinforcement learning agent, calculates the expected cumulative reward for each possible technical condition rating (classes I to V). The asymmetric reward function assigns different reward or penalty weights to different prediction results: a positive reward for correct prediction, a mild negative penalty for overestimating risk, and an exponentially severe negative penalty for underestimating risk (missed detection).

[0067] By defining a custom objective function for LightGBM, the first and second derivatives of the asymmetric reward function are calculated using the policy gradient algorithm. These results are then fed back into the node splitting process of the decision tree in the LightGBM model to force an adjustment to the weight distribution of the leaf nodes. Specifically, the asymmetric reward function is embedded into LightGBM's custom objective function. At each decision tree node split, the first and second gradients of this objective function with respect to the LightGBM model's predicted output are calculated. These gradient values ​​reflect the marginal impact of adjusting model parameters on the asymmetric reward function under the current prediction state.

[0068] Because the asymmetric reward function imposes an exponential penalty on the false negative error, the magnitudes of the calculated first and second gradients will increase dramatically when the LightGBM model mistakenly classifies a dangerous bridge as a safe one. These gradients are fed back into the node splitting process of the decision tree. The histogram algorithm, when selecting the optimal split point, prioritizes feature dimensions and splitting thresholds that significantly reduce the false negative error, thereby forcibly adjusting the weight distribution of the leaf nodes.

[0069] Through multiple rounds of iterative training, the LightGBM model learns a decision-making strategy: when faced with uncertain samples, it tends to output a higher rating (i.e., a conservative approach), because overestimating risk is only punished linearly, while underestimating risk is punished exponentially. Ultimately, for a given input feature tensor, the model outputs a technical condition rating that maximizes the expected value of the asymmetric reward function, thereby achieving the engineering safety goal of "better to report a false alarm than to miss a report."

[0070] Damage evolution in bridge structures is a highly nonlinear dynamic process, influenced by complex coupling effects from various external environments. In traditional AI modeling, a single model often struggles to handle data of different forms: pure deep learning performs well when processing high-frequency continuous time-series waveforms, but it is prone to overfitting and exhibits extremely poor generalization ability when processing discrete, sparse, static tabular data.

[0071] In one embodiment, the present invention proposes a divide-and-conquer and fusion deep feature hybrid architecture as a preferred internal implementation of a LightGBM model that incorporates reinforcement learning agents.

[0072] First, a deep residual neural network built on the PyTorch framework is used to abstract and extract deep implicit feature representations from high-frequency traffic load features layer by layer. The structure of this deep residual network is as follows: the input layer receives traffic load features, and the input space is mapped to a high-dimensional feature space by an initial feature mapping layer. To overcome the gradient vanishing and internal covariate shift caused by network deepening, a batch normalization layer is embedded in the network to pull the feature distribution back to a standard normal distribution, thereby accelerating the convergence speed of gradient descent; the ReLU activation function is used to introduce nonlinearity while avoiding the gradient vanishing problem caused by the Sigmoid function; a dropout mechanism with a dropout rate of 30% is introduced, which randomly cuts the connections of some neurons during each forward propagation, forcing the network to learn more robust redundant features and enhancing the model's fault tolerance when there is noise or local failure in the sensor; then, a deep feature compression layer extracts more high-order, abstract essential features of structural damage; finally, the decision output layer maps to a specified number of risk levels.

[0073] Secondly, static, discrete structural attribute data is processed by LightGBM, an ensemble learning model based on gradient boosting decision trees (GBDT). Finally, the high-order feature vectors extracted by the deep residual neural network (i.e., deep implicit feature representations abstracted layer by layer from traffic load features) are concatenated with the tabular features obtained after LightGBM processes the static, discrete data, and then fed into a fusion layer for joint decision-making. This hybrid approach of "representation learning + gradient boosting" perfectly balances a keen awareness of nonlinear dynamic features with extreme robustness to high-dimensional sparse tabular data.

[0074] Because the hyperparameter space of the LightGBM model, which incorporates reinforcement learning agents, is enormous (including the learning rate and hidden layer dimension of the deep network, as well as the tree depth, number of leaf nodes, and feature sampling rate of LightGBM), it exhibits typical high-dimensional, non-convex, and non-differentiable characteristics. Traditional grid search suffers from a severe "curse of dimensionality," leading to the exhaustion of computing power in the computing cluster; while random search relies entirely on luck and is inefficient.

[0075] To overcome the computing power bottleneck, in one embodiment, before splitting the nodes of the decision tree, a Bayesian optimization algorithm is introduced to automatically search for at least one hyperparameter among the learning rate, tree depth, and number of leaf nodes in the LightGBM model, so as to optimize the optimization process of node splitting weights.

[0076] Specifically, in this embodiment, the Optuna automated machine learning framework is used, and Bayesian optimization is implemented using the Tree-structured Parzen Estimator (TPE) algorithm. The TPE algorithm does not perform a blind search; it establishes a surrogate model to evaluate the posterior probability distribution of hyperparameter combinations, intelligently guiding the search towards regions with better historical evaluation results and lower exploration levels.

[0077] More importantly, this scheme combines median pruning and early stopping mechanisms: in the early stages of training, if the performance of a certain hyperparameter combination is found to be significantly lower than the median of all historical trials, the algorithm will terminate that trial, thereby concentrating computational resources on those truly promising hyperparameter combinations. This mechanism significantly shortens the overall model tuning cycle.

[0078] In this embodiment, the Bayesian search space for hyperparameters is first dynamically defined. For the LightGBM model, the objective function is set as a multi-class classification task with 5 risk levels, the evaluation metric is multi-class log loss, and a gradient boosting decision tree is used as the boosting type.

[0079] The Bayesian search space for hyperparameters is as follows: the number of leaf nodes adopts a uniform integer distribution, with a search range of 20 to 150; the learning rate adopts a uniform logarithmic distribution, with a search range of 1e-4 to 0.1, because the sensitivity of the learning rate differs at low and high orders of magnitude; the feature sampling rate adopts a uniform distribution, with a search range of 0.6 to 1.0, to prevent a single tree from over-relying on certain strong features and to enhance the model's robustness; the minimum number of samples per leaf node adopts a uniform integer distribution, with a search range of 10 to 100, which is a key parameter to prevent overfitting.

[0080] The training and validation data are converted to LightGBM's underlying Dataset binary format to accelerate memory access. During training, Optuna's integrated pruning callback function is injected to monitor the multi-class log loss metric on the validation set and perform early blocking and median pruning.

[0081] Model training is performed, with a maximum of 1000 iterations. After training, the predicted probability matrix on the validation set is calculated, and the macro-weighted AUC is used as the final evaluation metric to effectively address the problem of imbalanced samples at different risk levels. The AUC score is returned to the TPE algorithm to guide its next iteration search.

[0082] In the field of transportation infrastructure involving public safety, the black-box nature of machine learning models limits their engineering applications. When the LightGBM model outputs an assessment result of a bridge's high-risk condition, if it cannot provide specific characteristic evidence and damage location information that led to the assessment result, engineers are unlikely to accept the model's output conclusions, thus affecting the actual application effect of the monitoring system.

[0083] To address this, this invention integrates a SHAP interpretability module based on cooperative game theory into the final prediction output. The mathematical essence of SHAP is to calculate the expected marginal contribution of each input feature to the final prediction result across all possible feature combinations. Through the SHAP algorithm, the influence weight of each feature dimension on the final risk level determination can be accurately quantified. By automatically generating high-dimensional beehive or waterfall plots, engineers can intuitively see whether an increase in a certain feature value drives an increase or decrease in the risk level.

[0084] In one embodiment of the present invention, after feature fusion is completed and the RL-LightGBM evaluation model is constructed, the SHAP algorithm is used to calculate the marginal contribution of each dimension of the features in the multimodal engineering feature tensor, generating a feature importance ranking. For example... Figure 3The multimodal features of the bridge condition shown, ranked by feature importance weights, indicate that load features corresponding to dynamic weighing time-series traffic load data (e.g., WIM cumulative load equivalent) and defect indicators extracted from unstructured routine inspection reports (e.g., maximum longitudinal crack in the main girder) have the highest weights in the final safety assessment results. This conclusion is highly consistent with the mechanisms of "fatigue cumulative damage" and "degradation of main load-bearing components" in bridge engineering mechanics, proving the scientific validity and effectiveness of the constructed feature matrix.

[0085] Under normal circumstances, the data flow of bridge monitoring systems is relatively stable. However, when faced with extreme emergencies such as super typhoons, destructive earthquakes, or dense convoys of heavy vehicles crossing bridges, tens of thousands of sensor nodes on dozens or even hundreds of bridges in the area will simultaneously send massive amounts of alarm data streams to the cloud control center. If a traditional synchronous blocking web framework is used, the main thread will be blocked by slow I / O operations such as database writes or third-party SMS API calls, which can easily lead to thread pool exhaustion, service avalanche crashes, and ultimately, missed alarms.

[0086] To cope with high concurrency, this invention adopts the FastAPI high-performance asynchronous framework based on the ASGI specification at the cloud microservice deployment level. FastAPI utilizes Python 3's native async / await coroutine mechanism and event loop. When processing network requests, database I / O, or waiting for model inference results, the current coroutine will actively relinquish CPU control, enabling a single node to handle a throughput of tens of thousands of QPS.

[0087] At the data security and business decoupling level, the system uses the Pydantic library to perform strong type validation of input data, blocking malformed data or malicious injection. At the same time, it uses the BackgroundTasks component built into FastAPI to push time-consuming and latency-tolerant non-core link tasks such as issuing warning SMS messages and dispatching work orders into the background asynchronous coroutine queue for execution, ensuring that the core data receiving and model inference interfaces can return responses in milliseconds, ensuring the smooth operation of the IoT link.

[0088] In one embodiment of the present invention, after outputting the technical condition assessment level of the bridge, the method further includes: deploying a LightGBM model with reinforcement learning agent in a microservice cluster based on an asynchronous non-blocking I / O architecture, receiving multi-source heterogeneous bridge data collected in real time by sensors through an exposed RESTful API interface for real-time inference; and triggering a background asynchronous non-blocking alarm task when the output is determined to be in a high-risk state.

[0089] In one embodiment of the present invention, after outputting the technical condition assessment level of the bridge, the method further includes: deploying a LightGBM model with reinforcement learning agent in a microservice cluster based on an asynchronous non-blocking I / O architecture, receiving multi-source heterogeneous bridge data collected in real time by sensors through an exposed RESTful API interface for real-time inference; and triggering a background asynchronous non-blocking alarm task when the output is determined to be in a high-risk state.

[0090] It should be noted that the multi-source heterogeneous data of the bridge corresponds to the three types of data in S1, including: defect indicators extracted from unstructured scheduled inspection reports, traffic load characteristics extracted from dynamic weighing time-series traffic load data after filtering, and low-frequency regional meteorological and geological data. Specifically, the dynamic weighing time-series traffic load data is collected in real time by WIM sensors installed on the bridge; defect indicators are updated in real time by the monitoring system based on the scheduled inspection reports; and low-frequency regional meteorological and geological data is obtained by calling the meteorological bureau's API or querying geological databases. All of the above data is uniformly sent to the cloud microservice interface for real-time inference.

[0091] Specifically, a high-performance asynchronous framework based on the ASGI specification, FastAPI, is used to build cloud-based microservices. The FastAPI core application is initialized, automatically generating interactive interface documentation compliant with the OpenAPI specification. A strongly typed data gateway is built using Pydantic to define the bridge telemetry data model, including fields such as a unique bridge identifier, ISO 8601 standard timestamp, real-time observation of the maximum deflection of the main span, and the total energy integral value of structural vibration. The physical rationality of key indicators is then verified.

[0092] Define asynchronous, time-consuming tasks, such as calling a telecom operator's API to send emergency SMS messages or writing to a persistent log database. In the core high-concurrency data reception and prediction routing, validated clean data is input into a LightGBM model with a reinforcement learning agent for inference. When the model's inference output determines a high-risk state, the warning circuit breaker and degradation logic are triggered, pushing the time-consuming SMS notification and work order dispatch tasks into a background asynchronous coroutine queue for execution via the BackgroundTasks component. The core inference interface immediately returns a response.

[0093] At the same time, a global exception capture mechanism is set up to capture unpredictable engine exceptions, return standardized error codes, and trigger the circuit breaker or retry mechanism of the front-end monitoring panel and load balancer to prevent the cascading spread of faults.

[0094] To verify the effectiveness of this invention under complex working conditions, offline retrospective simulation tests were conducted using a historical anonymized dataset provided by a municipal smart transportation platform. This test set covered 1523 existing highway bridges under the city's jurisdiction, encompassing common bridge types with service lives ranging from 5 to 30 years (including simply supported hollow slab bridges, continuous beam bridges, and variable cross-section box girder bridges), and included various conventional and extreme working conditions such as heavy traffic and environmental erosion.

[0095] Under the traditional maintenance model, completing a comprehensive routine inspection of a large-scale bridge group is time-consuming and easily affected by subjective differences in manual inspection.

[0096] Using the method proposed in this invention, by connecting the bridge historical asset management database with the WIM dynamic weighing system, automated extraction of defect indicators and fusion calculation of multimodal engineering features were achieved. On edge computing nodes equipped with conventional industrial-grade edge AI accelerator cards (such as NVIDIA Jetson series or Tesla T4), the safety status assessment of all 1523 bridges can be completed within minutes, effectively improving assessment efficiency and providing objective data support for the initial screening by maintenance departments. Specific performance comparison results are shown in Table 3.

[0097] Table 3 Performance Comparison of Different Evaluation Algorithms / Methods

[0098] As shown in Table 3, the present invention achieves an overall accuracy of 91.3% in the five-category classification, which is superior to the 82.5% of traditional manual on-site inspection, the 78.4% of random forest, and the 88.6% of traditional LightGBM. In terms of evaluation time, the present invention requires only 8.5 seconds per 100 bridges, which is far lower than the 7200 seconds of traditional manual inspection. In terms of the core safety indicator of the false negative rate of high-risk bridges, the present invention achieves 0.5%, which is significantly better than the 4.2% of traditional manual inspection, the 12.5% ​​of random forest, and the 7.8% of traditional LightGBM.

[0099] Taking the actual evolution trajectory of a T-beam bridge across a river on a national highway in the aforementioned historical desensitized dataset as an example: when the system was back-processing the flood season data in August of a certain year, it triggered a Class V high-risk warning for the bridge, and the model output confidence level was 94.2%.

[0100] The underlying SHAP traceability of the system shows that the main triggering factor for the warning stems from the superposition of multiple features: on the one hand, the historical inspection reports of the bridge from the previous two years show that "there is a 0.35mm unclosed shear crack in the web at the mid-span"; on the other hand, the WIM data stream of the dynamic weighing time-series traffic load data shows that the bridge experienced a dense flow of overweight trucks in the 72 hours before the warning was triggered, and was in a rainy environment.

[0101] After verifying the relevant road maintenance and repair records, it was confirmed during a manual emergency inspection of the bridge after this period that the main beam web plate had developed diagonal cracks and the bearings were damaged, indicating that the bridge was in a dangerous state of significant deterioration in its load-bearing capacity. Traffic control and reinforcement were subsequently implemented.

[0102] This offline retrospective case demonstrates that the present invention can combine multi-source data to discover hidden dangers coupled with sudden changes that are difficult to detect with single-source data, providing early safety warnings for maintenance departments.

[0103] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention used to illustrate the technical solutions of the present invention, and are not intended to limit the invention, nor are they intended to limit the patent scope of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. That is to say, any changes or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but whose technical problems are still consistent with the present invention, should be included within the protection scope of the present invention. In addition, the direct or indirect application of the technical solutions of the present invention to other related technical fields are similarly included within the patent protection scope of the present invention.

Claims

1. A large-sample safety assessment method for bridges based on NLP and reinforcement learning, characterized in that, Includes the following steps: S1. Acquire multi-source heterogeneous historical data of bridges, including unstructured scheduled inspection reports, dynamic weighing time-series traffic load data, and low-frequency regional meteorological and geological data; S2. Construct a deep learning model based on Bert-BiLSTM-CRF, and use the BIO sequence labeling strategy to perform named entity recognition on unstructured inspection text reports, extracting disease indicators that include bridge disease types, disease locations and quantitative values. S3. Filter the dynamic weighing time series traffic load data to extract traffic load characteristics; S4. Align and fuse the extracted disease indicators with traffic load characteristics and low-frequency regional meteorological and geological data on the time axis to construct a multimodal engineering feature tensor; S5. Input the multimodal engineering feature tensor into the LightGBM model that incorporates a reinforcement learning agent, and output the technical condition rating of the bridge by maximizing the asymmetric reward function; In S5, the asymmetric reward function is set as follows: when the bridge safety assessment level predicted by the LightGBM model is lower than the actual risk level, a negative penalty that grows exponentially with the prediction error is applied; when the bridge safety assessment level predicted by the LightGBM model is higher than the actual risk level, a linear or constant penalty is applied; when the bridge safety assessment level predicted by the LightGBM model is equal to the actual risk level, a positive reward is given. By customizing the objective function of LightGBM, the first and second derivatives of the asymmetric reward function are calculated using the policy gradient algorithm. The calculation results are then fed back into the node splitting process of the decision tree in the LightGBM model to forcibly adjust the weight distribution of the leaf nodes in the decision tree.

2. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, In S1, when obtaining unstructured inspection text reports, optical character recognition technology is used in advance to extract the text from the scanned image, and regular expressions are used to filter out interfering characters. At the same time, information involving personnel privacy and sensitive locations is automatically desensitized and cleaned.

3. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, S2 specifically includes: The BERT model in deep learning is used to perform deep semantic encoding on the fixed-time text report to generate high-dimensional word vectors. The bidirectional long short-term memory network in the deep learning model is used to extract forward and backward sequence features from high-dimensional word vectors; By using the conditional random field in the deep learning model to learn the state transition probability matrix, global Viterbi constraint decoding is performed on the sequence features to output the BIO labeled sequence. Based on the BIO annotation sequence, defect indicators containing the type, location, and quantitative value of bridge defects are extracted from the scheduled inspection text report.

4. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, In S3, when filtering the dynamic weighing time-series traffic load data, a Butterworth bandpass filter is used for zero-phase filtering to remove environmental white noise and extract the vehicle impact equivalent axle load.

5. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, S4 specifically includes: Using the time axis of the dynamic weighing time series traffic load data corresponding to the traffic load characteristics as a benchmark, a forward filling strategy is implemented on the disease indicators and low-frequency regional meteorological and geological data to achieve alignment and fusion on the time axis. The aligned and fused disease indicators, traffic load characteristics, and low-frequency regional meteorological and geological data were uniformly converted into float64 floating-point type and normalized using Z-score standardization to obtain multimodal engineering feature tensors.

6. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, Before splitting nodes in the decision tree, a Bayesian optimization algorithm is introduced to automatically search for at least one hyperparameter in the LightGBM model, including the learning rate, tree depth, and number of leaf nodes, in order to optimize the process of finding the node splitting weights.

7. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, After assessing the technical condition of the bridge, the following are also included: The marginal contribution of each dimension of features in the multimodal engineering feature tensor is calculated using the SHAP game theory attribution algorithm, and the feature importance ranking is generated.

8. The bridge large-sample safety assessment method based on NLP and reinforcement learning according to claim 1, characterized in that, After assessing the technical condition of the bridge, the following are also included: The LightGBM model, which integrates reinforcement learning agents, is deployed in a microservice cluster based on an asynchronous non-blocking I / O architecture. It receives multi-source heterogeneous data from bridges collected in real time by sensors through an exposed RESTful API interface and performs real-time inference. When the inference indicates a high-risk state, an asynchronous non-blocking alarm task is triggered in the background.

Images