Forest fire hazard grade evaluation method based on reliable expansion confidence rule base

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to a forest fire hazard grade evaluation method based on a reliable expansion confidence rule base. The reliability modeling comprises the steps of building a forest fire hazard grade element attribute set and calculating the attribute matching degree of an extended confidence rule by using a Gaussian kernel function, and the reliability model reasoning comprises the step of calculating the comprehensive matching degree of an extended confidence rule base by using a copula function. Extended belief rule parameter optimization can be explained, and belief rule validity analysis is carried out by calculating extended belief rule profit and loss values. The forest fire hazard grade assessment method has good reliability in the aspect of forest fire hazard grade assessment, and can accurately and reliably assess the forest fire hazard grade according to the forest fire data.

Description

Technical Field

[0001] The invention belongs to the technical field of artificial intelligence and relates to a forest fire hazard level assessment method based on a reliable extended confidence rule base. Background Art

[0002] In recent years, forest fires have become more frequent, and responding to them has become increasingly challenging. In response to the needs of social development and security, people are paying more and more attention to the hazards of forest fires. Forest fire hazard level assessment is an important reference for the severity of fire damage, and it is also a prerequisite for organizing rescue and formulating resource allocation plans. Therefore, accurate and reliable assessment of fire hazard levels is crucial for responding to forest fires.

[0003] Thanks to the development of data science in the field of natural disasters, the current forest fire hazard level assessment method is mainly data-driven. Compared with expert experience, data-driven has great advantages of objectivity and accuracy, especially in the field of forest fire emergency management. If the fire department can accurately obtain forest fire hazard level information, it can issue early warnings and prepare response plans. As a popular expert system method, the extended confidence rule base model is suitable for natural disaster hazard assessment. It adopts the IF-THEN rule form, which is closer to human's natural understanding and expression of knowledge, and can describe the complex nonlinear relationship between input and output information. At the same time, fuzzy theory is used in the data input stage, and the modeling process is more flexible. In addition, the confidence rule base model is a white box method. The derivation process is visible throughout, and the results are traceable, which provides good interpretability for decision makers. The extended confidence rule base has achieved remarkable research results in many fields.

[0004] However, in the forest fire hazard level assessment, the traditional extended confidence rule base has defects in modeling reliability and model reasoning reliability, resulting in poor reliability of the assessment results, which is insufficient to provide reliable decision support for the fire department. Therefore, improving the modeling reliability and model reasoning reliability of the extended confidence rule base has become the top priority of forest fire hazard level assessment. Summary of the invention

[0005] The present invention provides a forest fire hazard level assessment method based on a reliable extended confidence rule base (REBRB), and elaborates on the specific method flow of reliability modeling and reliability model reasoning of the reliable extended confidence rule base in forest fire hazard level assessment, wherein the reliability modeling includes constructing a forest fire hazard level element attribute set and using a Gaussian kernel function to calculate the matching degree of the extended confidence rule attribute, and the reliability model reasoning includes using a copula function to calculate the comprehensive matching degree of the extended confidence rule base, and can explain the extended confidence rule parameter optimization and calculate the extended confidence rule profit and loss value to analyze the validity of the confidence rule. The present invention has good reliability in forest fire hazard level assessment, and can accurately and reliably assess the forest fire hazard level according to forest fire data.

[0006] The technical solution of the present invention is as follows:

[0007] A forest fire hazard level assessment method based on a reliable extended confidence rule base, the specific steps are:

[0008] Step 1: Construct a forest fire hazard level feature attribute set;

[0009] Forest fire events are composed of many factors that affect fires and the relationships between these factors. The degree of forest fire damage is the result of the interaction between hazard factors and hazard-bearing bodies in a specific disaster-prone environment. Therefore, the set of fire hazard level elements can be formally expressed as

[0010] Danger Level=φ(DIF,DAB,DPE,RL) (1)

[0011] Danger Level represents the forest fire hazard level, which is a nonlinear function of DIF, DAB, DPE, and RL, and is represented by the function symbol φ. R = {r d |d=1,…,D}.DIF represents the set of disaster factors, represents the attribute of the a-th hazard factor element, A represents the total number of hazard factor attributes, DAB represents the set of hazard-affected bodies, represents the attribute of the bth disaster-prone body element, B represents the total number of disaster-prone body attributes, DPE represents the set of disaster-prone environments, represents the cth disaster-prone environmental factor attribute, C represents the total number of disaster-prone environmental attributes, R represents the set of association relationships, and rl d represents the relationship between any two elements, and D represents the total number of relationships.

[0012] Given a set of attributes of all forest fire prerequisite elements And all attribute reference levels The initial utility value

[0013] Among them J a , J b , J c Respectively The total number of reference levels in .

[0014] In order to simplify the representation of the above set, the attribute set of the premise elements where the forest is located can be represented as V = {v1, v2, ..., v M ,in Given all attribute reference levels J m express The total number of reference levels, given the initial utility

[0015] Step 2: Use the Gaussian kernel function to calculate the extended confidence rule attribute matching degree;

[0016] To evaluate the forest fire hazard level, K groups of forest fire data are collected from the forest fire dataset. <v k , z k >(k=1,...,K), where v represents the kth group of forest fire feature attribute input value vectors, k，m represents the mth forest fire feature attribute value in the kth set of input value vectors, z k represents the output value of the kth forest fire hazard level; then, the utility theory is used to transform the information, and the K groups of data are transformed into the distributed confidence of the premise attributes and the result attributes. The calculation formula of the distributed confidence in the mth forest fire premise factor attribute is as follows:

[0017]

[0018] in Indicates v k，m Distribution confidence

[0019] Similarly, the output value z k The output attribute distribution confidence of forest fire hazard level can be calculated

[0020]

[0021] The obtained K groups of distributed confidences about the attributes of the premise forest fire elements and the attributes of the forest fire hazard results are used to calculate the antecedent similarity AS and the posterior similarity CS of each rule in the rule base:

[0022]

[0023] in, It represents the inconsistency between the mth attribute of the lth rule and the kth rule. σ is the adjustment parameter. The matching degree of the rule attributes can be adjusted by controlling the value of σ.

[0024]

[0025] Usually σ = 0.2, so that In its domain, it is similar to AS(R l , R k ) is close to a linear relationship.

[0026] Similarly, the calculation method of the subsequent similarity CS is

[0027]

[0028] Among them, dc l，k Indicates the inconsistency between the results of the lth rule and the kth rule,

[0029]

[0030] Assume that the input value vector x of a forest fire hazard assessment problem is p =(x p，1 , …, x p，M ), p = 1, ..., P, P is the total number of input vectors, and each input value in the input value vector is converted into a distributed representation using the formula:

[0031]

[0032] Next, input vector x p =(x p，1 …, x p，M ) and the individual matching degree of the m-th premise attribute of the k-th extension rule is calculated as follows:

[0033]

[0034] Among them, σ is a tuning parameter, and the rule attribute matching degree SS can be adjusted by controlling the value of σ. k (x p，m , v k，m ) is adjusted, usually taking σ=0.2, so that In its domain, it is similar to SS k (x p，m , v k，m ) is close to a linear relationship.

[0035] in, is the inconsistency between the input vector and the mth premise attribute of the kth extension rule

[0036]

[0037] Step 3: Use the copula function to calculate the comprehensive matching degree of the extended confidence rule;

[0038] Assume (v h , z h ) and (v m , z m ) are two sets of random variables of forest fire factors, and the corresponding Kendall rank correlation coefficient τ can be obtained. τ can measure the consistency of the change direction of the two random variables. τ=1, τ=-1 and τ→0 respectively represent the three situations where the change directions of the two random variables are completely consistent, completely opposite and completely independent.

[0039] The Frank copula function can be used to obtain the correlation between all forest fire element attributes.

[0040] Among them, λ∈[1,+∞], when λ→0, the variables tend to be more independent; when λ>0, it means that the variables are positively correlated; when λ<0, it means that the variables are negatively correlated.

[0041] There is a certain relationship between the parameters of the Copula function and the correlation index, that is, the Kendall coefficient can be uniquely represented by the copula function. Therefore, the copula parameters can be determined based on the Kendall rank correlation coefficient τ. The copula parameter λ is the correlation coefficient between the two factors. According to the property that the correlation coefficient between multiple factors is less than the correlation coefficient between any two factors, the minimum value of the correlation coefficient can be selected to calculate the estimated value of the multi-factor correlation coefficient.

[0042]

[0043] The cumulative attribute matching degree of the kth rule is actually the joint probability value obtained by multiplying the matching degrees of individual attributes. In other words, this is consistent with the result calculated by the copula function. Therefore, the comprehensive matching degree of the kth rule calculated according to the copula function is:

[0044]

[0045] According to the comprehensive matching degree S k , rule weight θ k and attribute weight δ m , the activation weight formula of the kth rule is γ k

[0046] Using the ER method to analyze the formula, the distributed confidence of the result attributes of forest fire hazards in all activation rules is synthesized into a new distributed confidence ρn . According to the input vector x p =(x p，1 , …, x p，M ), the forest fire hazard level can be obtained based on REBRB reasoning as f(x p ).

[0047] Step 4: Interpretable extended confidence rule parameter optimization;

[0048] In the construction of REBRB, the candidate grade utility values ​​of forest fire premise factor attributes, the grade utility values ​​of forest fire result attributes and attribute weights are parameters to be optimized, which are usually initialized by experts based on experience. In order to make the parameter optimization process objective and interpretable, the model of the parameter optimization method is given below.

[0049] REBRB parameter optimization: When the inference evaluation result of the REBRB model for each set of data is f(x p ), then the parameter optimization model of REBRB can be expressed as:

[0050]

[0051] Where E p Indicates the number of cases where the actual hazard level is not equal to the model output result. The objective function requires E p Minimize. lb m andub m Indicates the lower and upper bounds of the value in the mth premise attribute; lb n andub n Indicates the lower and upper bounds of the value in the result attribute; the interpretability of parameter optimization is reflected in the parameter restriction conditions in formula (15), which means that the grade utility value is monotonically increasing or monotonically decreasing, or monotonically increasing on the left side of the maximum grade utility and monotonically decreasing on the right side, which can ensure that there is no contradiction in grade utility. Then the DE algorithm can be used to optimize the parameters of the above model, and finally the optimized REBRB parameters are obtained.

[0052] Step 5: Calculate the profit and loss value of the extended confidence rule to analyze the effectiveness of the rule;

[0053] The purpose of REBRB rule validity analysis is to identify invalid rules and remove them from the rule base, retain valid rules for forest fire hazard level assessment, and thus improve the reliability of REBRB in forest fire hazard level assessment.

[0054] Rule validity analysis is a multi-iterative process. Assuming that there are L rules in the rule base, the profit and loss value of each rule needs to be calculated in each iteration:

[0055]

[0056] r w represents the wth iteration. When the profit and loss value is -1, Recorded as an invalid rule, when the profit and loss value is 1, Recorded as a valid rule, Indicates the current rule base Contribution obtained.

[0057]

[0058] Among them, z p represents the actual forest fire hazard level in the training set, Indicates that w In the iteration of round, the input vector is x p When the forest fire hazard level is obtained by using the l rules of REBRB to make inferences, Indicates that w In the iteration of round, the input vector is x p When using the l-1 rule of REBRB, that is, not including the R l The forest fire hazard level is obtained by reasoning and evaluating the rules. illustrate Its cumulative contribution is negative during the reasoning process of the dataset, lacking It will make the reasoning better; illustrate Its contribution is cumulatively positive during the reasoning process of the dataset, and it is missing It will make the inference result worse. w ) indicates that in r w In the iteration of the round, the parameter set of the optimized rules in REBRB It should be noted that the rule parameters are determined based on the performance of the entire rule base. Changes in the number of rules will also lead to changes in the parameters. After deleting invalid rules, all remaining rules in the rule base need to be optimized and updated again to ensure the validity of the rule parameters.

[0059] Figure 2The whole process of rule validity analysis is demonstrated. First, the DE evolutionary algorithm is used to optimize the rule parameters of the rules in the initial rule base. Then, the rule validity analysis is performed to calculate the profit and loss value of each rule. When the profit and loss value is not 1, the rule is recorded as an invalid rule and removed. When the profit and loss value is 1, the rule is recorded as a valid rule and put into the rule base. The above steps are repeated until there are no invalid rules in the rule base, and finally a valid rule base composed of valid rules is obtained. Specifically, for example, in the first round of iterations, the L rules in REBRB are used for reasoning, and each rule needs to be analyzed for validity. In this iteration, a total of l1 invalid rules are deleted, and the remaining L-l1 rules are further optimized by the DE optimization algorithm. In the second round of iterations, the L-l1 rules whose parameters have been optimized in the first round are also used for reasoning. Each rule needs to be analyzed for validity. In this iteration, a total of l2 invalid rules are deleted, and the remaining L-l1-l2 rules are further optimized by the DE optimization algorithm. This process is repeated until the Wth round, when all rules are analyzed for validity and no invalid rules are found, then all remaining L1-l1-l2, ..., -l W-1 The valid rules are organized into a valid rule base, and L1-l1-l2, ..., -l W-1 The reasoning is performed based on the rules, and the resulting reasoning evaluation result is used as the forest fire hazard level.

[0060] Beneficial effects of the present invention:

[0061] This method can enhance the reliability of forest fire hazard level assessment. On the one hand, it avoids the collapse of rules and ensures the smooth implementation of the method. On the other hand, it considers the role relationship of attributes and the effectiveness of rules, thereby improving the accuracy and reliability of forest fire hazard level assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of the steps of the forest fire hazard level assessment method based on a reliable confidence rule base;

[0063] Figure 2 Schematic diagram of the rule effectiveness analysis process;

[0064] Figure 3 A heat map of the geographical information of forest fire hazard level in a certain area based on a reliable confidence rule base. DETAILED DESCRIPTION

[0065] The evaluation method of the present invention can reliably evaluate the forest fire hazard level based on the forest fire element attribute information, and comprehensively evaluate the forest fire hazard level using forest fire data such as meteorological conditions, vegetation types, topography, fire characteristics and other factors, and can quickly adapt to environmental changes, timely update the evaluation results, and display the forest fire occurrence area in the form of intuitive charts and maps, which is convenient for users to understand and make decisions, and helps forest fire prevention and extinguishing technology companies to provide reliable forest fire hazard level evaluation services to forestry and fire departments in a timely and accurate manner, so as to formulate fire extinguishing preparations and resource deployment plans. The system design adopts a B / S structure, that is, a browser / server (Browser / Server) architecture, and is built based on the Flask framework, involving web technologies such as HTML and CSS.

[0066] The system architecture can be divided into the view layer, logic layer and data layer, as shown in Table 1:

[0067] Table 1 System architecture

[0068]

[0069] 1. Upload basic forest fire data

[0070] Weather data of the fire area can be obtained from the meteorological department, and fire spread data, geographic information data and vegetation data of the fire area can be obtained from forest fire infrared sensors, monitoring videos and remote sensing satellites, and these real-time data can be stored in the data layer. In addition, historical forest fire data is also uploaded to the data layer, and historical fires in the area or fires on similar terrain can be searched.

[0071] 2. The system analyzes and diagnoses the uploaded data

[0072] The experiment of this function requires the data layer, logic layer and view layer of the system to work together, as follows:

[0073] Step 1: The logic layer performs modal missing detection on the forest fire data uploaded by the data layer, and can use interpolation or historical data similarity calculation to fill in the missing data.

[0074] Step 2: Use the reliable confidence rule base method to perform reliability modeling, extract forest fire characteristics, generate a forest fire element attribute set, and convert all data into extended confidence rules. Use the Gaussian kernel function to calculate the extended confidence rule attribute matching degree.

[0075] Step 3: After the modeling is completed, the reliability model reasoning is then performed using the reliable confidence rule base method, and the comprehensive matching degree of the extended confidence rule is calculated using the copula function. After the calculation is completed, the historical fire data uploaded by the data layer is used to train the extended confidence rule parameters until the optimization stop condition is reached, and the optimized parameters are generated. Then, the rule profit and loss value is calculated to delete the invalid rules and organize an effective extended confidence rule base. The real-time forest fire data is passed into the reliable extended confidence rule base model to solve the forest fire hazard level.

[0076] Step 4: Upload the output hazard level results to the data layer for storage, and use ArcGIS to draw a geographic heat map of the forest fire hazard level results in the data layer. Figure 3 As shown, the view layer feeds back the visualized hazard level assessment results and the geographic information heat map of the fire area to the user, and displays the real-time information of the forest fire in the information bar.

[0077] 3. Users can visually view the forest fire hazard level

[0078] After the system completes the entire process of forest fire hazard level assessment, the assessment results are stored in the database. Users can click on the map to view the real-time forest fire hazard level results. Click on the fire area to learn more about the current basic fire information and areas of different hazard levels. Emergency plans can be initiated for different regions according to the hazard level results. Finally, the system assessment results can be used as a means to assist emergency decision-making.

Claims

1. A forest fire hazard level assessment method based on a reliable confidence rule base, characterized in that: The specific steps are as follows: Step 1: Construct a forest fire hazard level element set; The set of forest fire hazard levels is expressed as Danger Level=Φ(DIF,DAB,DPE,RL) (1) Given a set of attributes of all forest fire prerequisite elements And all attribute reference levels The initial utility value The attribute set of the premise elements where the forest is located is simplified as V = {v1, v2, ..., v M ),in Given all attribute reference levels Given the initial utility Step 2: Use the Gaussian kernel function to calculate the extended confidence rule attribute matching degree; Collect K groups of forest fire data from the forest fire dataset <v k , z k >(k=1,...,K),v k，m Yes k The mth attribute in ; K groups of data are converted into distributed confidence of premise attributes and result attributes. The calculation formula of distributed confidence in the mth forest fire premise factor attribute is as follows: Among them, v k，m Distribution confidence Similarly, the output value z k The output attribute distribution confidence of forest fire hazard level can be calculated The obtained K groups of distributed confidences about the attributes of the premise forest fire elements and the attributes of the forest fire hazard results are used to calculate the antecedent similarity AS and the posterior similarity CS of each rule in the rule base: in, It represents the inconsistency between the mth attribute of the lth rule and the kth rule. v is the adjustment parameter. The rule attribute matching degree can be adjusted by controlling the value of σ. Similarly, the calculation method of the subsequent similarity CS is Among them, dc l，k Indicates the inconsistency between the results of the lth rule and the kth rule, Suppose the input value vector x of a forest fire hazard assessment problem is p =(x p，1 , …, x p，M ), converting each input value in the input value vector into a distributed representation: Next, input vector x p =(x p，1 , …, x p，M ) and the individual matching degree of the mth premise attribute of the kth extension rule is: Where σ is a tuning parameter, and the rule attribute matching degree SS is adjusted by controlling the value of σ. k (x p，m , v k，m ) to make adjustments, is the inconsistency between the input vector and the mth premise attribute of the kth extension rule Step 3: Use the copula function to calculate the comprehensive matching degree of the extended confidence rule; The Frank copula function can be used to obtain the correlation between all forest fire element attributes. According to the property that the correlation coefficient between multiple factors is smaller than the correlation coefficient between any two factors, the minimum value among the correlation coefficients is selected to calculate the estimated value of the multi-factor correlation coefficient. The comprehensive matching degree of the kth rule is calculated according to the copula function: Step 4: Interpretable extended confidence rule parameter optimization; Step 5: Calculate the profit and loss value of the extended confidence rule to analyze the effectiveness of the rule.

2. A method for evaluating forest fire hazard level based on a reliable confidence rule base as claimed in claim 1, characterized in that: The step 4 can explain the extended confidence rule parameter optimization, and the specific operations are as follows; When the inference evaluation result of the REBRB model for each set of data is f(x p ), then the parameter optimization model of REBRB is expressed as: Where E p Indicates the number of cases where the actual hazard level is not equal to the model output result. The objective function requires E p Minimize; the interpretability of parameter optimization is reflected in the parameter constraints in formula (15). Then the DE algorithm can be used to optimize the parameters of the above model, and finally the optimized REBRB parameters are obtained.

3. A method for evaluating forest fire hazard level based on a reliable confidence rule base as claimed in claim 1 or 2, characterized in that: The step 5 calculates the profit and loss value of the extended confidence rule to analyze the effectiveness of the rule. The specific operation is as follows: Rule validity analysis is a multi-iterative process. Assume that there are L rules in the rule base. In each iteration, the profit and loss value of each rule needs to be calculated: r w represents the wth iteration; when the profit and loss value is -1, Recorded as an invalid rule, when the profit and loss value is 1, Recorded as a valid rule, Indicates the current rule base Contributions received; Among them, z p represents the actual forest fire hazard value in the training set, Indicates that w In the iteration of the round, the forest fire hazard level obtained is Indicates that w In the iteration of round, the Rth l The forest fire hazard level obtained by reasoning and evaluating the rules; ξ(r w ) indicates that in r w In the iteration of the round, the parameter set of the optimized rules in REBRB 4. A method for evaluating forest fire hazard level based on a reliable confidence rule base as claimed in claim 3, characterized in that: The illustrate Its cumulative contribution is negative during the reasoning process of the dataset, lacking It will make the reasoning better; illustrate Its contribution is cumulatively positive during the reasoning process of the dataset, and it is missing This will make the inference results worse.

Images