An underground road fire detection method based on BP neural network and D-S evidence theory

Abstract

A method for detecting underground road fires based on a BP neural network and Deductive evidence theory is proposed. To detect M types of underground road fires, information from N types of sensors needs to be collected. These M types of events are defined as a focal element set. The N types of sensors are defined as evidence sources. The data collected by the N types of sensors are preprocessed and then classified according to their source. Based on the data from different types of sensors, a basic probability distribution is performed using a BP neural network or a fuzzy membership function to obtain the judgment results of each type of sensor on the fire event, i.e., the single sensor decision result. All judgment results are re-weighted by calculating evidence support and focal element support. The weighted judgment results are fused using a fusion algorithm based on Deductive evidence theory to obtain the final probability of fire occurrence. The method of the present invention significantly improves detection latency and accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of fire detection technology, and also belongs to the field of data analysis and artificial intelligence, and particularly relates to an underground road fire detection method based on BP neural network and DS evidence theory. Background Art

[0002] As cities grow in size, the contradiction between surface land resources and the ever-increasing demand for transportation is becoming increasingly conflicting. Therefore, the development of urban transportation into underground spaces, promoting the harmonious complementarity of surface and underground road traffic functions, has become a common development trend in major cities at home and abroad. However, underground roads face complex internal and external environmental conditions, and with the surge in underground road traffic volume, the risk of safe operation has increased. Among various underground road disasters, fire is one of the most serious hazards. It not only poses a huge threat to personnel safety, but also causes serious damage to tunnel structures. Currently, accurate and timely detection of fire occurrence is an effective way to reduce the risk of such disasters. However, the commonly used fire detection methods are not ideal in special and complex scenarios, and they are difficult to meet the requirements in terms of detection stability, accuracy, and detection latency.

[0003] Existing fire detection solutions mainly fall into the following categories:

[0004] The first method uses a single sensor type to detect fire events. These sensors typically include temperature sensors, smoke sensors, and flame detectors. This method uses a single signal from a single sensor type to determine whether a fire has occurred. However, sensor results are susceptible to factors such as environmental conditions and sensor stability, leading to false alarms and poor accuracy.

[0005] The second method is to detect fire events through visual sensors. This approach primarily relies on combining visual sensors with artificial intelligence algorithms to identify fires. In complex fire scenarios, this approach typically faces three major challenges. First, relevant disaster data is difficult to collect, and the reliability of artificial intelligence algorithms is significantly reduced in the absence of corresponding scenario datasets. Second, the deployment of visual sensors is significantly affected by the application environment. Depending on the spatial location and observation points of the monitoring scene, multiple sensors may be required to monitor a single location. Third, achieving real-time performance for artificial intelligence image classification algorithms requires high hardware computing power, and this requirement is further increased when processing the results of multiple sensor sources simultaneously. Therefore, the practical feasibility of this approach in complex application scenarios is limited.

[0006] The third approach is to detect fire incidents through multi-sensor data fusion. Traditional sensor technology is already highly advanced, and public spaces are generally equipped with a large number of various sensors, making this approach highly feasible. However, most existing methods are designed for use in scenarios such as buildings, forests, and various safety work environments, and are less applicable in underground road scenarios. Furthermore, existing data fusion detection solutions still have certain flaws in data source selection and algorithm design, resulting in extended detection times and poor accuracy.

[0007] From the above content, it can be seen that the first two methods are difficult to meet the complex scenarios with high requirements for fire detection delay and stability in terms of stability, practical feasibility and cost. In comparison, the multi-sensor data fusion method is a more appropriate choice. Among them, how to process multi-sensor information is the most critical issue. However, the data fusion results obtained by the existing algorithms under some special conditions such as multi-sensor information conflict are difficult to meet the feasibility, accuracy, timeliness and other requirements of fire detection. Summary of the Invention

[0008] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an underground road fire detection method based on BP neural network and DS evidence theory, so as to provide highly reliable detection results even in the presence of a lot of misleading information and significantly reduce the delay of fire detection.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is:

[0010] A method for detecting underground road fires based on BP neural network and DS evidence theory includes the following steps:

[0011] Step 1: To detect M types of underground road fire events, there are N types of sensor information that need to be collected. These M types of events are defined as a focal element set F = {f1, f2, ..., f M}, M ≥ 2, f M represents the Mth focal element, that is, the Mth event; define N types of sensors as evidence sources E = {e1, e2, ..., e M}, N≥1, e N represents the Nth piece of evidence, that is, the data collected by the Nth type of sensor; after preprocessing the data collected by the N types of sensors, they are classified according to the data source;

[0012] Step 2: Based on the data from different types of sensors, basic probability distribution is performed based on BP neural network or fuzzy membership function to obtain the judgment results of each type of sensor on the fire event, that is, the single sensor decision result;

[0013] Step 3: Re-weight all judgment results by calculating evidence support and focal unit support. The evidence support refers to the degree of recognition of a certain evidence for the fire event judgment by other evidence. The more evidence the judgment results are close to the judgment result of the certain evidence, the greater the evidence support. The focal unit support refers to when all evidence determines that the probability of a certain fire event is higher than the probability of other fire events by exceeding a certain threshold, the higher the evidence support for the high-probability event, the higher the credibility, and the greater the focal unit support.

[0014] Step 4: The weighted judgment results are fused through a fusion algorithm based on DS evidence theory to obtain the final fire probability.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The implementation scheme of the present invention has relatively low requirements on hardware equipment. Compared with existing data fusion solutions, the method described in the present invention has obvious improvements in detection delay and accuracy.

[0017] The method of the present invention adopts strategies and corresponding effects for the following common error events:

[0018] 1. If the results obtained from data from different types of sensors tend to two extremes (for example, half of the sensors respond that there is no fire, and half of the sensors respond that there is a fire), the results will be weighted according to the reliability of the physical quantities monitored by the sensors (for example, the stability of oxygen and smoke changes in a fire is higher than that of carbon monoxide and temperature, so the results obtained by sensors monitoring oxygen and smoke are more reliable). This will avoid incidents where conclusions cannot be drawn due to polarized monitoring results, and at the same time make the results closer to the real events.

[0019] 2. If the results from a small subset of sensor types differ from those from other sensors, the algorithm will adhere to the "majority rule" and downweight the least likely alternative. This ensures that the final data fusion decision has a clear bias. (General methods often interfere with the fusion results when different sensor results are present, resulting in a lack of clear direction and useless information, hindering event judgment.) BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a framework diagram of the method of the present invention.

[0021] Figure 2 It is a flow chart of the method of the present invention.

[0022] Figure 3 It is a schematic diagram of the fuzzy membership function of smoke concentration of the present invention.

[0023] Figure 4 It is a schematic diagram of the fuzzy membership function of oxygen concentration difference of the present invention.

[0024] Figure 5 It is the fire detection probability change curve in the open fire scene.

[0025] Figure 6 It is the fire detection probability change curve in the smoldering fire scene. DETAILED DESCRIPTION

[0026] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0027] The underground road fire detection method of the present invention is mainly based on BP neural network and DS evidence theory. Its framework can be referred to Figure 1 ,The execution process of the method is from top to bottom, and the flow of data is also from top to bottom, ,transmitted from the bottom to the top, mainly including: information source acquisition part, ,data processing part, probability allocation part, weight correction part, and result fusion and situation decision ,part.

[0028] Among them, the information source acquisition part is mainly used to obtain the original data required for underground road fire detection. Due to the complex underground road environment, it is necessary to deploy various types of sensors, such as Figure 1 In the figure, type A sensors, type B sensors, and type C sensors are shown. Commonly used sensors include CO concentration sensors, temperature sensors, smoke sensors, and oxygen concentration sensors. These sensors collect parameters such as oxygen, carbon monoxide, smoke, and temperature. These observable quantities can only be increased, not decreased. The sensor type should be selected based on the detection target and detection scenario. The data collected by each sensor is then transmitted and aggregated.

[0029] The data processing part is mainly used for simple processing of original data to facilitate subsequent analysis, calculation and judgment. Its processing mainly includes obtaining parameters such as mean, variance, rate of change (with sampling time as the interval) and interval range. The processed data is still classified according to the data source.

[0030] The probability allocation component is the core of the present invention, relying primarily on a BP neural network and fuzzy membership functions for probability allocation. After receiving data processed by the data processing component, this component inputs the data into a BP neural network or fuzzy membership function based on the data source's susceptibility to interference (typically considered to be susceptible to unexpected factors outside of detection events or inherent instability during detection events). Ultimately, it outputs a fire event determination based on the sensor information.

[0031] The weight correction section, also a core component of the present invention, combines evidence support and focal element support to weight the judgment results corresponding to all sensor types. In this section, weights are assigned to the judgment results based on their reliability, increasing the influence of highly reliable judgment results on the final result and reducing the influence of less reliable judgment results on the final result.

[0032] In the result fusion and situation decision-making part, the judgment results are fused according to the modified weights through the fusion algorithm based on DS evidence theory, and a decision is made based on the fusion results. The decision result is the final probability of fire event occurrence.

[0033] refer to Figure 2 , the specific steps of the present invention are as follows:

[0034] Step 1, data collection

[0035] In order to detect M types of underground road fire events, there are N types of sensor information that need to be collected. These M types of events are defined as the focal element set F = {f1, f2, ..., f M}, M ≥ 2, f M Indicates the Mth focal element, that is, the Mth event. In the present invention, an event generally refers to no fire, smoldering fire or open fire. Define N types of sensors as evidence sources E = {e1, e2, ..., e N}, N≥1, e N Represents the Nth piece of evidence, that is, the data collected by the Nth type of sensor; the data collected by the N types of sensors are preprocessed to facilitate subsequent analysis and calculation, and after preprocessing, they are still classified according to the data source.

[0036] The pretreatment method of the present invention may be as follows:

[0037] First, the data is normalized by category, and the min-max standardization method is used to map the data to the interval [0,1]. Then, the change rate, mean, interval variance, and the absolute value of the difference between the observed value and the normal value are calculated for the normalized data.

[0038] The Min-max normalization formula is: Where x is the original data sample value, and y is the normalized data value. In the embodiment of the present invention, an existing data set is used in the design of the experiment, and the maximum value of the data sample is known. However, in actual events, the maximum value of the data is unknown. Therefore, it is necessary to manually set an appropriate maximum value for normalization during use.

[0039] Assume that the sampling rate of the sensor is KHz, that is, K samples per second.

[0040] Rate of change: That is, take the difference between adjacent sample values and divide it by the sampling rate.

[0041] The difference between the observed value and the normal value: DIF = x i -norm N , norm N is the normalized value of the normal values of N types of sensors.

[0042] Mean: Calculate the mean of 5 adjacent sample values.

[0043] Interval variance: Calculate the variance of 5 adjacent sample values.

[0044] Step 2: Probability Assignment

[0045] According to the data of different types of sensors, basic probability distribution is performed based on BP neural network or fuzzy membership function to obtain the judgment results of each type of sensor on the fire event, that is, the decision result of a single sensor.

[0046] In an embodiment of the present invention, the processed data is input into a BP neural network or fuzzy membership function trained with the corresponding sensor dataset. Specifically, if the data source is susceptible to unexpected factors outside of detection events or is inherently unstable during detection events, a basic probability distribution is performed based on the BP neural network; otherwise, a basic probability distribution is performed based on the fuzzy membership function. The inputs to the BP neural network are the difference between the observed value and the normal value, the average value, and the rate of change; the output is the probability of no fire, smoldering fire, and open fire.

[0047] The present invention specifically defines several criteria for determining which model to input data into: the stability of the observation during continuous observation events, the visibility of the observation's low-dimensional features, and the observation's impact on detection latency. More stable, reliable, and easily analyzed observations require simpler analysis and processing, using fuzzy membership functions, while more complex observations are processed using a BP neural network. Ultimately, both models output a relationship between sensor data and the probability of an event occurring, also known as a single sensor decision result.

[0048] According to the above indicators, in general, the data of CO concentration sensor and temperature sensor should be assigned basic probability based on BP neural network, while the data of smoke sensor and oxygen concentration sensor should be assigned basic probability based on fuzzy membership function.

[0049] (1) The BP neural network is used to construct the basic probability distribution function of temperature and CO concentration. The three preprocessed characteristic values of temperature and CO concentration (the difference between the observed value and the normal value, the average value, and the rate of change) are set as the input of the BP neural network, and the judgment probability of no fire, smoldering fire, and open fire is set as the output of the BP neural network.

[0050] (2) Use the fuzzy membership function to construct the basic probability distribution function of smoke concentration and O2 concentration difference, such as Figure 3 and Figure 4 As shown, there are three obvious nodes in the relationship between the observed values and different fire events. Preliminary detection results can be obtained by comparing the observed data with the node values.

[0051] Step 3: Weight correction.

[0052] The judgment results of all sensors need to be weighted to improve the accuracy of the final decision result. The weight correction of the present invention is based on the support of evidence and the support of focal element.

[0053] Evidence support refers to the degree to which other evidence supports a particular piece of evidence regarding a fire event. The more evidence the results of a particular piece of evidence are similar to the evidence in question, the greater the evidence support. Focus support refers to the degree to which other evidence supports a particular fire event, the greater the credibility of the evidence supporting the higher-probability event.

[0054] Specifically, first, the amount of information contained in the evidence is measured. Assume that each state x of a source is i The probability of occurrence is P i (i=1,2,…,n), then the exponential entropy of the source is:

[0055]

[0056] Then, the decision results of a single sensor are first constructed into an association matrix according to the weight correction algorithm, namely the IM matrix. The IM matrix describes the relationship between evidence and evidence, which is expressed as follows:

[0057]

[0058] The correlation matrix describes the relationship between evidence and evidence. The last element is r NN Instead of r MN , so it is an upper triangular matrix.

[0059] where r ij Indicates the probability m of M events occurring based on the i-th evidence iand the probability m of M events possibly occurring based on the jth evidence j The greater the difference, the greater the r ij The larger the value of r ij The calculation formula is as follows:

[0060]

[0061] in, Indicates evidence i The basic probability distribution function of Indicates evidence j The basic probability distribution function of x represents any element in the focus element set F; Indicates that through evidence i Determine event f x The probability of it happening, Indicates that through evidence j Determine event f x The probability of something happening.

[0062] Then the evidence support and focal element support are calculated respectively to obtain the weight coefficient.

[0063] The formula for calculating the evidence support is as follows:

[0064]

[0065] Due to r ij The larger the value, the greater the difference between the two pieces of evidence, and they cannot provide proof for each other. Therefore, the lower the mutual support between the two pieces of evidence. The definition of Sup(n) reflects the degree of recognition of all other evidence on the judgment of the nth piece of evidence regarding the fire incident. The more evidence the judgment results are close to the nth piece of evidence, the larger Sup(n) is.

[0066] When the following conditions are met, the cross-focal element support is considered when modifying the weights:

[0067]

[0068] ε is a custom constant;

[0069] This judgment condition primarily accounts for events where sensor detection is biased. By comparing all sensors' judgments for different fire states and no fire, if the comparison result is greater than ε, it can be assumed that some sensor detection results are inaccurate, and the impact of these sensor decisions on the final result needs to be reduced. (A larger ε value results in more reliable detection results, but also affects detection latency.) This significantly reduces the impact of erroneous data on the final result.

[0070] The weight coefficients after re-weighting are as follows:

[0071]

[0072] Among them, w k represents the kth piece of evidence e k The weight coefficient of Sup(i) is the i-th evidence e i The degree of evidence support.

[0073] Finally, we can get the corresponding matrix of the relationship between the evidence source and the focal element probability:

[0074]

[0075] Among them, P NM It refers to judging the probability of the Mth event occurring based on the Nth piece of evidence.

[0076] In the present invention, the basic probability distribution of the corrected evidence can be further expressed as follows:

[0077]

[0078] where m i is the basic probability distribution vector corresponding to evidence i, and M is the sum of the basic probability distribution vectors of all revised evidence.

[0079] Step 4: Result fusion and situation decision making

[0080] The weighted judgment results are fused through a fusion algorithm based on DS evidence theory to obtain the final fire occurrence probability.

[0081] Since it is a multi-evidence fusion, the fusion formula for N pieces of evidence is:

[0082]

[0083] Among them, m(f x ) is the event f x The probability of occurrence; A1 is any focal element corresponding to the first piece of evidence, A N is any focal element corresponding to the Nth piece of evidence; m1(A1) is the probability of A1 occurring as inferred from the first piece of evidence, m N (A N ) is A inferred from the Nth evidence N The probability of something happening;

[0084] In evidence theory, evidence is converted into event probability by relying on probability distribution function. The input of probability distribution function is the data of evidence source, and the output is the probability of various events. Any event that is expected to be inferred can be freely defined, but the evidence needs to be able to prove the possibility of the defined event to be valid. If there is an invalid focal element, then its probability is always 0. Therefore, the present invention defines the focal element corresponding to each piece of evidence separately. In the above formula, the focal elements that can be inferred by all evidence can also be set to be the same. The premise of this method is that all focal elements are independent of each other and there is no intersection.

[0085] According to the final fusion result, each focal element will have a probability value of occurrence, and a maximum value is set using the threshold method to determine whether a certain event (no fire, smoldering fire, or open fire) is likely to occur.

[0086] In the present invention, since the result of data fusion is only the probability of several possible events, a logic needs to be designed to select which event is most likely to occur. This embodiment uses a threshold method for judgment. For example, when the probability of no fire is higher than 0.8, it is determined that no fire has occurred; when the probability of smoldering fire is higher than 0.7 and the probability of no fire is lower than 0.1, it is determined that a smoldering fire has occurred; when the probability of open fire is higher than 0.7 and the probability of no fire is lower than 0.1, it is determined that an open fire has occurred. Only when the sum of the probabilities of smoldering fire and open fire is higher than 0.9 can a fire be determined to have occurred. Before that, it is in an uncertain state. The time from the occurrence of a fire to the system's determination that a fire has actually occurred is the detection delay.

[0087] The method of the present invention was verified and compared with the existing mainstream methods. The results are as follows Figure 5 and Figure 6 As shown in the figure, method 1 is the solution provided by the present invention; method 2 is a solution that uses JS distance to correct weights; and method 3 is a solution that does not correct weights. It can be seen that the method of the present invention improves the detection delay by 2-3 seconds.

Claims

1. A method for detecting underground road fire based on BP neural network and DS evidence theory, characterized in that: include: Step 1: To detect M types of underground road fire events, there are N types of sensor information that need to be collected. These M types of events are defined as a focal element set F = {f1, f2, ..., f M }, M ≥ 2, f M represents the Mth focal element, that is, the Mth event; define N types of sensors as evidence sources E = {e1, e2, ..., e N }, N≥1, e N represents the Nth piece of evidence, that is, the data collected by the Nth type of sensor; after preprocessing the data collected by the N types of sensors, they are classified according to the data source; Step 2: Based on the data from different types of sensors, basic probability distribution is performed based on BP neural network or fuzzy membership function to obtain the judgment results of each type of sensor on the fire event, that is, the single sensor decision result; Step 3: Re-weight all judgment results by calculating evidence support and focal unit support. The evidence support refers to the degree of recognition of a certain evidence for the fire event judgment by other evidence. The more evidence the judgment results are close to the judgment result of the certain evidence, the greater the evidence support. The focal unit support refers to when all evidence determines that the probability of a certain fire event is higher than the probability of other fire events by exceeding a certain threshold, the higher the evidence support for the high-probability event, the higher the credibility, and the greater the focal unit support. Step 4: The weighted judgment results are fused using a fusion algorithm based on DS evidence theory to obtain the final fire occurrence probability; Among them, in step 3, the decision result of a single sensor is first constructed into an association matrix according to the weight correction algorithm, namely the IM matrix; then the evidence support and the focal element support are calculated respectively to obtain the weight coefficient; finally, the corresponding matrix of the probability relationship between the evidence source and the focal element is obtained Among them, P NM It refers to judging the probability of the Mth event by the Nth piece of evidence; The constructed correlation matrix describes the relationship between evidence and evidence, which is expressed as follows: where r ij Indicates the probability m of M events occurring based on the i-th evidence i and the probability m of M events possibly occurring based on the jth evidence j The greater the difference, the greater the r ij The larger the value of r ij The calculation formula is as follows: in, Indicates evidence i The basic probability distribution function of Indicates evidence j The basic probability distribution function of x represents any element in the focus element set F; Indicates that through evidence i Determine event f x The probability of it happening, Indicates that through evidence j Determine event f x The probability of something happening; The calculation formula for the evidence support is as follows: When the following conditions are met, the cross-focal element support is considered when modifying the weights: ε is a custom constant; The weight coefficients after re-weighting are as follows: Among them, w k represents the kth piece of evidence e k The weight coefficient of Sup(i) is the i-th evidence e i The degree of evidence support.

2. The underground road fire detection method based on BP neural network and DS evidence theory according to claim 1 is characterized in that: The method of the step 1, pretreatment is as follows: First, the data is normalized by category and mapped to the interval [0,1]. Then, the rate of change, mean, interval variance, and the absolute value of the difference between the observed value and the normal value are calculated for the normalized data.

3. The underground road fire detection method based on BP neural network and DS evidence theory according to claim 1 is characterized in that: In step 2, if the data source is susceptible to unexpected factors of non-detection events or is unstable in detection events, basic probability distribution is performed based on the BP neural network; otherwise, basic probability distribution is performed based on the fuzzy membership function.

4. The underground road fire detection method based on BP neural network and DS evidence theory according to claim 1 is characterized in that: The sensors include: a CO concentration sensor, a temperature sensor, a smoke sensor, and an oxygen concentration sensor; the data of the CO concentration sensor and the temperature sensor are subjected to basic probability distribution based on a BP neural network; the data of the smoke sensor and the oxygen concentration sensor are subjected to basic probability distribution based on a fuzzy membership function; the input of the BP neural network is the difference between the observed value and the normal value, the average value, and the rate of change; the output is the judgment probability of no fire, smoldering fire, and open fire.

5. The underground road fire detection method based on BP neural network and DS evidence theory according to claim 1 is characterized in that: In step 4, the fusion formula of N evidences is: Among them, m(f x ) is the event f x The probability of occurrence; A1 is any focal element corresponding to the first piece of evidence, A N is any focal element corresponding to the Nth piece of evidence; m1(A1) is the probability of A1 occurring as inferred from the first piece of evidence, m N (A N ) is A inferred from the Nth evidence N The probability of something happening; 6. The underground road fire detection method based on BP neural network and DS evidence theory according to claim 1 is characterized in that: In step 4, based on the final fusion result, each focal element will have a probability value of occurrence, and a threshold method is used to set a maximum value to determine whether a certain event is likely to occur. The certain event refers to one of the three events: no fire, smoldering fire, and open fire.

Images