Safety accident prediction system and method based on construction site data analysis

Abstract

The invention discloses a safety accident prediction system and method based on construction site data analysis. The method comprises the following steps: screening basic sample data of typical historical accident cases to obtain target sample data; extracting element features of various three aspects (including an initial disaster-inducing factor, an accident disaster-bearing body damage state and an accident disaster-bearing body emergency response equipment operation state) from the target sample data, and constructing a hidden Markov network model based on the element features of the three aspects; inputting data to be detected, and outputting feedback result information through the Markov network model; according to the method, the state of the system in a period of time in the future can be predicted, so that the risk is accurately predicted, and corresponding measures are taken for intervention.

Description

Technical Field

[0001] The present invention relates to the field of construction accident prediction, and in particular to a safety accident prediction system and method based on construction site data analysis. Background Art

[0002] Construction safety is of paramount importance to the normal construction of buildings. Among many types of construction accidents, collapse accidents are the most harmful. Modern technical monitoring methods have played an important role in the prevention and monitoring of collapse accidents. The following common modern technical monitoring methods are currently used: Single-use support structure monitor: used to monitor the stress and deformation of the support structure at the construction site in real time. It can detect problems such as instability or overload of the support structure in advance, and warn of potential collapse risks.

[0003] Single-use soil and structural stability monitoring system: By monitoring soil pressure and structural stability, possible landslides, subsidence and other problems can be discovered in a timely manner to reduce the risk of collapse.

[0004] The application of these technical monitoring methods helps to improve the prevention and early warning capabilities of construction collapse accidents, thereby reducing the probability of accidents and ensuring the safety of workers' lives.

[0005] However, research has found that in reality the information collected by these monitoring methods is still limited, and some intelligent emergency disaster reduction equipment (such as emergency auxiliary support systems) can significantly reduce the risk of accidents after being activated.

[0006] At this time, traditional monitoring methods only focus on monitoring the initial factors that cause disasters, and do not fully consider the impact of disaster reduction intervention measures on accidents.

[0007] Due to the lack of systematic analysis of the effectiveness of disaster reduction equipment operation (i.e. the operating status of emergency response equipment of the accident-affected body), this may lead to inaccurate assessment of the severity of the disaster; at the same time, further research found that citing the Markov model is an important solution, but based on practical problems such as the complexity of data collection at the construction site, inconsistency of monitoring methods, and inaccurate probability values ​​of certain states in the state transition probability matrix, the use of the traditional citing Markov model for construction accident monitoring is prone to cause long-term output of overestimated accident risk levels, which in turn frequently triggers accident warnings (sometimes long-term underestimation of accident warning risks), resulting in reduced prediction accuracy. Summary of the invention

[0008] The purpose of the present invention is to provide a safety accident prediction system and method based on construction site data analysis, which solves the above-mentioned technical problems pointed out in the prior art.

[0009] The present invention provides a safety accident prediction system based on construction site data analysis, including a collection module, a screening module, a construction module, and a processing output module: The acquisition module is used to collect basic sample data of typical historical accident cases in advance; The screening module is used to screen the basic sample data of the typical historical accident cases to obtain target sample data; The construction module is used to extract the element characteristics of the target sample data including the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y), and then construct a hidden Markov network model based on the element characteristics of the three aspects, define the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and add connection marks between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; The processing output module is used to input the data to be detected and output feedback result information through the above Markov network model; The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability.

[0010] Preferably, as an implementable solution, the screening module is also used to perform initial screening on the construction safety accident report data contained in the basic sample data of typical historical accident cases; the collected basic sample data containing the construction safety accident report data is screened again to obtain the basic sample data containing the influencing factors as the target sample data.

[0011] Preferably, as an implementable solution, the construction module is also specifically used to obtain target sample data, and extract the initial disaster factor (C), the damage state of the accident disaster-bearing body (S), and the operating state of the emergency response equipment of the accident disaster-bearing body (Y) based on the target sample data; The construction module is also specifically used to extract the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y) based on the target sample data to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial hazard factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y); Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes a serious accident state (YW), a high-risk accident state (GW), and a potential dangerous accident state (QW); The construction module is also specifically used to obtain a hidden Markov network model λ through training based on the target sample data as a data set.

[0012] Preferably, as an implementable solution; the processing output module is also used to input the data to be detected, and obtain the observation sequence of the target time length T' of the data to be detected; confirm that the observation sequence of the target time length T' is the target observation sequence; wherein, the observation sequence of the target time length T' is O={o1,o2,...,oT'}; use the dynamic forward planning method to calculate the probability P(O|λ) of O appearing in the target observation sequence under the hidden Markov network model λ based on the target observation sequence.

[0013] Accordingly, the present invention provides a safety accident prediction method based on construction site data analysis, comprising the following steps: S10. Collect and obtain basic sample data of typical historical accident cases in advance; S20, screening the basic sample data of the typical historical accident cases to obtain target sample data; S30, then extracting the element features of three aspects (including the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y)) from the target sample data, and then constructing a hidden Markov network model based on the element features of the three aspects, defining the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and adding connection marks between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model; The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability.

[0014] Preferably, as an implementable solution, during the execution of S20, the basic sample data of the typical historical accident cases are screened to obtain target sample data, specifically including: S21. Initially screen the construction safety accident report data contained in the basic sample data of typical historical accident cases; typical historical accident cases are generally basic sample data containing construction safety accident reports, and also include basic sample data without construction safety accident reports, so preliminary screening is required; S22, re-screening the collected basic sample data containing the construction safety accident report data, filtering out the basic sample data lacking influencing factors, and obtaining the basic sample data containing influencing factors as the target sample data.

[0015] Preferably, as an implementable solution, during the execution of S30, a hidden Markov network model is constructed based on the element characteristics of the three aspects, and the accident state triggered by the influencing factors is defined as a network node on the hidden Markov network model, which specifically includes the following operations; S31. Obtain target sample data, and extract the initial disaster-causing factor (C), the damage status of the accident-affected body (S), and the operating status of the emergency response equipment of the accident-affected body (Y) based on the target sample data; S32. Based on the target sample data, the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y) are extracted to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial hazard factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y); Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes a serious accident state (YW), a high-risk accident state (GW), and a potential dangerous accident state (QW); S33, training based on the target sample data as a data set to obtain a hidden Markov network model λ; wherein λ=(A, B, π): Among them, A is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

[0016] Preferably, as an implementable scheme; S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model, specifically includes: S41, inputting the data to be detected, and obtaining an observation sequence of the target time length T' of the data to be detected; confirming that the observation sequence of the target time length T' is a target observation sequence; wherein the observation sequence of the target time length T' is O={o1, o2, ..., oT'}; S42. Using a dynamic forward planning method, based on the target observation sequence, calculate the probability P(O|λ) of O appearing in the target observation sequence under the hidden Markov network model λ.

[0017] Preferably, as an implementable scheme; S421, obtain the hidden Markov network model λ, define the partial observation sequence at time t as o1, o2…ot and the state as q i The probability of is called the forward probability; The forward probability is calculated as follows: α t (i)=P(o1,o2…o t ,i t =q i |λ), the forward probability α can be recursively calculated t (i) and the observed sequence probability P(O|λ); S422. At initialization, the probability of setting the first moment state to the i-th state is π i , the observation probability of obtaining the observation value o1 in the i-th state is b io1 , the forward probability of the first moment state being the i-th state is: ; S423, starting from time t=1, recursively calculate the forward probability at time t=2, and directly calculate multiple times to obtain the forward probability at the next time t+1: the above multiple calculations refer to starting from t=1, t=2... until the next time t+1 is obtained by multiple calculations, and finally calculating to time T'; The forward probability α of the next time t+1 t+1 (i) = the sum of the forward probability at time t over all states × the probability of the state at time t being observed; ; in, is the forward probability of state j at time t; is the transition probability from state j to state i; When it is the i-th state, we get o t+1 the probability of observing this observed value; S423, output the probability of the final observation sequence: The calculation formula for the probability of the final observation sequence is: ; The above calculation result is output as the probability P(O|λ) of the occurrence of O in the target observation sequence.

[0018] It is characterized in that, during the execution of S30, it also includes: S31', calculating the comprehensive bias rate of the target sample data, screening and obtaining the adjusted target sample data based on the comprehensive bias rate of the target sample data and the standard comprehensive bias rate threshold, and extracting the initial disaster-causing factors, the damage status of the accident-affected body, and the operating status of the emergency response equipment of the accident-affected body based on the adjusted target sample data; S32', based on the adjusted target sample data, extract the initial disaster-causing factors, the damage status of the accident-affected body, and the element characteristics of the operating status of the emergency response equipment of the accident-affected body to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial disaster factor, the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body; Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes serious accident state, high-risk accident state, and potential dangerous accident state; S33', implement the judgment of the first relay correction condition: judge whether there is a single indicator of the current sample distribution bias rate in the current comprehensive bias rate that is higher than the standard sample distribution bias rate threshold, if higher, detect the sample distribution bias rate range where the current sample distribution bias rate is located; call the mapping relationship table of the expert weighted weight B1j with preset historical state transition frequency and the sample distribution bias rate range; search and determine the expert weighted weight B1j of the historical state transition frequency based on the mapping relationship table; Then, based on the dynamic weighting of the state transition probability, the state transition probability is adjusted to: A'=A (A is the state transition probability matrix) × the expert weighted weight B1j of the historical state transition frequency: Implement the judgment of the second relay correction condition: further judge whether there is a situation in the current comprehensive bias rate that the current single indicator based on the feature importance bias rate is higher than the standard feature importance bias rate threshold, if so; return to step S31' to re-screen the adjusted target sample data; train based on the adjusted target sample data as a data set to obtain a hidden Markov network model λ'; Where λ'=(A,B,π): Among them, A' is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

[0019] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages: It can be seen from the analysis of the above-mentioned safety accident prediction system and method based on construction site data analysis provided by the present invention that, in specific application, basic sample data of typical historical accident cases are first collected and obtained; the basic sample data of the typical historical accident cases are screened to obtain target sample data; Then, the target sample data is extracted from the element features of three aspects (including the initial disaster-causing factor, the damage state of the accident-affected body, and the operating state of the emergency response equipment of the accident-affected body), and then a hidden Markov network model is constructed based on the element features of the three aspects, and the accident state triggered by the influencing factor is defined as a network node on the hidden Markov network model; and connection marks are added between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; Input the data to be detected, and output feedback result information through the above Markov network model; the feedback result information includes the accident state triggered by the influencing factors at the current time node and the corresponding evaluation probability: In specific implementation, the above technical solution is mainly based on the hidden Markov network model. By collecting historical accident case data, screening target sample data, extracting various element features, constructing a hidden Markov network model, and outputting feedback result information, it has at least the following technical effects: First, the solution can effectively analyze accidents, help us understand the causes and influencing factors of accidents, and predict the possible development trend of accidents. Second, it achieves more accurate prediction of risks, that is, by constructing a hidden Markov network model, the state of the system in the future can be predicted, so as to accurately predict risks and take corresponding measures to intervene. Since the solution can output various feedback result information, by analyzing the characteristics of elements such as the operating status of the emergency response equipment of the accident-bearing body, the efficiency of emergency response is improved and the decision-making level is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 An overall architecture diagram of a safety accident prediction system based on construction site data analysis provided in the first embodiment of the present invention; Figure 2 A schematic diagram of the operation flow of a safety accident prediction method based on construction site data analysis provided in the second embodiment of the present invention; Figure 3This is a schematic diagram of data demonstration of a specific embodiment of a safety accident prediction method based on construction site data analysis provided in Example 2 of the present invention.

[0021] Figure numerals: main controller 10, cloud server 20, information relay transmission system 30, accident scene collection system 40; collection module 11, screening module 12, construction module 13, processing output module 14. DETAILED DESCRIPTION

[0022] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0023] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0024] The present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings.

[0025] Embodiment 1 like Figure 1 As shown, the present invention proposes a safety accident prediction system based on construction site data analysis, including a main controller 10, a cloud server 20, an information transfer transmission system 30 and an accident site acquisition system 40; wherein the main controller 10 includes an acquisition module 11, a screening module 12, a construction module 13, and a processing output module 14: The acquisition module 11 is used to collect basic sample data of typical historical accident cases in advance; The screening module 12 is used to screen the basic sample data of the typical historical accident cases to obtain target sample data; The construction module 13 is used to extract the element characteristics of the target sample data including the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y), and then construct a hidden Markov network model based on the element characteristics of the three aspects, define the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and add connection marks between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; The processing output module 14 is used to input the data to be detected and output feedback result information through the above Markov network model.

[0026] The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability: The feedback result information output by the above hidden Markov network model may be the probability of the system being in a certain state or a certain event occurring at a certain time point or time period.

[0027] In the model, each node represents a possible state or event. For example, in accident analysis, a node can represent the possibility of an accident, the extent of equipment damage, the operation of emergency response equipment, and so on.

[0028] Through model analysis, we can calculate the probability that the system may be in different states in the future under the current state.

[0029] These probabilities can help us predict the possibility and development trend of accidents, so that we can take appropriate measures to avoid accidents or reduce the impact of accidents.

[0030] The hidden Markov network model adopted in the embodiment of the present invention can also output other relevant feedback result information, such as the transition probability of a certain state, the prediction of time series, etc.

[0031] This information is very helpful for studying the evolution laws of the system and formulating corresponding control strategies.

[0032] The screening module is also used to perform initial screening on the construction safety accident report data contained in the basic sample data of typical historical accident cases; wherein, the typical historical accident cases are generally basic sample data containing construction safety accident reports, and also include basic sample data not containing construction safety accident reports, so preliminary screening is required; the collected basic sample data containing construction safety accident report data are screened again, the basic sample data lacking influencing factors are filtered out, and the basic sample data containing influencing factors are obtained as the target sample data.

[0033] Preferably, as an implementable solution, the construction module is also specifically used to obtain target sample data, and extract the initial disaster factor (C), the damage state of the accident disaster-bearing body (S), and the operating state of the emergency response equipment of the accident disaster-bearing body (Y) based on the target sample data; The construction module is also specifically used to extract the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y) based on the target sample data to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial hazard factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y); Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes a serious accident state (YW), a high-risk accident state (GW), and a potential dangerous accident state (QW); The construction module is further specifically used to obtain a hidden Markov network model λ by training based on the target sample data as a data set; It should be noted that collapse accidents will continue to develop with the passage of time, and their development is greatly affected by time, especially the situation at the previous moment, and changes with the continuous collapse and the passage of time. Therefore, the fragments of the accident state triggered by each influencing factor are placed on the timeline to determine the evolution law of the accident state triggered by the influencing factors.

[0034] Preferably, as an implementable solution; the processing output module is further used to input the data to be detected and obtain the observation sequence of the target time length T' of the data to be detected; confirm that the observation sequence of the target time length T' is the target observation sequence; wherein the observation sequence of the target time length T' is O={o1,o2,...,oT'}; using a dynamic forward planning method, based on the target observation sequence, calculate the probability P(O|λ) of the occurrence of O in the target observation sequence under the hidden Markov network model λ; The above technical solution is mainly based on the hidden Markov network model. By collecting historical accident case data, screening target sample data, extracting various factor features, a hidden Markov network model is constructed and feedback result information is output.

[0035] The following are the main technical effects of this technical solution: First, this solution can effectively analyze accidents, help us understand the causes and influencing factors of accidents, and predict the possible development trends of accidents.

[0036] Second, it achieves more accurate prediction of risks. That is, by constructing a hidden Markov network model, the state of the system in the future can be predicted, thereby accurately predicting risks and taking appropriate measures to intervene.

[0037] Since the scheme can output various feedback result information, the efficiency of emergency response and the level of decision-making can be improved by analyzing the characteristics of factors such as the operating status of the emergency response equipment of the accident-affected body. In summary, the technical scheme is fast, accurate and reliable, and can effectively analyze accidents, accurately predict risks, improve decision-making levels, and provide strong technical support for accident management and emergency response.

[0038] Embodiment 2 like Figure 2 As shown, accordingly, the present invention also proposes a safety accident prediction method based on construction site data analysis, comprising the following operating steps: S10. Collect and obtain basic sample data of typical historical accident cases in advance; S20, screening the basic sample data of the typical historical accident cases to obtain target sample data; S30, then extracting the element features of three aspects (including the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y)) from the target sample data, and then constructing a hidden Markov network model based on the element features of the three aspects, defining the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and adding connection marks between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model; The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability.

[0039] Hidden Markov random fields are often used to analyze random processes. Each node in the random field represents a state that the random system may be in at a certain moment. The random variable corresponding to the node represents the probability that the system is in that state at a certain moment. The probability of the system being in each state at the initial moment, each node calculates the probability of the system being in that state at the next moment based on the prior probability of its adjacent nodes. With each iteration, the probability distribution of the observed variables will generate a time series that can describe the random process, which we call the observation sequence.

[0040] The Hidden Markov Model (HMM) adds a set of observed random variables, and a hidden Markov chain will produce an unobservable random process. The prior results of each node in the chain will produce an observable probability distribution on the observed variables. The embodiments of the present application can analyze the random process on the hidden Markov chain with the help of observed variables.

[0041] Prioritize as an implementable solution; during the execution of S20, the basic sample data of the typical historical accident cases are screened to obtain target sample data, specifically including: S21. Initially screen the construction safety accident report data contained in the basic sample data of typical historical accident cases; typical historical accident cases are generally basic sample data containing construction safety accident reports, and also include basic sample data without construction safety accident reports, so preliminary screening is required; S22, re-screening the collected basic sample data containing the construction safety accident report data, filtering out the basic sample data lacking influencing factors, and obtaining the basic sample data containing influencing factors as the target sample data.

[0042] It should be noted that regarding the acquisition of data sources, collapse accident cases are collected from the annual construction safety accident reports compiled by the China Construction Industry Association, the Ministry of Emergency Management, the China Construction Safety Standardization Management Network, or the "Annual Report on Production Safety Accident Statistics", which are the basic original data of the main typical historical accident cases.

[0043] The above-mentioned typical historical accident cases generally record information such as the time of the accident, the level of the accident, the course of the accident, the operating status of the emergency response equipment of the accident-affected body (i.e. the operating status of the emergency auxiliary support system), and the damage status of the accident-affected body.

[0044] Among them, the damage state of the accident disaster-bearing body, wherein the accident disaster-bearing body can be a support frame, floor slab or surrounding soil during the construction process, etc.; the damage state of the accident disaster-bearing body is illustrated by way of example, it can be said to be the damage state of the support frame, specifically, it can be, for example, the fracture, deformation or displacement level of the support frame; however, the corresponding operation state of the emergency response equipment of the accident disaster-bearing body is that the support frame itself will have an active response support equipment startup processing, at this time, the operation state of the emergency response equipment of the accident disaster-bearing body can be the support force level index after a certain active response support equipment is started, or other aspects reflect the index of the operation of the emergency response equipment of the accident disaster-bearing body.

[0045] That is, filter out the basic sample data lacking influencing factors, and obtain the construction safety accident cases (basic sample data) containing influencing factors as the target sample data (that is, the basic sample data lacking influencing factors refers to accident cases without clear initial disaster factors (C), and some construction sites do not have the intelligent emergency response equipment in the background technology installed, so it is impossible to obtain the operating status of the emergency response equipment of the accident-bearing body (Y); these construction safety accident cases need to be filtered); For the collapse accident cases collected in the construction safety accident reports of typical historical accident cases, the TF-IDF algorithm (based on Python's jieba word segmentation package) can be used in the above process to classify and summarize the influencing factors of the collapse accidents, and structure the accident case text; the embodiment of the present application preferably selects the above-mentioned influencing factors as three aspects, namely, the initial disaster-causing factor (C), the damage state of the accident disaster-bearing body (S), and the operating state of the emergency response equipment of the accident disaster-bearing body (Y), which will not be repeated here.

[0046] Prioritize as an implementable solution; during the execution of S30, a hidden Markov network model is constructed based on the three aspects of the element characteristics, and the accident state triggered by the influencing factors is defined as a network node on the hidden Markov network model, which specifically includes the following operations; S31. Obtain target sample data, and extract the initial disaster-causing factor (C), the damage status of the accident-affected body (S), and the operating status of the emergency response equipment of the accident-affected body (Y) based on the target sample data; S32. Based on the target sample data, the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y) are extracted to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to be {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including initial hazard factor (C), damage state of accident hazard-bearing body (S), and operation state of emergency response equipment of accident hazard-bearing body (Y); for example, use the monitoring system to monitor various hazard factors, damage state of accident hazard-bearing body and operation state of emergency response equipment at the construction site.

[0047] In order to assess the risk of collapse accidents, the observed state sequence is analyzed based on the hidden Markov model (HMM), and the observed state set is set as follows: Observed state set V: v1: initial disaster factor state (such as soil pressure, supporting structure force, etc.).

[0048] v2: Damage status of the accident-bearing body (such as structural cracks, deformation, etc.). v3: Operation status of emergency response equipment (such as alarm equipment working status, support equipment working status).

[0049] For example, set the number of states M: M = 3, corresponding to v1, v2, v3. The data corresponding to the observation state are as follows: The key observation state data of the construction site at 5 time steps (assuming the unit time step is 1 hour), see Figure 3 ; Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes a serious accident state (YW), a high-risk accident state (GW), and a potential dangerous accident state (QW); S33, training based on the target sample data as a data set to obtain a hidden Markov network model λ; wherein λ=(A, B, π): Among them, A is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

[0050] In addition, for example, this case is marked as case 100: Analyzing the above case 100, the above influencing factors trigger three accident states: serious accident state, high-risk accident state, and potential dangerous accident state.

[0051] If a target sample data is obtained and it is found that there is an influencing factor that will directly lead to the occurrence of an accident, but the accident state triggered by this influencing factor has only three forms: serious accident state (YW), high-risk accident state (GW), and potential dangerous accident state (QW).

[0052] Analysis of the target sample data shows that the accident state triggered by the influencing factor is controlled by three aspects of observation states; the three aspects of observation states mentioned above are the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y); it is concluded that the accident state triggered by the influencing factor will only be affected by the initial disaster factor (C), the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body (Y). It is assumed that the observation state set V is {v1, v2, V3} at this time, and it can also be assumed that the observation state set V is changed to {C1, S1, Y1}, then the number of observation states M at this time is 3; Further analysis of the target sample data revealed that the accident state triggered by the influencing factor is controlled by three aspects of observation factors; however, after being affected by the above observation factors, the accident state triggered by the influencing factor has only three forms: for example: serious accident state (YW), high-risk accident state (GW), and potential dangerous accident state (QW). Therefore, the state set Q={q1,q2,q3} of the accident state triggered by the influencing factor can also be assumed to be the state set Q={serious accident state (YW), high-risk accident state (GW), potential dangerous accident state (QW)}, and the number of states of the accident state triggered by the influencing factor is N=3; Based on the above case 100, the above target sample data of the accident state triggered by the influencing factors of the case 100 is obtained, and its state transition probability matrix is ​​known to be: ; The observation probability matrix is: ; The initial state probability matrix is: π = (0.2, 0.4, 0.4).

[0053] That is, the above process is calculated through target sample data, so multiple target sample data are used as data sets for training to obtain the hidden Markov network model λ; In addition, some other information can be obtained through the above target sample data, that is, I is a state sequence represented by a time length of T, and O is the corresponding observation sequence; I={i1,i2,...,iT}; O={o1,o2,...,on}; Then the hidden Markov network model λ performs subsequent prediction of the data to be detected. Assuming that the observed state sequence of the observed state is {C1, Y1, S1, Y2, S2, Y2} observed in one day, the probability of such an observed state sequence being {C1, Y1, S1, Y2, S2, Y2} is calculated in this embodiment; in this way, the hidden Markov network model λ can predict the calculated value of the probability of the observed state sequence being {C1, Y1, S1, Y2, S2, Y2} under the combined effect of the damage state of the accident-affected body and the offset of the operation influence of the emergency response equipment of the accident-affected body. For details, please refer to the subsequent operations.

[0054] Preferably, as an implementable scheme; S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model, specifically includes: S41, inputting the data to be detected, and obtaining an observation sequence of the target time length T' of the data to be detected; confirming that the observation sequence of the target time length T' is a target observation sequence; wherein the observation sequence of the target time length T' is O={o1, o2, ..., oT'}; S42. Using a dynamic forward planning method, based on the target observation sequence, calculate the probability P(O|λ) of O appearing in the target observation sequence under the hidden Markov network model λ.

[0055] The specific example 100 of this embodiment uses three elements to describe an HMM model (Hidden Markov Network Model) with N states and M observations: initial state probability matrix: that is, describing an N-dimensional vector, in which a vector value represents the probability that the random process is in the corresponding state at the initial moment; state transition probability matrix A: N × N matrix, in which a[i,j] represents the probability of transitioning from state i to j; observation probability matrix B: N×M matrix, in which b[i, k] represents the probability of generating observation value k when in state j; The above hidden Markov network model is expressed as: λ=(A,B,π); The state set Q={q1,q2,...,qn} of the accident state triggered by the above-mentioned influencing factors; The above observation state set V is V={v1,v2,...,vm}; A is the state transition probability matrix: ; where a ij =P(i t+1 =q j |i t =q i ), where a ij is in state q at time t i Under the condition that time t+1 is transferred to state q j probability; B is the observation probability matrix: Among them, b ik =P(o t =v k | i t =q i ) , b ik is in state q at time t i Generate observation v under the condition k probability.

[0056] π is the initial state probability vector: π=(π i ); Among them, π i =P(i1=q i ), π i At time t=1, it is in state q i probability.

[0057] Preferably, as an implementable method; S42, using a dynamic forward planning method, based on the target observation sequence, calculate the probability P(O|λ) of O appearing in the target observation sequence (recorded as the target observation sequence O={o1,o2,...,oT'}) under the hidden Markov network model λ, specifically including the following steps: S421, obtain the hidden Markov network model λ, define the partial observation sequence at time t as o1, o2…ot and the state as q i The probability of is called the forward probability; The forward probability is calculated as follows: α t (i)=P(o1,o2…o t ,i t =q i |λ), recursively calculate the forward probability α t (i) and the observed sequence probability P(O|λ); S422. At initialization, the probability of setting the first moment state to the i-th state is π i , the observation probability of obtaining the observation value o1 in the i-th state is b io1 , the forward probability of the first moment state being the i-th state is: ; S423, starting from time t=1, recursively calculate the forward probability at time t=2, and directly calculate multiple times to obtain the forward probability at the next time t+1: the above multiple calculations refer to starting from t=1, t=2... until the next time t+1 is obtained by multiple calculations, and finally calculating to time T'; The forward probability α of the next time t+1 t+1 (i) = the sum of the forward probability at time t over all states × the probability of the state at time t being observed; ; in, is the forward probability of state j at time t; is the transition probability from state j to state i; When it is the i-th state, we get o t+1 the probability of observing this observed value; S423, output the probability of the final observation sequence: The calculation formula for the probability of the final observation sequence is: ; In fact: P(O|λ)= α T’ (1) +α T’ (2) + ...+ α T’ (n); The above calculation result is output as the probability P(O|λ) of the occurrence of O in the target observation sequence.

[0058] Embodiment 3 Further research has found that the technical solution implemented in the above-mentioned embodiment 2 is still unable to accurately predict the probability of influencing factors triggering an accident state. The reason for this is that there are problems with the target sample data.

[0059] Among them, the target sample data involves multiple characteristics such as the initial disaster factor (C), the damage status of the accident disaster-bearing body (S), the operating status of the emergency response equipment of the accident disaster-bearing body (Y), etc. These characteristics are characteristic data obtained by different monitoring methods. For example, the support structure monitor is used to monitor the deformation of the support structure (that is, monitor the damage status of the accident disaster-bearing body).

[0060] The pressure monitoring system is used to monitor the structural pressure data of the accident-bearing body (that is, it also monitors the damage status of the accident-bearing body).

[0061] Therefore, this researcher believes that the bias rate of the target sample data, that is, the target sample data may be biased (such as uneven distribution of data from different monitoring methods).

[0062] To this end, the embodiment of the present application introduces the calculation of the bias assessment rate BiasRate, which is used to measure the contribution of the monitoring means to the target observation sequence and adjust the corresponding observation probability.

[0063] The bias rate is not only based on the uneven distribution of sample data, but also combined with the importance of features and the time correlation of the observation sequence, which can also cause negative effects.

[0064] For this purpose, the embodiment of the present application introduces a comprehensive bias rate hat(BiasRate)_i , which involves a sample distribution bias rate, a feature importance bias rate, and a time correlation bias rate.

[0065] In order to consider the different contributions of multiple data to accident risk assessment, a dynamic bias rate estimation is introduced. The bias rate is not only based on the number of samples, but also dynamically combines the importance of features and the time correlation of the observation sequence, and then step S30 is modified again.

[0066] Calculation of sample distribution bias rate: Bias rate formula: BiasRate_i = number of samples of technical means i / total number of samples.

[0067] Calculation of feature importance bias rate: Use information gain (IG) to calculate the contribution of each monitoring method to the accident status classification: ;in, is the entropy of the accident state (accident state is the influencing factor triggering the accident state), is the conditional entropy after given monitoring feature Xi.

[0068] Time correlation bias rate: introduce a weight function for the time dimension in the sample, and dynamically adjust the bias rate based on the time correlation of the observation sequence: ;in, is the time decay parameter, T' represents the target time length of the observation sequence, and t represents the current time point.

[0069] Implement the calculation of comprehensive bias rate; combine the above three bias rates to dynamically adjust the weight distribution of factor characteristics: ; Among them: α1, α2, α3 are weight coefficients, satisfying α1+α2+α3=1.

[0070] The comprehensive bias rate is a value calculated by comprehensively calculating the time correlation bias rate, the sample distribution bias rate and the feature importance bias rate. Therefore, even if the target sample data with a higher comprehensive bias rate is screened, see step S31' for details, it is still necessary to perform a specific bias rate analysis in the specific comprehensive bias rate, so as to ensure that the training of the hidden Markov network model λ is more reliable. The reason is that the time correlation bias rate, the sample distribution bias rate and the feature importance bias rate in the comprehensive bias rate will affect the core parameters of the Markov network model.

[0071] Among them, the sample distribution bias rate directly affects the probability distribution of state transition, but state transition is essentially the relationship between latent variables. A sample distribution bias rate that is too high may cause overfitting.

[0072] To this end, the adjustment of the state transition probability can be constrained by combining other information (such as the expert weighted weight B1j of the historical state transition frequency) to avoid simple linear adjustments to the latent variable relationship that only rely on the sample distribution bias rate.

[0073] First, a mapping relationship table between the expert weighted weight B1j of the historical state transition frequency and the sample distribution bias rate range is preset; detect whether the single indicator of the current sample distribution bias rate is higher than the standard sample distribution bias rate threshold, if higher, detect the sample distribution bias rate range where the current sample distribution bias rate is located; then determine the expert weighted weight B1j of the historical state transition frequency; then adjust the state transition probability based on dynamic weighting of the state transition probability: A'=A (A is the state transition probability matrix) × the expert weighted weight B1j of the historical state transition frequency: If the number of samples of the same state (or similar state) is large and the corresponding sample distribution bias rate is high, the transition probability will be amplified, which needs to be reduced by multiplying the expert weight of the historical state transition frequency; conversely, the transition probability of the state with a small number of samples will be reduced.

[0074] Among them, the observation probability matrix B also needs to be dynamically optimized: based on the feature importance bias rate, the contribution of different monitoring methods is recalculated; in complex scenarios, the feature importance of different monitoring methods may change dynamically, which will lead to frequent adjustments to the observation probability matrix, thereby increasing the computational complexity.

[0075] Therefore, when the single indicator based on the feature importance bias rate is higher than the standard feature importance bias rate threshold, it is necessary to return to step S31' to re-screen the adjusted target sample data, so that the observation probability matrix B can be de-redundant to ensure the processing of subsequent steps.

[0076] Therefore, in the third embodiment, during the execution of S30, a hidden Markov network model is constructed based on the element features, and the accident state triggered by the influencing factors is defined as a network node on the hidden Markov network model, which specifically includes the following operations: S31', based on the comprehensive bias rate of the target sample data, screen and obtain the adjusted target sample data (select the target sample data whose comprehensive bias rate is less than the standard comprehensive bias rate threshold), and extract the initial disaster-causing factors, the damage status of the accident-affected body, and the operating status of the emergency response equipment of the accident-affected body based on the adjusted target sample data; S32', based on the adjusted target sample data, extract the initial disaster-causing factors, the damage status of the accident-affected body, and the element characteristics of the operating status of the emergency response equipment of the accident-affected body to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial disaster factor, the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body; Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes serious accident state, high-risk accident state, and potential dangerous accident state; S33', determine the first relay correction condition: determine whether there is a single indicator of the current sample distribution bias rate in the current comprehensive bias rate that is higher than the standard sample distribution bias rate threshold, if higher, detect the sample distribution bias rate range where the current sample distribution bias rate is located; call the mapping relationship table of the expert weighted weight B1j with preset historical state transition frequency and the sample distribution bias rate range; based on the mapping relationship table, search and determine the expert weighted weight B1j of the historical state transition frequency; the current comprehensive bias rate is small and not safe, and the comprehensive bias rate is composed of three parameters, so it does not mean that all the single indicators are low; therefore, it is necessary to add the first relay correction condition and processing process; Then, based on the dynamic weighting of the state transition probability, the state transition probability is adjusted to: A'=A (A is the state transition probability matrix) × the expert weighted weight B1j of the historical state transition frequency: that is, the state transition probability matrix is ​​re-corrected.

[0077] Determine the second relay correction condition: further determine whether there is a situation in the current comprehensive bias rate where the single indicator based on the feature importance bias rate is higher than the standard feature importance bias rate threshold, if so; return to step S31' to re-screen the adjusted target sample data; train based on the adjusted target sample data as a data set to obtain a hidden Markov network model λ'; Where λ'=(A,B,π): Among them, A' is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

[0078] The embodiment of the present application comprehensively considers the impact of the multi-dimensional factors of the dynamic bias rate on the state transition probability of samples and intermediate processes; the bias rate includes sample distribution, feature importance and time correlation. This embodiment introduces a collaborative dynamic adjustment method, and then clarifies how the dynamic bias rate affects the training of the hidden Markov model and the dynamic adaptability of state prediction and the impact of sample data. Among them, clarifying how the dynamic bias rate affects the training and state prediction of the hidden Markov model: refers to the adjustment of the state transition probability matrix and the observation probability matrix. The impact of sample data is to initially screen and select target sample data whose comprehensive bias rate is less than the standard comprehensive bias rate threshold as the original data source.

[0079] Embodiment 4 Further research found that the technical solution implemented in the above-mentioned embodiment 2 still cannot accurately predict the probability of the influencing factors triggering the accident state. The reason lies in the problem of the state transition probability matrix. When the probability values ​​of certain states in the state transition probability matrix are inaccurate, the model will be overly sensitive to data anomalies. In this regard, the cumulative probability model is introduced; After S424, the cumulative probability model is introduced to iteratively calculate the probability of the final observation sequence, which specifically includes: S424. Define the cumulative probability matrix C as an extension of the state transition probability matrix A. Each element C_t(i, j) represents the cumulative transition probability from state i to state j. According to the historical cumulative probability and the current state transition probability, the cumulative probability matrix is ​​updated. Its definition is as follows: C_t(i, j) = α × A(i, j) + (1 - α) ×C_{t-1}(i, j); where A(i, j) is the transition probability in the original state transition probability matrix; C_{t-1}(i, j) is the cumulative transition probability of the previous time step; α is the smoothing coefficient; For example, in the hidden Markov model, the state transition probability matrix A represents the probability of transitioning from state i to state j. After the cumulative probability model is introduced, first, the cumulative probability matrix C is defined; the cumulative probability matrix C is an extension of the state transition probability matrix A, and each element C_t(i, j) represents the cumulative transition probability from state i to state j, which is defined as follows: C_t(i, j) =α× A(i, j)+(1-α) × C_{t-1}(i, j); where A(i, j) is the transition probability in the original state transition probability matrix.

[0080] C_{t-1}(i, j) is the cumulative transition probability of the previous time step. α: Smoothing coefficient, which controls the weight distribution of historical information and current information, and its value range is [0, 1].

[0081] Initialize the cumulative probability matrix: At t=1, the cumulative probability matrix is ​​initialized to the state transition probability matrix: C_1(i, j) = A(i, j); dynamically update the cumulative probability matrix, that is, for each time step t, update the cumulative probability matrix C_t(i, j) according to the historical cumulative probability and the current state transition probability (that is, the transition probability A(i, j) in the original state transition probability matrix).

[0082] S425, Regularization of cumulative probability: In order to prevent the accumulation of deviations in the cumulative probability, C_t(i, j) needs to be regularized to ensure that the sum of the cumulative probability of each row is 1: C_t(i, j) = C_t(i, j) / Σ_j C_t(i,j); S426. Improve the recursive formula of the hidden Markov model: Use the cumulative probability matrix to replace the original state transfer probability matrix and update the forward probability calculation of the hidden Markov model: S4261. Execute initialization forward probability: α1(i) = π_i × b_i(o1); S4262. Execute recursive forward probability: αt+1(i) = Σ_j [αt(j) ×C_t(j, i)] × b_i(ot+1); where C_t(j, i) is the cumulative probability of transitioning from state j to state i in the cumulative probability matrix.

[0083] b_i(ot+1) is the observation probability at time t+1 in the observation sequence.

[0084] S4263. Calculate the probability of the final observation sequence: P(O|λ) = Σ_i αT'(i); Explanation: Because the cumulative probability model incorporates historical information (i.e., the probability values ​​of some states are inaccurate), the short-term impact of outliers will be smoothed.

[0085] Adjusting the smoothing coefficient α can control the balance between historical information and current information, making the model insensitive to data anomalies.

[0086] The technical solution adopted in the embodiment of the present invention is that the cumulative probability model effectively reduces the instantaneous impact of abnormal data on state transition through a time smoothing mechanism (α smoothing coefficient), making the model more robust.

[0087] At the same time, the cumulative probability model can capture the long-term dependence characteristics in the state transfer process, which is particularly important in complex construction scenarios.

[0088] For example, the stress state of a building's supporting structure may gradually deteriorate over a long period of time, and the cumulative probability model can better reflect this trend.

[0089] The introduction of the above-mentioned cumulative probability model only requires the expansion of the state transfer probability matrix, with limited increase in computational complexity, and is easy to implement in the existing hidden Markov model framework. The optimized technical solution overcomes the shortcomings of the state transfer probability matrix and significantly improves the accuracy and robustness of the model's prediction of the accident state trigger probability.

[0090] This solution is applicable to a variety of complex scenarios in construction. It can more accurately predict the risks of accidents such as collapse and provide more reliable decision-making support for construction safety.

[0091] In summary, the above-mentioned safety accident prediction system based on construction site data analysis provided by the present invention can output feedback result information through the above-mentioned Markov network model, and finally can accurately understand the situation of the accident state triggered by the influencing factors at the current time node and the corresponding prediction probability of the situation, thereby improving the detection level of the data to be detected at the current accident site and ensuring the production safety of the construction site.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. A person skilled in the art may modify the technical solutions described in the above embodiments, or replace part or all of the technical features therein with equivalents. However, these modifications or replacements 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.

Claims

1. A safety accident prediction system based on construction site data analysis, characterized in that: Including acquisition module, screening module, construction module, processing output module: The acquisition module is used to collect basic sample data of typical historical accident cases in advance; The screening module is used to screen the basic sample data of the typical historical accident cases to obtain target sample data; The construction module is used to extract the element features including the initial disaster-causing factor, the damage state of the accident-affected body and the operating state of the emergency response equipment of the accident-affected body from the target sample data, and then construct a hidden Markov network model based on the element features, define the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and add connection identifiers between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; The processing output module is used to input the data to be detected and output feedback result information through the Markov network model; The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability.

2. A safety accident prediction method based on construction site data analysis, characterized in that: The steps are as follows: S10. Collect and obtain basic sample data of typical historical accident cases in advance; S20, screening the basic sample data of the typical historical accident cases to obtain target sample data; S30, extracting the element features including the initial disaster-causing factor, the damage state of the accident-affected body and the operating state of the emergency response equipment of the accident-affected body from the target sample data, and then constructing a hidden Markov network model based on the element features, defining the accident state triggered by the influencing factor as a network node on the hidden Markov network model; and adding connection marks between adjacent network nodes, thereby forming a chain model diagram of the hidden Markov network model; S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model; The feedback result information includes the accident status triggered by the influencing factors at the current time node and the corresponding evaluation probability.

3. A safety accident prediction method based on construction site data analysis according to claim 2, characterized in that: During the execution of S20, the basic sample data of the typical historical accident cases are screened to obtain target sample data, specifically including: S21. Initially screen the construction safety accident report data contained in the basic sample data of typical historical accident cases; S22, re-screening the collected basic sample data containing the construction safety accident report data, filtering out the basic sample data lacking influencing factors, and obtaining the basic sample data containing the influencing factors as the target sample data.

4. A safety accident prediction method based on construction site data analysis according to claim 3, characterized in that: During the execution of S30, a hidden Markov network model is constructed based on the element features, and the accident state triggered by the influencing factors is defined as a network node on the hidden Markov network model, which specifically includes the following operations; S31. Obtain target sample data, and extract initial disaster-causing factors, damage status of the accident-affected body, and operating status of emergency response equipment of the accident-affected body based on the target sample data; S32. Extracting the initial disaster-causing factors, the damage status of the accident-affected body, and the element characteristics of the operating status of the emergency response equipment of the accident-affected body based on the target sample data to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial disaster factor, the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body; Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes serious accident state, high-risk accident state, and potential dangerous accident state; S33, training based on the target sample data as a data set to obtain a hidden Markov network model λ; wherein λ=(A, B, π): Among them, A is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

5. A safety accident prediction method based on construction site data analysis according to claim 4, characterized in that: S40, inputting the data to be detected, and outputting feedback result information through the above Markov network model, specifically including: S41, inputting the data to be detected, and obtaining an observation sequence of the target time length T' of the data to be detected; confirming that the observation sequence of the target time length T' is a target observation sequence; wherein the observation sequence of the target time length T' is O={o1, o2, ..., oT'}; S42. Using a dynamic forward planning method, based on the target observation sequence, calculate the probability P(O|λ) of O appearing in the target observation sequence under the hidden Markov network model λ.

6. A safety accident prediction method based on construction site data analysis according to claim 5, characterized in that: S421, obtain the hidden Markov network model λ, define the partial observation sequence at time t as o1, o2…ot and the state as q i The probability of is called the forward probability; The forward probability is calculated as follows: α t (i)=P(o1,o2…o t ,i t =q i |λ), the forward probability α can be recursively calculated t (i) and the observed sequence probability P(O|λ); S422. At the time of initialization, the probability of setting the first moment state to the i-th state is π i , the observation probability of obtaining the observation value o1 in the i-th state is b io1 , the forward probability of the first moment state being the i-th state is: ; S423, starting from time t=1, recursively calculate the forward probability at time t=2, and directly calculate multiple times to obtain the forward probability at the next time t+1: the above multiple calculations refer to starting from t=1, t=2... until the next time t+1 is obtained by multiple calculations, and finally calculating to time T'; The forward probability α of the next time t+1 t+1 (i) = the sum of the forward probability at time t over all states × the probability of the state at time t being observed; ; in, is the forward probability of state j at time t; is the transition probability from state j to state i; When it is the i-th state, we get o t+1 the probability of observing this observed value; S423, output the probability of the final observation sequence: The calculation formula for the probability of the final observation sequence is: ; The above calculation result is output as the probability P(O|λ) of the occurrence of O in the target observation sequence.

7. The safety accident prediction method based on construction site data analysis according to claim 3 is characterized in that: During the execution of S30, it also includes: S31', calculating the comprehensive bias rate of the target sample data, screening and obtaining the adjusted target sample data based on the comprehensive bias rate of the target sample data and the standard comprehensive bias rate threshold, and extracting the initial disaster-causing factors, the damage status of the accident-affected body, and the operating status of the emergency response equipment of the accident-affected body based on the adjusted target sample data; S32', based on the adjusted target sample data, extract the initial disaster-causing factors, the damage status of the accident-affected body, and the element characteristics of the operating status of the emergency response equipment of the accident-affected body to construct a hidden Markov network model of the accident state triggered by the influencing factors; the accident state triggered by the influencing factors has three states, namely, a serious accident state, a high-risk accident state, and a potential dangerous accident state; Set the state set V of the observation state to {v1, v2, ..., vM}, then the number of observation states is M; and {v1, v2, ..., vM} involves three type elements including the initial disaster factor, the damage state of the accident-affected body (S), and the operating state of the emergency response equipment of the accident-affected body; Set the state set Q={q1,q2,...,qn} of the accident state triggered by the influencing factors; the number of states of the accident state triggered by the influencing factors is N; and {q1,q2,...,qn} includes serious accident state, high-risk accident state, and potential dangerous accident state; S33', implement the judgment of the first relay correction condition: judge whether there is a single indicator of the current sample distribution bias rate in the current comprehensive bias rate that is higher than the standard sample distribution bias rate threshold, if higher, detect the sample distribution bias rate range where the current sample distribution bias rate is located; call the mapping relationship table of the expert weighted weight B1j with preset historical state transition frequency and the sample distribution bias rate range; search and determine the expert weighted weight B1j of the historical state transition frequency based on the mapping relationship table; Then, based on the dynamic weighting of the state transition probability, the state transition probability is adjusted to: A'=A (A is the state transition probability matrix) × the expert weighted weight B1j of the historical state transition frequency: Implement the judgment of the second relay correction condition: further judge whether there is a situation in the current comprehensive bias rate that the current single indicator based on the feature importance bias rate is higher than the standard feature importance bias rate threshold, if so; return to step S31' to re-screen the adjusted target sample data; train based on the adjusted target sample data as a data set to obtain a hidden Markov network model λ'; Where λ'=(A,B,π): Among them, A' is the state transition probability matrix; B is the observation probability matrix; π is the initial state probability matrix.

8. The safety accident prediction method based on construction site data analysis according to claim 6 is characterized in that: After S424, the cumulative probability model is introduced to iteratively calculate the probability of the final observation sequence, including: S424. Define the cumulative probability matrix C as an extension of the state transition probability matrix A. Each element C_t(i, j) represents the cumulative transition probability from state i to state j. The cumulative probability matrix is ​​updated according to the historical cumulative probability and the current state transition probability. It is defined as follows: C_t(i, j) =α×A(i, j) + (1-α)× C_{t-1}(i, j); where A(i, j) is the transition probability in the original state transition probability matrix; C_{t-1}(i, j) is the cumulative transition probability of the previous time step; α is the smoothing coefficient; S425, performing regularization processing on C_t(i, j); S426. Use the cumulative probability matrix to replace the original state transfer probability matrix and update the forward probability calculation of the hidden Markov model: S4261. Execute initialization forward probability: α1(i)=π_i × b_i(o1); S4262, perform recursive forward probability: αt+1(i) =Σ_j [αt(j) ×C_t(j, i)] ×b_i(ot+1); where C_t(j, i) is the cumulative probability of transitioning from state j to state i in the cumulative probability matrix; b_i(ot+1) is the observation probability at time t+1 in the observation sequence; S4263. Calculate the probability of the final observation sequence: P(O|λ)=Σ_iαT'(i).

Images