High-altitude tunnel fire automatic early warning method based on intelligent sensor network
By constructing data feature factors and synchronization indexes in high-altitude tunnel fire early warning and selecting an LSTM prediction algorithm with an appropriate activation function, the problem of the LSTM prediction algorithm's slow response to early fire signals is solved, and earlier and more reliable fire early warning is achieved.
Patent Information
- Application Number
- CN202511324239.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-17
- Publication Date
- 2025-12-23
AI Technical Summary
Existing LSTM prediction algorithms are slow to respond to early fire signals in high-altitude tunnel fire early warning systems, failing to effectively capture weak fire signals, resulting in delayed warnings and reduced safety.
By constructing data weak amplitude trend change characteristic factors, data synchronization index, and data cross-correlation degree within a time window, the confidence level of the initial fire situation is obtained. An appropriate activation function is selected as the input gate of the LSTM prediction algorithm to improve the ability to keenly capture the initial signals of a fire.
It enables earlier and more reliable automatic fire warnings, improves the ability to identify early fire signals, and ensures the timeliness of personnel evacuation and rescue.
Smart Images
Figure CN121191256A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks. Background Technology
[0002] Due to the thin air, low oxygen content, and significant differences in air pressure and temperature at high altitudes, fires often rely primarily on radiative heat transfer. Heat accumulates within the enclosed, narrow tunnel spaces, and once a fire enters its growth and spread phase, the temperature rise and delayed spread are extremely rapid, drastically reducing evacuation and rescue time. Furthermore, the terrain and geological conditions at high altitudes often result in curved tunnels, causing uneven smoke flow and flame radiation. These curves can create areas prone to high-temperature accumulation and low-visibility blind spots, further exacerbating the dangers of evacuation and rescue. Therefore, early detection and identification of fires in high-altitude curved tunnels are crucial to gain valuable time for evacuation and provide a proactive approach to rescue and ventilation control.
[0003] Traditional methods typically use predictive algorithms (such as LSTM) to model sensor data, creating predictive models that forecast environmental data for future moments to detect fires early and deploy rapid response measures. However, in high-altitude environments, the low pressure in tunnels leads to thin oxygen, resulting in flame heights slightly higher than at normal pressure. Heat diffusion is slow, local temperature rise and smoke concentration changes are gradual, and fires are primarily radiation-driven. Early fires exhibit a slower rate of heat release and smaller fluctuations. Because the LSTM prediction algorithm uses the sigmoid activation function by default for its input gates, which mathematically modulates data exponentially, the output tends to saturate when the value is close to 0 or 1. This results in small gradient changes, making it insensitive to small data fluctuations. It can easily compress the data in the input gates to near-zero values, ultimately preventing early, weak fire signals from being effectively written into the cell states. Consequently, the LSTM model fails to capture early trends effectively, resulting in sluggish output predictions that react slowly to early signals, delayed early warnings, and reduced safety.
[0004] Therefore, improving the ability of prediction models to capture early signs of fires, and thus improving the accuracy of prediction models, has become an urgent problem to be solved. Summary of the Invention
[0005] In view of this, embodiments of the present invention provide an automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks, in order to solve the problem of how to improve the ability of prediction models to sensitively capture early fire signals, thereby improving the accuracy of prediction models.
[0006] This invention provides an automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks. The method includes the following steps: For any monitoring point in the tunnel, obtain the multidimensional environmental parameters of that monitoring point at each moment; Construct a time window of a preset length containing the current moment, and obtain a weak amplitude trend change characteristic factor of data change within the time window of any monitoring point based on the multidimensional environmental parameters within the time window of any monitoring point. A preset number of reference monitoring points are obtained for any given monitoring point. The multidimensional environmental parameters of any given monitoring point and its reference monitoring points within the time window are arranged into at least two data matrices according to their respective dimensions. Based on the similarity characteristics of the data change trends in each data matrix, the data synchronization index between any given monitoring point and its reference monitoring points is obtained. Based on the data correlation characteristics in each data matrix, the degree of data cross-correlation between any given monitoring point and its reference monitoring points is obtained. Based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree, the initial fire confidence level of any monitoring point at the current time is obtained. Based on the initial fire confidence level, the activation function of the input gate in the LSTM prediction algorithm is selected to obtain the LSTM prediction model, which is used to predict the fire at any monitoring point at the next time.
[0007] Preferably, the step of obtaining a data weak amplitude trend change characteristic factor to characterize the data change within the time window of any monitoring point based on multidimensional environmental parameters within the time window of any monitoring point includes: The multidimensional environmental parameters within the time window are divided into at least two environmental parameter sequences according to their respective dimensions. The difference between the maximum and minimum values in each environmental parameter sequence is accumulated to obtain the accumulated environmental parameter difference value. The accumulated environmental parameter difference value is normalized to obtain the normalized accumulated environmental parameter difference value. The difference between the constant 1 and the normalized accumulated environmental parameter difference value is obtained to determine the degree of data change of any monitoring point within the time window. For any environmental parameter sequence, a coordinate graph of the environmental parameter sequence is constructed. The horizontal axis of the coordinate graph is time, and the vertical axis is the environmental parameter at each time in the environmental parameter sequence. The environmental parameter sequence is fitted to obtain a fitted straight line. The distance between each environmental parameter in the environmental parameter sequence and the fitted straight line is obtained in the coordinate graph, and the data deviation value of the environmental parameter sequence is obtained accordingly. The data deviation values of each environmental parameter sequence are accumulated to obtain the accumulated data deviation value. The accumulated data deviation value is then normalized to obtain the normalized accumulated data deviation value. The difference between the constant 1 and the normalized accumulated data deviation value is obtained to determine the degree of data deviation of any monitoring point in the time window. Based on the average between the degree of data change and the degree of data deviation, a weak amplitude trend change characteristic factor for data change within a time window for any monitoring point is obtained.
[0008] Preferably, the step of forming at least two data matrices based on the multidimensional environmental parameters of any monitoring point and its reference monitoring point within the time window according to their respective dimensions includes: For any environmental parameter sequence at any monitoring point, environmental parameters belonging to the same dimension as the environmental parameter sequence at each reference monitoring point are obtained to form a reference environmental parameter sequence. The environmental parameter sequence and its reference environmental parameter sequence are combined to form a data matrix. The number of rows in the data matrix is the number of the environmental parameter sequence and its reference environmental parameter sequence, and the number of columns is the number of environmental parameters in the environmental parameter sequence.
[0009] Preferably, obtaining the data synchronization index between any monitoring point and its reference monitoring point based on the similarity characteristics of data change trends in each data matrix includes: Obtain the rank of each data matrix, accumulate the difference between the number of columns and the rank of each data matrix to obtain the data synchronization degree, and normalize the data synchronization degree to obtain the first data synchronization index. For any reference monitoring point, in any data matrix, obtain the difference sequence between the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point, and accumulate the difference data in the difference sequence to obtain the accumulated difference value between the monitoring point and the reference monitoring point in the data matrix. The cumulative difference between any monitoring point and any reference monitoring point in each data matrix is summed to obtain the total cumulative difference between any monitoring point and any reference monitoring point; The total difference between any monitoring point and each reference monitoring point is accumulated to obtain the degree of difference of any monitoring point. The degree of difference is then normalized to obtain the second data synchronization index. The data synchronization index between any monitoring point and its reference monitoring point is obtained based on the average of the first data synchronization index and the second data synchronization index.
[0010] Preferably, obtaining the degree of cross-correlation between any monitoring point and its reference monitoring point based on the data correlation characteristics in each data matrix includes: For any reference monitoring point, in any data matrix, perform cross-correlation operation on the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point to obtain the cross-correlation value. Calculate the sum of the cross-correlation value and the constant 1 to obtain the cross-correlation index between the monitoring point and the reference monitoring point in the data matrix. The cross-correlation index between any monitoring point and any reference monitoring point in each data matrix is accumulated to obtain the accumulated value of the cross-correlation index between any monitoring point and any reference monitoring point; The cumulative cross-correlation index values of any monitoring point and each reference monitoring point are summed to obtain the total cumulative cross-correlation index value between any monitoring point and its reference monitoring points. The product of the total cumulative cross-correlation index value and one-quarter is calculated to obtain the degree of data cross-correlation between any monitoring point and its reference monitoring points.
[0011] Preferably, obtaining the initial fire confidence level of any monitoring point at the current moment based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree includes: If the weak amplitude trend change characteristic factor of the data is greater than or equal to the preset weak amplitude trend change characteristic factor threshold of the data, then the mean between the data synchronization index and the data cross-correlation degree is obtained to obtain the initial fire confidence of any monitoring point at the current time. If the weak amplitude trend change characteristic factor of the data is less than the preset threshold of the weak amplitude trend change characteristic factor of the data, then the constant 0 is recorded as the initial fire confidence level of any monitoring point at the current moment.
[0012] Preferably, the step of selecting the activation function of the input gate in the LSTM prediction algorithm based on the initial fire confidence level includes: If the initial fire confidence level is greater than the preset initial fire confidence level threshold, then the Hard Sigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm. If the initial fire confidence level is less than or equal to the preset initial fire confidence level threshold, then the Sigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm.
[0013] Preferably, the automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks further includes: If the number of reference monitoring points for any monitoring point is less than a preset number, the activation function of the input gate in the LSTM prediction algorithm will not be replaced, and the LSTM prediction algorithm will be used to predict the fire at any monitoring point at the next time step.
[0014] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: This invention targets any monitoring point in a tunnel, acquiring multidimensional environmental parameters of that monitoring point at each moment; constructing a time window of a preset length including the current moment; acquiring a weak amplitude trend change feature factor of data within the time window of any monitoring point, based on the multidimensional environmental parameters of that monitoring point; acquiring a preset number of reference monitoring points for that monitoring point; forming at least two data matrices by arranging the multidimensional environmental parameters of the monitoring point and its reference monitoring points within the time window according to their respective dimensions; acquiring a data synchronization index of the monitoring point and its reference monitoring points based on the similarity characteristics of the data change trends in each data matrix; acquiring the data cross-correlation degree of the monitoring point and its reference monitoring points based on the data correlation characteristics in each data matrix; acquiring the initial fire confidence level of the monitoring point at the current moment based on the weak amplitude trend change feature factor, the data synchronization index, and the data cross-correlation degree; and selecting the activation function of the input gate in the LSTM prediction algorithm based on the initial fire confidence level to obtain an LSTM prediction model for predicting the fire at the monitoring point at the next moment. Specifically, based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the degree of data cross-correlation, the initial fire confidence level of any monitoring point at the current moment is obtained. It is then determined whether the data of any monitoring point exhibits the characteristics of the initial stage of a fire. Based on the initial fire confidence level, the activation function of the input gate in the LSTM prediction algorithm is selected to obtain the LSTM prediction model. When the data of any monitoring point exhibits the characteristics of the initial stage of a fire, more weak signals belonging to the initial stage of a fire can be written into the cell units, enabling the LSTM prediction model to identify abnormal signs in the budding stage of a fire earlier, improving the sensitivity to the initial fire signals, and thus achieving faster and more reliable automatic fire early warning. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of an automatic early warning method for high-altitude tunnel fires based on an intelligent sensor network, provided in Embodiment 1 of the present invention. Detailed Implementation
[0017] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.
[0018] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.
[0019] To illustrate the technical solution of the present invention, specific embodiments are described below.
[0020] See Figure 1 This is a flowchart of an automatic early warning method for high-altitude tunnel fires based on an intelligent sensor network, as provided in Embodiment 1 of the present invention. Figure 1 As shown, the method may include: Step S101: For any monitoring point in the tunnel, obtain the multidimensional environmental parameters of that monitoring point at each moment.
[0021] Due to the thin air, low oxygen content, and significant differences in air pressure and temperature at high altitudes, fires often rely primarily on radiative heat transfer. Heat accumulates within the enclosed, narrow tunnel spaces, and once a fire enters its growth and spread phase, the temperature rise and delayed spread are extremely rapid, drastically reducing evacuation and rescue time. Furthermore, the terrain and geological conditions at high altitudes often result in curved tunnels, causing uneven smoke flow and flame radiation. These curves can create areas prone to high-temperature accumulation and low-visibility blind spots, further exacerbating the dangers of evacuation and rescue. Therefore, early detection and identification of fires in high-altitude curved tunnels are crucial to gain valuable time for evacuation and provide a proactive approach to rescue and ventilation control.
[0022] In this embodiment, multiple sensors are deployed at each monitoring point inside the tunnel, including but not limited to temperature sensors, smoke concentration sensors, CO gas sensors, CO2 gas sensors, and wind speed and direction sensors. Environmental parameters at each monitoring point are collected in real time, and the data is preprocessed for automatic fire early warning.
[0023] Data preprocessing is an existing technology, and will be briefly described here: (1) Missing value processing: Linear interpolation or moving average method is used to fill in the missing data caused by the instantaneous communication problem of the sensor; (2) Noise filtering: Low-pass filtering is used to remove high-frequency noise caused by electromagnetic interference or environmental jitter; (3) Abnormal data processing: Outliers of environmental parameters collected by the sensor are automatically identified by box plot method and replaced with data from the previous moment; (4) Data standardization and synchronization: Environmental parameters of different types and dimensions are normalized / standardized, and multi-source data are synchronized by timestamp.
[0024] Since the method for issuing fire warnings at each monitoring point is the same, this embodiment takes any one monitoring point as an example. It acquires multi-dimensional environmental parameters at each monitoring point at any given time, including temperature, smoke concentration, CO gas concentration, CO2 gas concentration, wind speed, and wind direction, for automatic fire warning analysis at any monitoring point. In this embodiment, the acquisition frequency for each environmental parameter after synchronization is 1 Hz. This is not limited and can be set according to the specific implementation scenario.
[0025] Traditional methods typically use predictive algorithms (such as LSTM) to model sensor data, creating predictive models that forecast environmental data for future moments to detect fires early and deploy rapid response measures. However, in high-altitude environments, the low pressure in tunnels leads to thin oxygen, resulting in flame heights slightly higher than at normal pressure. Heat diffusion is slow, local temperature rise and smoke concentration changes are gradual, and fires are primarily radiation-driven. Early-stage fires exhibit a slower rate of heat release and smaller fluctuations. Because the LSTM prediction algorithm uses the sigmoid function by default for the input gate activation function, which mathematically changes exponentially with the data, the output tends to saturate when the value is close to 0 or 1. This results in small gradient changes, making it insensitive to small data fluctuations. It easily compresses the data in the input gate to near-zero values, ultimately preventing early, weak fire signals from being effectively written into the cell states. Consequently, the LSTM model fails to capture early trends effectively, resulting in sluggish output predictions that react slowly to early signals, delayed early warnings, and reduced safety.
[0026] Therefore, this embodiment analyzes the multidimensional environmental parameters in the input window of the LSTM prediction algorithm to obtain the initial fire confidence level of any monitoring point at the current moment, determines whether any monitoring point has the characteristics of the initial stage of fire, and then selects the activation function of the input gate in the LSTM prediction algorithm based on the initial fire confidence level to obtain the LSTM prediction model, thereby improving the ability to sensitively capture the initial fire signal and thus achieving faster and more reliable automatic fire early warning.
[0027] Step S102: Construct a time window of a preset length containing the current time, and obtain a weak amplitude trend change characteristic factor of data for characterizing the data change within the time window of any monitoring point based on the multidimensional environmental parameters within the time window of any monitoring point.
[0028] First, the environmental parameters in the input window of the LSTM prediction algorithm are analyzed. In this embodiment, a time window with a preset length of 90 seconds containing the current time is constructed, that is, a window containing 90 multi-dimensional environmental parameters is constructed as the input window of the LSTM prediction algorithm. There is no limitation here, and it can be set according to the specific implementation scenario.
[0029] Because the fire source in the tunnel is small in the early stage of a fire in a high-altitude tunnel, the heat released by the fire source is limited and no large-scale smoke plume is formed. Therefore, various environmental parameters show slight changes. At the same time, the thin air caused by high altitude will result in a slower combustion rate and gas convection than at low altitude, which will make the upward trend of data change even slower. Overall, it shows a weak trend that gradually increases over time.
[0030] Therefore, based on the multidimensional environmental parameters within a time window of any monitoring point, a weak amplitude trend change characteristic factor can be obtained to characterize the data changes within that time window, thus determining whether any monitoring point exhibits characteristics of the initial stage of a fire. The method for obtaining the weak amplitude trend change characteristic factor is as follows: The multidimensional environmental parameters within the time window are divided into at least two environmental parameter sequences according to their respective dimensions. The difference between the maximum and minimum values in each environmental parameter sequence is accumulated to obtain the accumulated environmental parameter difference value. The accumulated environmental parameter difference value is normalized to obtain the normalized accumulated environmental parameter difference value. The difference between the constant 1 and the normalized accumulated environmental parameter difference value is obtained to determine the degree of data change of any monitoring point within the time window. For any environmental parameter sequence, a coordinate graph of the environmental parameter sequence is constructed. The horizontal axis of the coordinate graph represents time, and the vertical axis represents the environmental parameter at each time in the environmental parameter sequence. The environmental parameter sequence is linearly fitted using the least squares method to obtain a fitted straight line. The linear fitting using the least squares method is an existing technology and will not be elaborated here. The distance between each environmental parameter in the environmental parameter sequence and the fitted straight line is obtained from the coordinate graph, and the corresponding data deviation value of the environmental parameter sequence is obtained. The data deviation values of each environmental parameter sequence are accumulated to obtain the accumulated data deviation value. The accumulated data deviation value is then normalized to obtain the normalized accumulated data deviation value. The difference between the constant 1 and the normalized accumulated data deviation value is obtained to determine the degree of data deviation of any monitoring point in the time window. Based on the average between the degree of data change and the degree of data deviation, a weak amplitude trend change characteristic factor for data change within a time window for any monitoring point is obtained.
[0031] In one implementation, the formula for calculating the weak amplitude trend change characteristic factor of the data is: Wherein, S is a weak amplitude trend change characteristic factor used to characterize the data change within a time window of any monitoring point; It is the maximum value in the j-th environmental parameter sequence; is the minimum value in the j-th environmental parameter sequence; M is the number of environmental parameter sequences, that is, the number of dimensions of the multidimensional environmental parameters; is the distance between the t-th environmental parameter in the j-th environmental parameter sequence and its fitted line; n is the number of data in the environmental parameter sequence, i.e., the length of the time window; This is the normalization function.
[0032] It should be noted that, For any monitoring point, the degree of data change within a time window, The smaller the value, the smaller the magnitude of data change in the j-th environmental parameter sequence, the weaker the trend of data change in the j-th environmental parameter sequence, and the larger S is. The degree of data deviation at any monitoring point within a time window. The smaller the value, the closer the environmental parameter in the j-th environmental parameter sequence is to its fitted straight line, the more obvious the linear trend of the j-th environmental parameter sequence, and the larger S becomes.
[0033] Thus, a weak amplitude trend change characteristic factor for data changes within a time window for any of the monitoring points is obtained.
[0034] Step S103: Obtain a preset number of reference monitoring points for any monitoring point; assemble at least two data matrices according to the dimensions of the multidimensional environmental parameters of any monitoring point and its reference monitoring points within the time window; obtain the data synchronization index between any monitoring point and its reference monitoring points based on the similarity characteristics of the data change trends in each data matrix; and obtain the degree of data cross-correlation between any monitoring point and its reference monitoring points based on the data correlation characteristics in each data matrix.
[0035] In step S102, a weak amplitude trend change characteristic factor of data is obtained to characterize the data change within the time window of any monitoring point. If the weak amplitude trend change characteristic factor of data is large, it is considered that the data change of the current monitoring point may be caused by the initial stage of the fire. However, other factors (such as vehicle accidents, sensor abnormalities, etc.) may also cause the weak amplitude trend change characteristic factor of data at any monitoring point to be large. Therefore, it is necessary to further determine whether the weak amplitude trend change characteristic factor of data at any monitoring point is large due to the initial stage of the fire.
[0036] When a fire occurs, the heat and smoke from combustion first accumulate in the local space around the fire source. The nearest monitoring point will be the first to receive this change. Secondly, due to airflow guidance and smoke convection, the hot smoke generated by the fire will spread downstream in the tunnel, causing a delay in the response of data from nearby monitoring points. Therefore, the next lower monitoring point is obtained as a reference monitoring point to analyze whether any monitoring point exhibits characteristics of the initial stage of a fire. Since the most common monitoring point interval in high-altitude tunnels is 30m, the preset number of reference monitoring points in this embodiment is set to 2. This is not a limitation and can be set according to the specific implementation scenario.
[0037] Specifically, if the number of reference monitoring points for any monitoring point is less than 2, the activation function of the input gate in the LSTM prediction algorithm is not replaced, and the LSTM prediction algorithm is used to predict the fire at any monitoring point at the next time step.
[0038] The weak radiant heat and smoke generated by the early fire will first affect a group of monitoring point sensors locally, and then affect the sensors of the next monitoring point along the wind direction, showing similar weak changes. This weak synchronous change leads to an increase in the correlation between various environmental parameters. At the same time, the data of the sensor closest to the fire source will change first, while the data of the neighboring sensor will change relatively later.
[0039] Therefore, for any environmental parameter sequence at any monitoring point, environmental parameters belonging to the same dimension as the environmental parameter sequence at each reference monitoring point are obtained to form a reference environmental parameter sequence. The environmental parameter sequence and its reference environmental parameter sequences are then combined to form a data matrix. The number of rows in the data matrix is equal to the number of rows in the environmental parameter sequence and its reference environmental parameter sequence, and the number of columns is equal to the number of environmental parameters in the environmental parameter sequence. Similarly, at least two data matrices are obtained. Then, based on the similarity of the data change trends in each data matrix, a data synchronization index between any monitoring point and its reference monitoring point is obtained to determine whether the weak amplitude trend change characteristic factor of the data at any monitoring point is large due to the initial stage of a fire.
[0040] The method for obtaining the data synchronization index between any monitoring point and its reference monitoring point based on the similarity characteristics of the data change trends in each data matrix is as follows: Obtain the rank of each data matrix, accumulate the difference between the number of columns and the rank of each data matrix to obtain the data synchronization degree, and normalize the data synchronization degree to obtain the first data synchronization index. For any reference monitoring point, in any data matrix, obtain the difference sequence between the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point, and accumulate the difference data in the difference sequence to obtain the accumulated difference value between the monitoring point and the reference monitoring point in the data matrix. The cumulative difference between any monitoring point and any reference monitoring point in each data matrix is summed to obtain the total cumulative difference between any monitoring point and any reference monitoring point; The total difference between any monitoring point and each reference monitoring point is accumulated to obtain the degree of difference of any monitoring point. The degree of difference is then normalized to obtain the second data synchronization index. The data synchronization index between any monitoring point and its reference monitoring point is obtained based on the average of the first data synchronization index and the second data synchronization index.
[0041] In one embodiment, the formula for calculating the data synchronization index between any monitoring point and its reference monitoring point is: Where P is the data synchronization index between any monitoring point and its reference monitoring point; Let be the number of columns in the j-th data matrix; Let M be the rank of the j-th data matrix; M is the number of data matrices, which is also the number of dimensions of the multidimensional environment parameters. For any monitoring point, let t be the environmental parameter of the j-th environmental parameter sequence. Let t be the t-th environmental parameter in the t-th reference environmental parameter sequence of the ith reference monitoring point; n is the number of data in the environmental parameter sequence, which is also the length of the time window; The number of monitoring points is for reference. This is the normalization function.
[0042] It should be noted that, The first data synchronization index, The larger the value, the smaller the column rank of the j-th data matrix, and the higher the linear correlation between the columns of the j-th data matrix. This means that the data from any monitoring point and its reference monitoring point show relatively synchronous changes within the same time window. The larger the value, the larger P becomes; This is the second data synchronization index. Because data changes caused by a fire have a time delay, data from the same type of sensor at different monitoring points can vary significantly at the same time. The larger the value, the greater the difference between the data from any monitoring point and its i-th reference monitoring point in the j-th data matrix (the j-th type of sensor). The larger it is, the larger P will be.
[0043] Furthermore, since data changes caused by fire propagate downstream due to wind direction, there is a certain time delay between any monitoring point and its nearest reference monitoring point. Therefore, the cross-correlation degree between any monitoring point and its reference monitoring point (i.e., the cross-correlation degree between any monitoring point and its reference monitoring point under a certain delay time) can be obtained based on the data correlation characteristics in each data matrix. In this embodiment, the time delay between the current monitoring point and its nearest reference monitoring point is obtained through FDS (Fire Dynamics Simulator, a common fire dynamics numerical simulation software in fire early warning, which can accurately simulate the propagation of heat, smoke, and airflow generated by a fire, thereby calculating the sensor response time at any location). This time delay is used as the theoretical time delay T, i.e., the theoretical time delay between any monitoring point and its nearest reference monitoring point. FDS is existing technology and will not be elaborated here.
[0044] The method for obtaining the cross-correlation degree between any monitoring point and its reference monitoring point based on the data correlation characteristics in each data matrix is as follows: For any reference monitoring point, in any data matrix, a cross-correlation operation is performed between the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point to obtain a cross-correlation value. The cross-correlation operation is an existing technology and will not be described in detail here. The cross-correlation value is calculated and added to a constant 1 to obtain the cross-correlation index between the monitoring point and the reference monitoring point in the data matrix. The cross-correlation index between any monitoring point and any reference monitoring point in each data matrix is accumulated to obtain the accumulated value of the cross-correlation index between any monitoring point and any reference monitoring point; The cumulative cross-correlation index values of any monitoring point and each reference monitoring point are summed to obtain the total cumulative cross-correlation index value between any monitoring point and its reference monitoring points. The product of the total cumulative cross-correlation index value and one-quarter is calculated to obtain the degree of data cross-correlation between any monitoring point and its reference monitoring points.
[0045] In one embodiment, the formula for calculating the cross-correlation between the data from any monitoring point and its reference monitoring point is as follows: Where Q represents the degree of cross-correlation between the data of any monitoring point and its reference monitoring point; M represents the number of reference monitoring points; M represents the number of data matrices. The cross-correlation value is obtained by performing a cross-correlation operation on the environmental parameter sequence of any monitoring point in the j-th data matrix and the reference environmental parameter sequence of the i-th reference monitoring point in the j-th data matrix. Used to limit the calculation results to 0 to 1.
[0046] It should be noted that, Let be the cross-correlation index between any monitoring point and the i-th reference monitoring point in the j-th data matrix. The larger the value, the more consistent the subtle changes in the data at any monitoring point are with the subtle changes in the data at the i-th reference monitoring point over a certain delay time, and the larger the value of Q.
[0047] Thus, we obtain the data synchronization index between any monitoring point and its reference monitoring point, as well as the data cross-correlation degree between any monitoring point and its reference monitoring point.
[0048] Step S104: Based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree, obtain the initial fire confidence level of any monitoring point at the current time. Based on the initial fire confidence level, select the activation function of the input gate in the LSTM prediction algorithm to obtain the LSTM prediction model, which is used to predict the fire at any monitoring point at the next time.
[0049] After obtaining the data synchronization index between any monitoring point and its reference monitoring point, the data cross-correlation degree between any monitoring point and its reference monitoring point, and the weak amplitude trend change characteristic factor used to characterize the data change within the time window of any monitoring point, the initial fire confidence level of any monitoring point at the current moment is obtained based on the weak amplitude trend change characteristic factor, the data synchronization index, and the data cross-correlation degree. It is then determined whether any monitoring point exhibits the characteristics of the initial stage of a fire. Based on the initial fire confidence level, the activation function of the input gate in the LSTM prediction algorithm is selected to obtain the LSTM prediction model, which is used to predict the fire at any monitoring point at the next moment.
[0050] The method for obtaining the initial fire confidence level of any monitoring point at the current moment based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree is as follows: If the weak amplitude trend change characteristic factor of the data is greater than or equal to the preset weak amplitude trend change characteristic factor threshold, it is considered that the data change at the current monitoring point may be due to the initial stage of the fire. The mean between the data synchronization index and the data cross-correlation degree is obtained to obtain the initial fire confidence level of any monitoring point at the current moment. If the weak amplitude trend change characteristic factor of the data is less than the preset weak amplitude trend change characteristic factor threshold, it is considered that the data change at the current monitoring point is not caused by the initial stage of the fire. Therefore, the constant 0 is recorded as the initial fire confidence level of any monitoring point at the current moment.
[0051] In one embodiment, the formula for calculating the initial fire confidence level of any monitoring point at the current moment is: Where F is the initial fire confidence level of any monitoring point at the current moment; P is the data synchronization index between any monitoring point and its reference monitoring point; Q is the data cross-correlation degree between any monitoring point and its reference monitoring point; S is the weak amplitude trend change characteristic factor of data used to characterize the data change within the time window of any monitoring point; a is the preset threshold of the weak amplitude trend change characteristic factor of data. Since setting the preset threshold of the weak amplitude trend change characteristic factor of data too large will lead to overly strict identification of the initial fire and cause some initial manifestations to be not identified in time, while setting it too small will lead to overly sensitive data changes caused by some environmental interference, resulting in false identification, a is set to 0.75 in this embodiment. There is no limitation here, and it can be set according to the specific implementation scenario.
[0052] It should be noted that when When P and Q are larger, it means that the data changes of any monitoring point are more consistent with the characteristics of the early stage of a fire, and F is larger.
[0053] Furthermore, if the initial fire confidence level is greater than a preset initial fire confidence threshold, the HardSigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm; if the initial fire confidence level is less than or equal to the preset initial fire confidence threshold, the Sigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm. Then, based on the activation function of the input gate in the LSTM prediction algorithm, an LSTM prediction model is obtained to predict the fire at any monitoring point at the next moment. Since setting the preset initial fire confidence threshold too high may ignore the characteristics of the initial fire phase, resulting in weak signals from the actual initial fire phase not being written into the cell unit, while setting it too low may cause the LSTM prediction model to be overly sensitive, amplifying and capturing minor interferences, leading to false alarms, the preset initial fire confidence threshold is set to 0.7 in this embodiment. This is not a limitation and can be set according to the specific implementation scenario.
[0054] It is worth noting that the main purpose of this scheme is to select the activation function of the input gate in the LSTM prediction algorithm. Based on the activation function of the input gate in the LSTM prediction algorithm, the LSTM prediction model is obtained. Predicting the fire at any monitoring point at the next moment is an existing technology and will not be elaborated here.
[0055] In summary, this embodiment of the invention acquires multidimensional environmental parameters of any monitoring point in a tunnel at each moment; constructs a time window of a preset length including the current moment; acquires a weak amplitude trend change feature factor to characterize the data changes within the time window of any monitoring point based on the multidimensional environmental parameters within the time window; acquires a preset number of reference monitoring points for any monitoring point; combines the multidimensional environmental parameters of any monitoring point and its reference monitoring points within the time window into at least two data matrices according to their respective dimensions; acquires a data synchronization index between any monitoring point and its reference monitoring points based on the similarity characteristics of the data change trends in each data matrix; acquires the data cross-correlation degree between any monitoring point and its reference monitoring points based on the data correlation characteristics in each data matrix; acquires the initial fire confidence level of any monitoring point at the current moment based on the weak amplitude trend change feature factor, the data synchronization index, and the data cross-correlation degree; and selects the activation function of the input gate in the LSTM prediction algorithm based on the initial fire confidence level to obtain an LSTM prediction model for predicting the fire at any monitoring point at the next moment. Specifically, based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the degree of data cross-correlation, the initial fire confidence level of any monitoring point at the current moment is obtained. It is then determined whether the data of any monitoring point exhibits the characteristics of the initial stage of a fire. Based on the initial fire confidence level, the activation function of the input gate in the LSTM prediction algorithm is selected to obtain the LSTM prediction model. When the data of any monitoring point exhibits the characteristics of the initial stage of a fire, more weak signals belonging to the initial stage of a fire can be written into the cell units, enabling the LSTM prediction model to identify abnormal signs in the budding stage of a fire earlier, improving the sensitivity to the initial fire signals, and thus achieving faster and more reliable automatic fire early warning.
[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. An automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks, characterized in that, The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks includes: For any monitoring point in the tunnel, obtain the multidimensional environmental parameters of that monitoring point at each moment; Construct a time window of a preset length containing the current moment, and obtain a weak amplitude trend change characteristic factor of data change within the time window of any monitoring point based on the multidimensional environmental parameters within the time window of any monitoring point. A preset number of reference monitoring points are obtained for any given monitoring point. The multidimensional environmental parameters of any given monitoring point and its reference monitoring points within the time window are arranged into at least two data matrices according to their respective dimensions. Based on the similarity characteristics of the data change trends in each data matrix, the data synchronization index between any given monitoring point and its reference monitoring points is obtained. Based on the data correlation characteristics in each data matrix, the degree of data cross-correlation between any given monitoring point and its reference monitoring points is obtained. Based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree, the initial fire confidence level of any monitoring point at the current time is obtained. Based on the initial fire confidence level, the activation function of the input gate in the LSTM prediction algorithm is selected to obtain the LSTM prediction model, which is used to predict the fire at any monitoring point at the next time.
2. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 1, characterized in that, The step of obtaining a weak amplitude trend change characteristic factor for data changes within the time window of any monitoring point based on multidimensional environmental parameters within that time window includes: The multidimensional environmental parameters within the time window are divided into at least two environmental parameter sequences according to their respective dimensions. The difference between the maximum and minimum values in each environmental parameter sequence is accumulated to obtain the accumulated environmental parameter difference value. The accumulated environmental parameter difference value is normalized to obtain the normalized accumulated environmental parameter difference value. The difference between the constant 1 and the normalized accumulated environmental parameter difference value is obtained to determine the degree of data change of any monitoring point within the time window. For any environmental parameter sequence, a coordinate graph of the environmental parameter sequence is constructed. The horizontal axis of the coordinate graph is time, and the vertical axis is the environmental parameter at each time in the environmental parameter sequence. The environmental parameter sequence is fitted to obtain a fitted straight line. The distance between each environmental parameter in the environmental parameter sequence and the fitted straight line is obtained in the coordinate graph, and the data deviation value of the environmental parameter sequence is obtained accordingly. The data deviation values of each environmental parameter sequence are accumulated to obtain the accumulated data deviation value. The accumulated data deviation value is then normalized to obtain the normalized accumulated data deviation value. The difference between the constant 1 and the normalized accumulated data deviation value is obtained to determine the degree of data deviation of any monitoring point in the time window. Based on the average between the degree of data change and the degree of data deviation, a weak amplitude trend change characteristic factor for data change within a time window for any monitoring point is obtained.
3. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 2, characterized in that, The step of forming at least two data matrices based on the dimensions of the multidimensional environmental parameters of any monitoring point and its reference monitoring point within the time window includes: For any environmental parameter sequence at any monitoring point, environmental parameters belonging to the same dimension as the environmental parameter sequence at each reference monitoring point are obtained to form a reference environmental parameter sequence. The environmental parameter sequence and its reference environmental parameter sequence are combined to form a data matrix. The number of rows in the data matrix is the number of the environmental parameter sequence and its reference environmental parameter sequence, and the number of columns is the number of environmental parameters in the environmental parameter sequence.
4. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 3, characterized in that, The step of obtaining the data synchronization index between any monitoring point and its reference monitoring point based on the similarity characteristics of the data change trends in each data matrix includes: Obtain the rank of each data matrix, accumulate the difference between the number of columns and the rank of each data matrix to obtain the data synchronization degree, and normalize the data synchronization degree to obtain the first data synchronization index. For any reference monitoring point, in any data matrix, obtain the difference sequence between the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point, and accumulate the difference data in the difference sequence to obtain the accumulated difference value between the monitoring point and the reference monitoring point in the data matrix. The cumulative difference between any monitoring point and any reference monitoring point in each data matrix is summed to obtain the total cumulative difference between any monitoring point and any reference monitoring point; The total difference between any monitoring point and each reference monitoring point is accumulated to obtain the degree of difference of any monitoring point. The degree of difference is then normalized to obtain the second data synchronization index. The data synchronization index between any monitoring point and its reference monitoring point is obtained based on the average of the first data synchronization index and the second data synchronization index.
5. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 3, characterized in that, The step of obtaining the cross-correlation degree between any monitoring point and its reference monitoring point based on the data correlation characteristics in each data matrix includes: For any reference monitoring point, in any data matrix, perform cross-correlation operation on the environmental parameter sequence of the monitoring point and the reference environmental parameter sequence of the reference monitoring point to obtain the cross-correlation value. Calculate the sum of the cross-correlation value and the constant 1 to obtain the cross-correlation index between the monitoring point and the reference monitoring point in the data matrix. The cross-correlation index between any monitoring point and any reference monitoring point in each data matrix is accumulated to obtain the accumulated value of the cross-correlation index between any monitoring point and any reference monitoring point; The cumulative cross-correlation index values of any monitoring point and each reference monitoring point are summed to obtain the total cumulative cross-correlation index value between any monitoring point and its reference monitoring points. The product of the total cumulative cross-correlation index value and one-quarter is calculated to obtain the degree of data cross-correlation between any monitoring point and its reference monitoring points.
6. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 1, characterized in that, The step of obtaining the initial fire confidence level of any monitoring point at the current moment based on the weak amplitude trend change characteristic factor of the data, the data synchronization index, and the data cross-correlation degree includes: If the weak amplitude trend change characteristic factor of the data is greater than or equal to the preset weak amplitude trend change characteristic factor threshold of the data, then the mean between the data synchronization index and the data cross-correlation degree is obtained to obtain the initial fire confidence of any monitoring point at the current time. If the weak amplitude trend change characteristic factor of the data is less than the preset threshold of the weak amplitude trend change characteristic factor of the data, then the constant 0 is recorded as the initial fire confidence level of any monitoring point at the current moment.
7. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 1, characterized in that, The step of selecting the activation function for the input gate in the LSTM prediction algorithm based on the initial fire confidence level includes: If the initial fire confidence level is greater than the preset initial fire confidence level threshold, then the Hard Sigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm. If the initial fire confidence level is less than or equal to the preset initial fire confidence level threshold, then the Sigmoid function is selected as the activation function of the input gate in the LSTM prediction algorithm.
8. The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks according to claim 1, characterized in that, The automatic early warning method for high-altitude tunnel fires based on intelligent sensor networks also includes: If the number of reference monitoring points for any monitoring point is less than a preset number, the activation function of the input gate in the LSTM prediction algorithm will not be replaced, and the LSTM prediction algorithm will be used to predict the fire at any monitoring point at the next time step.