A dynamic bayesian network construction risk prediction method based on sliding window adaptive learning and particle filtering

Abstract

The application relates to a dynamic Bayesian network construction risk prediction method based on a sliding window adaptive learning and a particle filter, and relates to the technical fields of engineering construction safety management and intelligent prediction. The method comprises the following steps: inputting a risk factor state; constructing a dynamic Bayesian network topological structure; updating a CPT based on a sliding window; recursively reasoning a risk state based on a particle filter; predicting a risk based on a forward Monte Carlo propagation; and evaluating a risk grade. The method can efficiently establish a precise time sequence mapping relationship between a risk state and multi-source observation, and realizes global state estimation and uncertainty quantification through a dynamic Bayesian network combined with a particle filter, and realizes rapid response and online update of distribution drift through a sliding window adaptive learning, so that the prediction efficiency and precision and the early warning timeliness are comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of engineering construction safety management and intelligent prediction technology, specifically a dynamic Bayesian network construction risk prediction method based on sliding window adaptive learning and particle filtering. Background Technology

[0002] Construction risk prediction, as a key technical aspect of modern engineering construction safety management, is widely used in various complex working conditions and large- and medium-sized engineering projects due to its supporting role in accident prevention, resource allocation, and organizational coordination. On-site data sources encompass multi-channel information flows including IoT sensing, machinery operation, environmental monitoring, and progress and quality management, exhibiting characteristics of high frequency, heterogeneity, time-varying nature, and incompleteness. Therefore, how to accurately predict and promptly warn of risk states during construction under the background of non-stationary data flows and noise interference has become a key technical issue for ensuring construction safety and improving management efficiency.

[0003] In existing technologies, construction risk prediction typically employs rule-based judgments based on standard thresholds and expert experience, classification / regression models trained offline on historical samples, and causal and temporal tools such as static Bayesian networks, fault trees, or hidden Markov models. Engineering practice often involves establishing static or quasi-static models and performing parameter calibration, threshold tuning, and iterative calculations to obtain risk scores and early warning strategies. However, high-dimensional, multi-source data requires processing and iteration in large-scale scenarios, placing high demands on computing resources and consuming significant time, making it difficult to meet the needs of rapid on-site response. Due to model simplification and parameter uncertainty, estimates under static or linear assumptions are prone to deviating from actual working conditions, resulting in insufficient accuracy and stability of early warnings. Risk threshold and strategy adjustments mainly rely on empirical trial and error, lacking systematization and automation, and easily introducing human error. Complex nonlinear, time-varying, and partially observable mapping relationships exist between the risk state during the construction phase and multi-source observations, which traditional methods struggle to accurately characterize and quantify for uncertainty. As projects expand in scale and become more digitalized, risk prediction problems are increasingly characterized by multi-dimensionality, high complexity, and conceptual drift. Fixed window and conventional parameter update strategies are prone to getting stuck in local optima or model aging in high-dimensional search and dynamic environments, making it difficult to guarantee the efficiency and quality of early warnings.

[0004] In summary, existing technologies for construction risk prediction suffer from drawbacks such as long computation time, insufficient accuracy and robustness, reliance on manual threshold tuning, and difficulty in handling nonlinear and non-stationary problems. Therefore, a new intelligent prediction method is urgently needed to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a construction risk prediction method based on a dynamic Bayesian network using sliding window adaptive learning and particle filtering. This method can efficiently establish a precise temporal mapping relationship between risk states and multi-source observations. By combining a dynamic Bayesian network with particle filtering, it achieves global state estimation and uncertainty quantification. Furthermore, the sliding window adaptive learning enables rapid response and online updates to distribution drift, thereby comprehensively improving prediction efficiency, accuracy, and early warning timeliness.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a dynamic Bayesian network construction risk prediction method based on sliding window adaptive learning and particle filtering, comprising the following steps:

[0007] Step 1: Input risk factor status;

[0008] ① Construct risk factor status: Determine and define risk factors based on the actual construction scenario requirements. The overall state vector of all risk factors at time 1 ;

[0009] ② Constructing a discrete state set: Establish a discrete state set for each risk factor. The specific states of risk factors are all taken from the corresponding discrete state set;

[0010] Step 2: Constructing the dynamic Bayesian network topology;

[0011] ① A two-slice DBN structure is adopted: adjacent time slices are used. The structure is used to model and represent the time series process, treating each moment as a time slice. Within a time slice, the state of risk factors at that moment is considered a variable. Within the same time slice, acyclic unidirectional edges represent the immediate effects between variables. Across time slices, only the state from... point to The forward edges divide the parent nodes corresponding to the target risk factor into cross-slice parent nodes and same-slice parent nodes. The cross-slice parent node is the state of the target risk factor itself in the previous time slice, and the same-slice parent node is the state of the risk factor that has a direct impact on the state of the target risk factor in the current time slice.

[0012] ② Temporal decomposition of joint distribution: Define a joint distribution for the entire time period and decompose it into an initial joint distribution. With the transfer distribution The product;

[0013] ③ Determination of the initial joint distribution: Calculate the conditional distribution of risk factor states under a given parent configuration based on the historical sample set, and construct the initial CPT;

[0014] ④ Determining the transition distribution: Decompose the joint distribution according to the topological structure to obtain the distribution under a given parent configuration. Conditional distribution of risk factor states at any given time;

[0015] Step 3: CPT update based on sliding window;

[0016] ① Define a sliding window: Set a fixed-length window that shifts over time, and only samples within the window participate in CPT updates; Define an equivalent count baseline to connect the initial CPT with the weighted count recursion;

[0017] ②Weighted count recursion: Apply a forgetting factor to the historical weighted count to decay it, and then add the count of newly added samples in the sliding window to obtain the current weighted count;

[0018] ③Conditional probability estimation: Smooth the weighted count, calculate the conditional probability under a given parent configuration, and write the result back to update the CPT;

[0019] Step 4: Risk state recursive reasoning based on particle filtering;

[0020] ①State observation model:

[0021] Will A quantitative relationship is established between the risk factor state at any given time and the observation results collected by the sensing system, and a likelihood observation model is constructed.

[0022] ②Particle prediction: Based on the transition distribution between two adjacent time slices, from The particle state sampling at time t gives a given particle Predicted samples at each time point;

[0023] ③Weight update: Weights are calculated based on the likelihood observation model to determine the unnormalized weights of the particles, and then each weight is normalized.

[0024] ④ Degradation detection and resampling: Calculate the number of valid samples. When the number of valid samples is less than the resampling threshold, perform resampling and reset the weights.

[0025] ⑤ Posterior approximation: The posterior distribution is approximated as a finite number of particles and their weights, thereby expressing the probability distribution through a computable set of samples;

[0026] Step 5: Risk prediction based on forward Monte Carlo propagation;

[0027] ① Particle forward roll: For each particle, starting from the current state, the future is obtained by progressively sampling according to the transition model. The forward path of time, To predict the step size;

[0028] ② Probability of future events: A weighted statistical set of particles yields the probability of future events. The predicted probability of the state of the risk factor;

[0029] ③ Output: Obtain the probability result corresponding to the earliest out-of-limit step. ;

[0030] Step Six: Risk Level Assessment;

[0031] ① Severity of Loss Score: Based on economic loss, downtime, and personal injury, scores are calculated for each item and weighted to obtain a comprehensive score. ;

[0032] ② Risk Value Calculation and Classification: Defining Risk Value ,according to The value determines the risk level.

[0033] Furthermore, in step one, the overall state vector With discrete state set They are represented as follows:

[0034]

[0035]

[0036] In the formula, Indicates the first Each risk factor in The state at any given moment, , For time slice index, , express The A specific state, .

[0037] Furthermore, in step two, the process of determining the initial joint distribution and transition distribution through joint distribution time series decomposition is as follows:

[0038] In the topology, the state of a cross-slice parent node is represented as: The state of the parent node in the same piece is represented as ;

[0039] The joint distribution expression for the entire time period is defined as follows:

[0040]

[0041] In the formula, Indicates the first to the last The complete state record of risk factors for each time slice, with the risk factor state vectors within each time slice arranged in chronological order. This represents the joint probability distribution of the state vectors of all risk factors over all time periods.

[0042] The initial joint distribution is represented as follows:

[0043]

[0044] Let the historical sample set used for initialization be . Then we have:

[0045]

[0046] In the formula, for The Middle The state of each risk factor, in Timing calculations are more efficient than cross-slice parent configurations. In order to be in The number of samples that meet the specific state requirements at any given time. The smoothing parameter is then substituted into the specific state to obtain the corresponding conditional probability. For different parent configurations and child values, the obtained conditional probabilities are filled in with the parent configuration as column data and the child value as row data. The sum of the probabilities in each row is equal to 1, thus obtaining the initial CPT.

[0047] The transition distribution is represented as follows:

[0048]

[0049] in, For a given parent configuration The conditional distribution.

[0050] Furthermore, in step three, the equivalent counting baseline is represented as:

[0051]

[0052] In the formula, As the initial row for the weighted counting recursion, This represents the total weight of the row;

[0053] The weighted counting recursive representation is as follows:

[0054]

[0055] In the formula, In order to be in The number of samples that meet the specific state requirements at any given time. As of Historical weighted count of moments Forgetting factor;

[0056] Conditional distribution Smoothing is performed as follows:

[0057]

[0058] In the formula, As of The weighted count that satisfies the specific state requirements at each moment is then substituted into the specific state to obtain the corresponding conditional probability, and the updated count is then used to calculate the conditional probability. Write the results back to update CPT.

[0059] Furthermore, in step four, the state observation model and particle prediction specifically include:

[0060] The likelihood observation model is defined as follows:

[0061]

[0062] In the formula, For the observation result vector, For the first Observation results of each channel, ;

[0063] In a given particle After the state at time point, according to the transition distribution of two adjacent time slices, in topological order... Factor-wise sampling yields the following predicted sample for a given particle at this moment:

[0064]

[0065] In the formula, Indicates the first The particle corresponds to the first Each risk factor in The state at time t is the prediction sample. Indicates the first The particle corresponds to the first Each risk factor in The state at time t is a priori sample. The number of particles.

[0066] Furthermore, in step four, the degradation discrimination, resampling, and posterior approximation specifically involve:

[0067] Define the number of valid samples It is expressed as follows:

[0068]

[0069] In the formula, for The normalized weights of particles at time t, when , If the resampling threshold is reached, then resampling is performed and the weights are reset to [value]. ;

[0070] The probability distribution is expressed using a computable set of samples as:

[0071]

[0072] In the formula, For the specific state of the target, It is an indicator function if and only if it satisfies The indicator function is set to 1 if the condition is met, and 0 otherwise.

[0073] Furthermore, in step five, the forward path is represented as follows: , For the first Individual particles The values ​​at time points, weighted statistically for the particle set, are shown as follows: , To predict probabilities.

[0074] Furthermore, in step six, the specific scoring items are as follows:

[0075] Economic loss score Based on the amount of loss The amount of 10,000 yuan is determined and defined as follows:

[0076]

[0077] Downtime rating According to the work stoppage time The heavens are determined and defined as follows:

[0078]

[0079] Personnel Injury Score Determined based on the number of casualties, and defined as:

[0080]

[0081] Overall score .

[0082] Furthermore, in step six, the risk levels are divided into four levels from low to high, with level 1 corresponding to... Level 2 corresponds Level 3 corresponds Level 4 corresponds to .

[0083] Compared with existing technologies, the beneficial effects of this invention are as follows: Based on considering the temporal correlation of risk factors, parameter drift, and uncertainty propagation, this invention constructs a high-precision, real-time updated construction risk prediction model. This model is used for online fusion and temporal inference of multi-source heterogeneous data from construction sites under non-stationary, partially observable, and noise interference conditions, enabling real-time assessment and early warning of risk status. It has the following technical effects:

[0084] 1. Significantly improved risk prediction accuracy: By introducing a sliding window weighted counting recursion and an exponential forgetting update mechanism for conditional probability tables, this invention can reflect the latest distribution characteristics of risk factors at the construction site in real time, avoiding the prediction bias caused by outdated data in static probability tables in existing technologies. Calculation examples show that, under the same dataset, the prediction accuracy of this invention is about 10% higher than that of the traditional fixed-parameter DBN method.

[0085] 2. High prediction efficiency, meeting real-time monitoring requirements: This invention adopts a two-piece DBN structure and a particle filter sequential update algorithm, which decomposes the complex joint probability calculation into local conditional probability updates, significantly reducing the amount of computation. The time taken for a single prediction calculation is 40% to 50% shorter than that of the traditional Monte Carlo global inference method, and risk prediction and warning can be achieved within a minute-level construction monitoring cycle.

[0086] 3. Reduced reliance on manual labor and simplified operation: The parameter updates, particle weight normalization, and risk level determination of this invention are all completed automatically by algorithms, avoiding the process of extensive manual parameter tuning and trial and error in existing methods, reducing human error, and improving the operability and stability of field applications. Attached Figure Description

[0087] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0088] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0089] A dynamic Bayesian network construction risk prediction method based on sliding window adaptive learning and particle filtering is proposed, and its process combines... Figure 1 As shown, it includes the following steps:

[0090] Step 1: Input risk factor status;

[0091] ① Constructing the status of risk factors:

[0092] Risk factors are determined based on the actual construction scenario requirements and defined. The overall state vector of all risk factors at time 1 It is expressed as follows:

[0093]

[0094] In the formula, Indicates the first Each risk factor in The state at any given moment, , For the number of risk factors, For time slice index, , This represents the total number of time slices.

[0095] ② Construct a set of discrete states:

[0096] Establish a discrete state set for each risk factor. , All are taken from this set , means as follows:

[0097]

[0098] In the formula, express The A specific state, , This represents the total number of specific states.

[0099] Step 2: Construction of Dynamic Bayesian Network (DBN) Topology;

[0100] ① A two-piece DBN structure is adopted:

[0101] The two-piece DBN structure utilizes adjacent time slices The structure is used to model the time series process, and each moment is regarded as a time slice, which contains the state of the risk factor at that moment as a variable.

[0102] Within the same time slice, one-way edges represent the immediate effects between variables, but these edges cannot be connected end-to-end to form a loop (that is, starting from a variable along an edge, it cannot loop back to itself); if spanning multiple time slices, only settings from... point to Forward edges are used to represent lag effects. Skipping intermediate time slices and connecting edges from back to front are not allowed. The topology and parameters of the pair of adjacent time slices are treated as templates, and repeated copying and splicing along the time axis can form a complete time series network.

[0103] To solidify the direct impact relationships between various risk factors on the construction site into a calculable and verifiable structure, the following concepts are introduced:

[0104] Parent set: The set of risk factors that directly affect the current state of a certain target risk factor, determined by the topology network;

[0105] Parent node: It is the direct source of influence of the target risk factor state and the starting point of a unidirectional edge in the topology;

[0106] Parent configuration: A vector consisting of the states of each parent node at a given moment;

[0107] Child node: The state of the target risk factor modeled under a given parent configuration; the endpoint of a unidirectional edge in the topology.

[0108] Child value: The state of the child node at a certain moment.

[0109] In this topology, the parent nodes corresponding to child nodes (i.e., the target risk factor states) are divided into two categories: cross-slice parent nodes and same-slice parent nodes. A cross-slice parent node is the target risk factor state itself from the previous time slice, and its state is represented in the topology as follows: (Cross-slice parent configuration); The parent node within the same slice is the state of the risk factor that directly affects the state of the target risk factor within the current time slice, and its state is represented as follows: (Same as parent configuration).

[0110] ② Temporal decomposition of joint distribution:

[0111] because Representing the Each risk factor in The state at any given time is a random variable. It is a discrete distribution. Indicates the first The specific state of each risk factor It is a specific probability value.

[0112] To represent the temporal dependencies of the structure, the joint distribution expression for the entire time period is defined as follows:

[0113]

[0114] In the formula, Indicates the first to the last The complete state record of risk factors for each time slice, with the risk factor state vectors within each time slice arranged in chronological order. This represents the joint probability distribution of the state vectors of all risk factors over all time periods. As the initial joint distribution, the set of risk factor states is determined in The initial uncertainty at time is a distribution rather than a fixed probability value. Only when each risk factor state is replaced with a specific state can a specific probability be obtained. The transition distribution of adjacent time slices represents the temporal dependency where the current time step is determined solely by the previous time step.

[0115] ③ Determination of the initial joint distribution:

[0116] The joint distribution is decomposed according to a predetermined topological structure to obtain the initial joint distribution, as follows:

[0117]

[0118] Taking the complete record of the states of all risk factors within the system at a specific moment as a sample, each sample can be represented as a global state vector with a specific state. Let the historical sample set used for initialization be... Then we have:

[0119]

[0120] In the formula, For a given parent configuration Conditional distribution, for The Middle The state of each risk factor, in Timing calculations are more efficient than cross-slice parent configurations. In order to be in Always satisfied The sample count required for specific states, In order to be in Always satisfied The sample count required for specific states, This is a smoothing parameter used to avoid zero probability and small sample jitter.

[0121] Substituting the specific state yields the corresponding conditional probability:

[0122]

[0123] In the formula, for The specific set of states The specific state label for the sub-value represents a specific state that this count is targeting. Similarly, It is the specific status label configured in the same parent segment. It is the specific status label of the cross-chip parent configuration. In order to be in Always satisfied The sample count results, In order to be in Always satisfied The sample count results.

[0124] For different parent configurations and child values, the obtained conditional probabilities are filled into the table below with the parent configuration as the column data and the child value as the row data. The sum of the probabilities in each row is equal to 1, thus obtaining the initial conditional probability table (CPT).

[0125]

[0126] ④ Determination of the transfer distribution:

[0127] The joint distribution is decomposed according to a given topological structure to obtain the transition distribution, as shown below:

[0128]

[0129] in, For a given parent configuration The conditional distribution.

[0130] Step 3: CPT update based on sliding window;

[0131] ① Define a sliding window:

[0132] To limit the data range upon which CPT updates depend, the concept of a sliding window is introduced, which refers to a fixed-length interval on the timeline that changes with the current time. The data is shifted forward synchronously, and only samples falling within the interval are included in the statistics and updates of CPT; the rest of the data are no longer directly involved in the current estimate.

[0133] For time slice index The length is defined as And the step size is (Sky) sliding window for:

[0134]

[0135] The initial CPT cannot be directly used for counting recursion. To integrate it into the counting recursion, an equivalent counting baseline is defined:

[0136]

[0137] In the formula, This serves as the equivalent counting baseline and the initial row for the weighted counting recursion. The total weight of this row is determined by the historical sample set. The number of samples in the sample is determined by the sample size.

[0138] ② Weighted counting recursion:

[0139] To account for the non-stationarity of the data, a forgetting factor is first applied to the historical weighted counts. This reflects the attenuation, and by superimposing the counts of newly added samples within the sliding window, we obtain the weighted counting recursive expression:

[0140]

[0141] In the formula, In order to be in Always satisfied The sample count required for specific states, As of Always satisfied The weighted count required for a specific state, i.e., the historical weighted count.

[0142] ③Conditional probability estimation:

[0143] After obtaining the weighted counts, data smoothing is performed on each row to ensure numerical stability and avoid zero probability, thus adjusting the conditional distribution. Smoothing is performed as follows:

[0144]

[0145] In the formula, As of Always satisfied Weighted counts required for specific states, As of Always satisfied Weighted counts required for specific states.

[0146] Substituting each of the specific states yields the corresponding conditional probabilities:

[0147]

[0148] In the formula, As of Always satisfied The weighted count results, As of Always satisfied The weighted count results.

[0149] ④ Write the result back:

[0150] The updated Stored as the latest CPT for subsequent inference and prediction.

[0151] Step 4: Risk state recursive reasoning based on particle filtering;

[0152] Particle filtering is a state estimation algorithm used to recursively estimate the true state of a system when the system state cannot be directly observed and there is uncertainty.

[0153] ①State observation model:

[0154] At construction sites, the status of risk factors cannot be directly observed; data must be acquired using on-site sensing systems to infer their effects. To... A quantitative relationship is established between the risk factor state at any given time and the observation results collected by the sensing system. The likelihood observation model is defined as follows:

[0155]

[0156] In the formula, For the observation result vector, For the first The observation results for each channel are independent of each other. , The total number of observation channels indicates the number of channels the system can access. Different types of sensors are used to acquire data.

[0157] ②Particle prediction:

[0158] In a given particle After the state at time point, according to the transition distribution of two adjacent time slices, in topological order... Factor-wise sampling yields the following predicted sample for a given particle at this moment:

[0159]

[0160] In the formula, Indicates the first The particle corresponds to the first Each risk factor in The state at time step, i.e., the predicted sample obtained by sampling from the transition model. Indicates the first The particle corresponds to the first Each risk factor in The state at a given time, i.e., the prior sample. To show obedience distributed, The number of particles.

[0161] ③ Weight update:

[0162] After obtaining the predicted samples, weighting is required based on the likelihood observation model. The particle weights from the previous time step are multiplied by the likelihood calculated under the likelihood observation model at the current time step to obtain the unnormalized weights. Then, the sum of the unnormalized weights of all particles is used as a constant to normalize each weight, as shown below:

[0163]

[0164]

[0165] In the formula, for Particle weights at time step for Particle weights at time t, for Normalized weights of particles at time step This is the normalization constant.

[0166] ④ Degradation detection and resampling:

[0167] In estimating event probabilities, the fewer the number of particles, the greater the weight of the minority particles, and thus the larger the variance of the estimate. The effective sample size is defined as follows. It is expressed as follows:

[0168]

[0169] A small number of effective samples indicates that the weighted estimation becomes inaccurate. When the number of effective samples is too small and the particle weights are too concentrated, a resampling operation should be performed to prevent the estimation results from becoming biased and to maintain sample diversity. , If the resampling threshold is reached, then resampling is performed and the weights are reset to [value]. .

[0170] ⑤ Posterior approximation:

[0171] The true posterior distribution of a system typically lacks an analytical form and cannot be directly calculated. Through posterior approximation, the theoretical posterior distribution is approximated as a finite set of particles and their weights. This allows the complex probability distribution to be expressed using a computable sample set, thus enabling operable numerical inference, as shown below:

[0172]

[0173] In the formula, For the specific state of the target, It is an indicator function if and only if it satisfies The indicator function is set to 1 if the condition is met, and 0 otherwise.

[0174] Step 5: Risk prediction based on forward Monte Carlo propagation;

[0175] ①Particle forward roll:

[0176] For each particle, its forward path is obtained by step-by-step sampling from its current state according to the transition model, as shown below:

[0177]

[0178] In the formula, For the first Individual particles The value at time, The prediction step size represents the number of forward predictions.

[0179] ② Probability of future events:

[0180] In obtaining the particle After the forward prediction samples are obtained, the particle set is weighted and statistically analyzed by comparing the sample states with the target states, as shown below:

[0181]

[0182] In the formula, To predict the probability, if and only if the following condition is met: The indicator function is set to 1 if the condition is met, and 0 otherwise.

[0183] ③Result Output:

[0184] Get the Future The risk probability distribution of each step is used to obtain the probability result corresponding to the earliest out-of-limit step. This provides input for risk level assessment.

[0185] Step Six: Risk Level Assessment;

[0186] ① Loss severity score:

[0187] The severity of the loss is scored based on the economic damage, downtime, and personal injury caused by the risk, as detailed below:

[0188] Economic loss score Based on the amount of loss (Unit: 10,000 yuan) is determined and defined as follows:

[0189]

[0190] Downtime rating According to the work stoppage time (Unit: days) Determined, defined as:

[0191]

[0192] Personnel Injury Score Determined based on the number of casualties, and defined as:

[0193]

[0194] The scores from each item are combined into a comprehensive score using a fixed weighting. , represented as:

[0195]

[0196] ② Risk value calculation and classification:

[0197] Define risk value According to risk value The area is divided into four risk levels, from low to high, with level 1 corresponding to... Level 2 corresponds Level 3 corresponds Level 4 corresponds to .

[0198] The corresponding measures for each risk level are as follows:

[0199]

[0200] In summary, this invention, by collecting multi-source heterogeneous data from construction sites and combining dynamic Bayesian network modeling, particle filter sequential estimation, and sliding window adaptive learning, comprehensively considers the complexity of the environment, materials, processes, and organizational management during the construction phase, as well as practical factors such as non-stationarity, partial observability, and noise interference. It proposes a high-precision and efficient online intelligent prediction method for construction risk prediction and early warning. This method can achieve rapid model updates and uncertainty quantification under conditions of continuous data arrival, providing scientific support for on-site risk management, resource allocation, and emergency response. It effectively improves prediction accuracy and timeliness, reduces false alarms and false negatives, and thus ensures the structural safety and operational reliability of engineering construction.

[0201] Example

[0202] This embodiment focuses on the construction site of the Tongshunhe Bridge, and uses dynamic construction risk factor prediction to ensure safety and structural stability during the construction process.

[0203] Step 1: Input risk factors;

[0204] ① Constructing the status of risk factors:

[0205]

[0206] The number of risk factors is , Indicates personnel behavior, Indicates device status. Indicates the state of the environment. Indicates a stage in the process.

[0207] ② Construct a set of discrete states:

[0208] In this case, the specific status of each risk factor is as follows:

[0209]

[0210]

[0211]

[0212]

[0213] Step 2: Constructing the dynamic Bayesian network topology;

[0214] Construct a conditional probability table:

[0215] Child nodes: Device status at any time ;

[0216] Parent node: The device state at the previous moment. The current environmental state (equivalent to) This directly affects the value of the device status;

[0217] .

[0218] Historical frequency (from historical sample set data):

[0219] The counts for the device status values ​​of Normal / Warning / Fault are 12 / 6 / 2 respectively;

[0220] The counts for the device status values ​​of Normal / Warning / Fault are 6 / 8 / 6 respectively;

[0221] The device status values ​​are normal, warning, and fault, and the counts are 3, 5, and 12 respectively.

[0222] The normalized results are as follows, satisfying the condition that the sum of the row probabilities equals 1.000:

[0223]

[0224] Step 3: CPT update based on sliding window;

[0225] In this case study, the length of the sliding window is calculated. Set to 24 days, step size Forgetting factor (1 day) Set to 0.94, number of particles Set to 100, total weight Take 10, parent configuration is The sub-value is Substituting the result from step two, we obtain the equivalent count at the initial time:

[0226]

[0227] By performing a weighted counting recursion, the weighted effective count is obtained as follows:

[0228]

[0229] After estimating the conditional probability, the result is written back to obtain the updated CPT:

[0230]

[0231] Step 4: Risk state recursive reasoning based on particle filtering;

[0232] ① Channel likelihood (conditional independent product: look up the table for each particle and each channel and multiply):

[0233] Observation results:

[0234] Channel likelihood probability table:

[0235]

[0236] Select only the parent configuration that satisfies , The 10 particles are displayed:

[0237]

[0238] ② Weight update:

[0239] At the previous moment, all particles had equal weight. Calculate particle weights:

[0240]

[0241] Normalize the weights:

[0242]

[0243]

[0244] check: The calculation is correct.

[0245] ③ Effective particle count and resampling trigger determination:

[0246]

[0247]

[0248]

[0249] Resampling threshold Taking 0.5 and substituting it into the formula, we get: , Therefore, resampling is not triggered in this round.

[0250] ④ Posterior approximation:

[0251]

[0252] Step 5: Risk Prediction Based on Forward Monte Carlo Propagation

[0253] Target event: Device status is faulty ( The prediction step size is .

[0254] ① First step prediction:

[0255]

[0256]

[0257] ② Second step prediction:

[0258]

[0259]

[0260] ③ Earliest overstep and lead time:

[0261] Set management threshold ,

[0262] The earliest over-limit step ,

[0263] The probability result for the second prediction step is 0.2060, which exceeds the threshold (0.2060 ≥ 0.20), indicating the earliest step exceeding the threshold. .

[0264] Step Six: Risk Level Assessment

[0265] ① Loss severity score:

[0266] enter: Ten thousand yuan; Five people were injured, but none were disabled.

[0267] Conversion: , , ;

[0268] get: .

[0269] ② Risk value calculation and classification:

[0270] exist Place, .

[0271] The final risk level was determined to be Level 2; the response strategy was to strengthen prevention and incorporate it into the safety management plan.

[0272] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0273] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks, characterized in that: Includes the following steps: Step 1: Input risk factor status; ① Construct risk factor status: Determine and define risk factors based on the actual construction scenario requirements. The overall state vector of all risk factors at time 1 ; ② Constructing a discrete state set: Establish a discrete state set for each risk factor. The specific states of risk factors are all taken from the corresponding discrete state set; Step 2: Constructing the dynamic Bayesian network topology; ① A two-slice DBN structure is adopted: adjacent time slices are used. The structure is used to model and represent the time series process, treating each moment as a time slice. Within a time slice, the state of risk factors at that moment is considered a variable. Within the same time slice, acyclic unidirectional edges represent the immediate effects between variables. Across time slices, only the state from... point to The forward edges divide the parent nodes corresponding to the target risk factor into cross-slice parent nodes and same-slice parent nodes. The cross-slice parent node is the state of the target risk factor itself in the previous time slice, and the same-slice parent node is the state of the risk factor that has a direct impact on the state of the target risk factor in the current time slice. ② Temporal decomposition of joint distribution: Define a joint distribution for the entire time period and decompose it into an initial joint distribution. With the transfer distribution The product; ③ Determination of the initial joint distribution: Calculate the conditional distribution of risk factor states under a given parent configuration based on the historical sample set, and construct the initial CPT; ④ Determining the transition distribution: Decompose the joint distribution according to the topological structure to obtain the distribution under a given parent configuration. Conditional distribution of risk factor states at any given time; Step 3: CPT update based on sliding window; ① Define a sliding window: Set a fixed-length window that shifts over time, and only samples within the window participate in CPT updates; Define an equivalent count baseline to connect the initial CPT with the weighted count recursion; ②Weighted count recursion: Apply a forgetting factor to the historical weighted count to decay it, and then add the count of newly added samples in the sliding window to obtain the current weighted count; ③Conditional probability estimation: Smooth the weighted count, calculate the conditional probability under a given parent configuration, and write the result back to update the CPT; Step 4: Risk state recursive reasoning based on particle filtering; ①State observation model: Will A quantitative relationship is established between the risk factor state at any given time and the observation results collected by the sensing system, and a likelihood observation model is constructed. ②Particle prediction: Based on the transition distribution between two adjacent time slices, from The particle state sampling at time t gives a given particle Predicted samples at each time point; ③Weight update: Weights are calculated based on the likelihood observation model to determine the unnormalized weights of the particles, and then each weight is normalized. ④ Degradation detection and resampling: Calculate the number of valid samples. When the number of valid samples is less than the resampling threshold, perform resampling and reset the weights. ⑤ Posterior approximation: The posterior distribution is approximated as a finite number of particles and their weights, thereby expressing the probability distribution through a computable set of samples; Step 5: Risk prediction based on forward Monte Carlo propagation; ① Particle forward roll: For each particle, starting from the current state, the future is obtained by progressively sampling according to the transition model. The forward path of time, To predict the step size; ② Probability of future events: A weighted statistical set of particles yields the probability of future events. The predicted probability of the state of the risk factor; ③ Output: Obtain the probability result corresponding to the earliest out-of-limit step. ; Step Six: Risk Level Assessment; ① Severity of Loss Score: Based on economic loss, downtime, and personal injury, scores are calculated for each item and weighted to obtain a comprehensive score. ; ② Risk Value Calculation and Classification: Defining Risk Value ,according to The value determines the risk level.

2. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 1, characterized in that: In step one, the overall state vector With discrete state set They are represented as follows: In the formula, Indicates the first Each risk factor in The state at any given moment, , For time slice index, , express The A specific state, .

3. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 2, characterized in that: In step two, the process of determining the initial joint distribution and transition distribution through joint distribution time series decomposition is as follows: In the topology, the state of a cross-slice parent node is represented as: The state of the parent node in the same piece is represented as ; The joint distribution expression for the entire time period is defined as follows: In the formula, Indicates the first to the last The complete state record of risk factors for each time slice, with the risk factor state vectors within each time slice arranged in chronological order. This represents the joint probability distribution of the state vectors of all risk factors over all time periods. The initial joint distribution is represented as follows: Let the historical sample set used for initialization be . Then we have: In the formula, for The Middle The state of each risk factor, in Timing calculations are more efficient than cross-slice parent configurations. In order to be in The number of samples that meet the specific state requirements at any given time. The smoothing parameter is then substituted into the specific state to obtain the corresponding conditional probability. For different parent configurations and child values, the obtained conditional probabilities are filled in with the parent configuration as column data and the child value as row data. The sum of the probabilities in each row is equal to 1, thus obtaining the initial CPT. The transition distribution is represented as follows: in, For a given parent configuration The conditional distribution.

4. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 3, characterized in that: In step three, the equivalent counting baseline is represented as follows: In the formula, As the initial row for the weighted counting recursion, This represents the total weight of the row; The weighted counting recursive representation is as follows: In the formula, In order to be in The number of samples that meet the specific state requirements at any given time. As of Historical weighted count of moments Forgetting factor; Conditional distribution Smoothing is performed as follows: In the formula, As of The weighted count that satisfies the specific state requirements at each moment is then substituted into the specific state to obtain the corresponding conditional probability, and the updated count is then used to calculate the conditional probability. Write the results back to update CPT.

5. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 4, characterized in that: In step four, the state observation model and particle prediction specifically refer to: The likelihood observation model is defined as follows: In the formula, For the observation result vector, For the first Observation results of each channel, ; In a given particle After the state at time point, according to the transition distribution of two adjacent time slices, in topological order... Factor-wise sampling yields the following predicted sample for a given particle at this moment: In the formula, Indicates the first The particle corresponds to the first Each risk factor in The state at time t is the prediction sample. Indicates the first The particle corresponds to the first Each risk factor in The state at time t is a priori sample. The number of particles.

6. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 5, characterized in that: In step four, the degradation discrimination, resampling, and posterior approximation are specifically as follows: Define the number of valid samples It is expressed as follows: In the formula, for The normalized weights of particles at time t, when , If the resampling threshold is reached, then resampling is performed and the weights are reset to [value]. ; The probability distribution is expressed using a computable set of samples as: In the formula, For the specific state of the target, It is an indicator function if and only if it satisfies The indicator function is set to 1 if the condition is met, and 0 otherwise.

7. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 6, characterized in that: In step five, the forward path is represented as follows: , For the first Individual particles The values ​​at time points, weighted statistically for the particle set, are shown as follows: , To predict probabilities.

8. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 1, characterized in that: In step six, the specific scores are as follows: Economic loss score Based on the amount of loss The amount of 10,000 yuan is determined and defined as follows: Downtime rating According to the work stoppage time The heavens are determined and defined as: Personnel Injury Score Determined based on the number of casualties, and defined as: Overall score .

9. The construction risk prediction method based on sliding window adaptive learning and particle filtering using dynamic Bayesian networks according to claim 8, characterized in that: In step six, the risk levels are divided into four levels from low to high, with level 1 corresponding to... Level 2 corresponds Level 3 corresponds Level 4 corresponds to .

Images