CN103116961B - A kind of confined space fire detection alarm system based on Electronic Nose Technology and method - Google Patents

Abstract

Based on confined space fire detection alarm system and the method for Electronic Nose Technology, comprising: sampling pipe, filtration unit, three gas sensors, temperature sensor, sensor power unit, sensor control unit, signal gathering unit, signal condition unit, monitoring host computer, acoustic-optic alarm, flowmeter, vacuum pump, three-way solenoid valve and exhaust gas processing devices.The present invention can on-line analysis, study, intelligent decision, so that at pole early detection fire, and reduces wrong report, rate of failing to report, reduces the infringement of fire to confined space personnel and equipment to a greater extent.

Description

A fire detection and alarm system and method for confined spaces based on electronic nose technology

technical field

The invention belongs to the technical field of fire detection, and in particular relates to a fire detection and alarm system and method in a confined space based on electronic nose technology.

Background technique

Confined space refers to an unconventional limited space with restricted entrance and exit, poor natural ventilation, and relative isolation from the outside world. Common confined spaces mainly include some power distribution cabinets, machine rooms, material warehouses, aircraft cargo compartments and space capsules. Once a fire occurs in such a confined space, the generated smoke, poisonous gas and heat will inevitably accumulate in a large amount in a short period of time, causing great damage to personnel and equipment. Therefore, early fire detection and alarm is particularly important.

At present, in the confined space, the alarm is mainly aimed at the change of the visible smoke concentration during the fire, and it is relatively easy to be interfered by dust and water vapor in the air to cause false alarms and false alarms. Fire signature Gases are produced very early in a fire, prior to the appearance of visible smoke. Moreover, the concentration changes of characteristic gases such as CO and _CO2 produced by real fires are very regular, which is conducive to detection. However, the gas sensor is selective (that is, cross-sensitive to multiple gases), and the use of a single sensor for fire characteristic gas detection and alarm is easily affected by other gases or environmental factors, causing false alarms. The use of electronic nose technology can solve this problem well. Electronic nose technology mainly consists of three parts: gas sensor array, signal preprocessing and pattern recognition. The pattern recognition algorithms currently used in the electronic nose technology in the field of fire detection mainly include BP neural network, support vector machine (SVM) and so on. This type of algorithm takes a long time to train and has a complex structure, which is not conducive to updating the network model online, resulting in false positives and false negatives.

Contents of the invention

The technical solution of the present invention is to overcome the deficiencies of the prior art, and provide a confined space fire detection and alarm system and method based on electronic nose technology, which extracts the change rules of fire characteristic gases O ₂ , CO, and CO ₂ , and conducts real-time online analysis and learning. Intelligent judgment, so as to detect fire at an early stage, reduce false alarm and false alarm rate, and minimize fire damage to personnel and equipment in confined spaces.

Technical solution of the present invention: a confined space fire detection and alarm system based on electronic nose technology, its structure is shown in Figure 1, it is controlled by a sampling tube, a filter device, three gas sensors, a temperature sensor, a sensor power supply unit, and a sensor Unit, signal acquisition unit, signal conditioning unit, monitoring host, sound and light alarm device, flow meter, vacuum pump, three-way solenoid valve, exhaust gas treatment device, etc.

The gas sensor collects O ₂ , CO, CO ₂ and other signals of the monitored site through the sampling tube, and after data preprocessing, input the trained PNN network model to predict the probability of fire occurrence U _i . If the fire occurrence probability is greater than the given threshold U _d1 , start recording the warning time t(i), otherwise, input the data into the online training program of the PNN network model. Make a second judgment, if the fire occurrence probability is greater than or equal to the given threshold value U _d2 , then directly turn on the sound and light alarm device and connect the exhaust gas treatment device; otherwise, judge whether the early warning time t(i) is greater than or equal to the given threshold value t _d : Yes, turn on the sound and light alarm device and connect the exhaust gas treatment device; if not, input the data into the online training program of the PNN network model. Fig. 2 is a flow chart of the real-time fire identification algorithm in the present invention.

Based on the electronic nose technology, the present invention uses three cross-sensitive gas sensors for O ₂ , CO, and CO ₂ and a temperature sensor to form a sensor array to collect data, and uses a probabilistic neural network (PNN) as its pattern recognition algorithm. In order to reduce the complexity of the model and the difficulty of training, a filter device is added at the end of the sampling tube and a flow meter is used to control the gas flow rate through the sensor. The filter device has two purposes: (1) to filter the smoke particles in the sampled gas and prolong the life of the sensor; (2) to dry the sampled gas to reduce the influence of water vapor on the gas sensor. When determining the optimal gas flow rate, multiple groups of mixed gas samples of O ₂ , CO, and CO ₂ can be configured, the flow rate can be changed, and the gas sensor response curves at multiple flow rates can be obtained for judgment.

The present invention uses a probabilistic neural network model (PNN) for pattern recognition. PNN is based on Bayesian classification theory decision-making, and uses the Parzen window method to estimate the probability density function, which has many excellent properties: (1) PNN training speed is fast, which is conducive to real-time application; (2) Decision-making can achieve Bayesian optimality; ( 3) Strong fault tolerance; (4) When adding new training samples, it is not necessary to retrain the network.

Before using the PNN network model, it needs to be trained offline. The offline training process is shown in Figure 3. Firstly, the on-site fire hazards in the confined space are investigated, and a simulated fire experiment is carried out based on this. Obtain simulated fire site data through gas sensors and temperature sensors. Since the present invention is devoted to real-time monitoring, only gas sensor response value S _i , gas sensor change rate ΔS _i and temperature response value T _i can be selected as corresponding features for preprocessing. The data preprocessing method also directly affects the working characteristics of the system. Using array normalization can achieve good results, and the calculation formula is:

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; /</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, Xi _,j represents the jth eigenvalue in the i-th measurement, Represents the transformed value. normalized data Carry out PCA principal component analysis, through dimension reduction, find N (N<j) orthogonal characteristic variables, that is, principal components, so that they can reflect the data The main feature of the algorithm is to compress the size of the original data matrix. When the contribution rate of the first N principal components reaches 90%, it is considered that the N principal components can reflect The main features of are used to fit the original data. A new sample set X={ _xi,n |n=1,2,...,N} is formed by using the first N principal components, ie feature parameters, and applied to the training of the new PNN network model. When the output is expected to meet the requirements, the training is terminated, and the formed PNN network model is input to the monitoring host.

The present invention can write the obtained data into the PNN network model training sample and retrain it while real-time on-line monitoring, thereby optimizing the PNN network model. The online training process is shown in Figure 4. After the monitoring site data is judged by fire, it is sent to the online training program. For online programs, when the system is stable for a long time, the data written into the training samples will tend to be the same, which wastes storage space and reduces the normalization ability of the model. Therefore, it is necessary to judge whether the system is in a steady state. Set the threshold θ, when it is satisfied:

‖X(i+1)-X(i)‖<θ

It is considered that the system has entered a steady state, so the online training is stopped and the PNN network model remains unchanged. When the system is not in a steady state, the output of the fire monitoring main program is used as the expected output of the sample, which is added to the training sample set of the PNN network model for online training until the expected output of the test sample meets the conditions. The obtained new PNN network model is rewritten into the monitoring host for subsequent monitoring.

When the fire occurrence probability is around 0.5, it is unreasonable to directly judge whether a fire has occurred. The invention introduces the early warning time to intelligently judge the fire probability, and increases the reliability of the alarm. The warning time is defined as follows:

$\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\end{array}\right.$

u(·) represents a unit step signal, i∈[0,T). T is the cycle time, when i=T, reset the warning time. When the fire occurrence probability U _i is greater than U _d1 , start the warning time timer t(i). If U _i is greater than U _d2 , start the sound and light alarm directly, connect the exhaust gas treatment device and reset the warning time t(i); otherwise, further judgment is required. Determine whether the warning time t(i) is greater than or equal to the given threshold t _d , if so, start the alarm, record and reset the warning time t(i); if not, input the data into the online training program and reset the warning time t( i).

A fire detection and alarm method in a confined space based on electronic nose technology, characterized in that the implementation steps are as follows:

(1) First conduct fire accident investigation in confined spaces, find out the main fire risks, and conduct simulation experiments for these main risks.

(2) According to the data obtained from the simulation experiment, the PNN network model is trained offline. After normalizing the data, perform PCA principal component analysis and apply it to the training of the new PNN network model. When the output is expected to meet the requirements, the training is terminated, and the formed PNN network model is input to the monitoring host.

(3) Write the trained PNN network model into the monitoring host. Place the system sampling tube in the confined space to be monitored, connect the three-way solenoid valve to the air, turn on the sensor power supply unit and control unit, turn on the vacuum pump to start pumping air, and start the monitoring host. The data collected from the monitoring site is used to obtain the probability of fire occurrence through the PNN network model, and combined with the early warning time judgment mechanism for real-time fire identification.

(4) Use the real-time data obtained from the monitoring site to train the PNN network model online. Firstly, it is judged whether the system is in a steady state. When the system is in a steady state, stop online training and keep the PNN network model unchanged; when the system is not in a steady state, add data to the training sample set of the PNN network model for online training to obtain a new PNN network model, and Rewrite the monitoring host for subsequent monitoring.

The advantage of the present invention compared with prior art is:

(1) The present invention utilizes the electronic nose technology to deal with the gas concentration changes before the smoke in the fire, and the alarm time is earlier than the conventional smoke detectors and video smoke detectors routinely used in confined spaces, and reduces the rate of false alarms and missed alarms.

(2) The present invention adopts a PNN network model with a stronger fault tolerance rate and is more conducive to online training, and introduces steady-state judgment to realize real-time monitoring and online update of fires in confined spaces.

(3) The present invention introduces an early warning time judgment mechanism to make the method more intelligent and increase the reliability of the alarm.

Description of drawings

Fig. 1 is the structure schematic diagram of the fire detection and alarm system in confined space based on the electronic nose technology of the present invention;

Fig. 2 is the flow chart of real-time fire recognition algorithm among the present invention;

Fig. 3 is the flow chart of off-line training of probabilistic neural network (PNN) in the present invention;

Fig. 4 is a flow chart of the online training of the probabilistic neural network (PNN) in the present invention.

Detailed ways

As shown in Figure 1, the fire detection and alarm system for confined spaces based on the electronic nose technology of the present invention includes: a sampling tube, a filtering device, three gas sensors, a temperature sensor, a sensor power supply unit, a sensor control unit, a signal acquisition unit, and a signal conditioning unit , monitoring host, sound and light alarm device, flow meter, vacuum pump, three-way solenoid valve and exhaust gas treatment device; add a filter device at the end of the sampling pipe and use a flow meter to control the gas flow rate through the three gas sensors.

The sensor power supply unit provides power for the three gas sensors and temperature sensors, and its output terminals are connected to the sensors; the sensor control unit controls the opening and closing of the three gas sensors and temperature sensors, and its output terminals are connected to the sensors; the three sensors are respectively _O2 sensors , CO sensor and _CO2 sensor; the signal acquisition unit is used to collect the data of the three gas sensors and the temperature sensor, the input end is connected to the sensor, and the output end is connected to the signal conditioning unit; the signal conditioning unit can amplify the collected signal , filtering, its input end is connected with the signal acquisition unit, and its output end is connected with the monitoring host; the monitoring host is equipped with a fire monitoring main program and a PNN neural network online training algorithm, which can identify fires in real time and give an alarm, online update PNN neural network and The data is stored, and its output end is connected with an audible and visual alarm device and a three-way solenoid valve.

The sampling tube collects and transmits on-site gas data in a confined space, one end of which is placed in the confined space, and the other end is connected to a filter device; the filter device filters water vapor and smoke particles in the gas to reduce the impact on the sensor, one end of which is connected to the sampling tube, and the other end Connect the gas analysis chamber through the pipeline; the gas analysis chamber is used to place three gas sensors and temperature sensors, the other end of which is connected to the flowmeter through the pipeline; the flowmeter controls the flow rate in the pipeline, and the other end is connected to the vacuum pump through the pipeline; the vacuum pump pumps air , to provide power for system sampling, the other end of which is connected to the three-way solenoid valve through the pipeline; the three-way solenoid valve controls the flow direction of the exhaust gas, and the other end is connected to the exhaust gas treatment device through the pipeline; the exhaust gas treatment device can process CO and CO ₂ in the sampled gas and other harmful gases to prevent pollution.

Under normal conditions, the three-way solenoid valve is connected to the air, the flow meter controls the pumping speed of the vacuum pump, and the gas in the confined space passes through the sampling tube and enters the gas analysis chamber through the filter device. Three gas sensors and one temperature sensor are placed in the gas analysis chamber, and the power supply unit and control unit ensure their normal operation. The signals collected by the four sensors are processed by the signal acquisition unit and the signal conditioning unit, and then sent to the monitoring host for analysis. When a fire occurs, the solenoid valve is connected to the exhaust gas treatment device, and the monitoring host controls the sound and light alarm device to give an alarm.

The method of the present invention is applied to some equipment rooms, and the steps are as follows:

(1) First conduct a fire accident investigation on this type of confined space, and find that the main fire risk is the overheating or short circuit of non-metallic materials such as cables and circuit boards. Therefore, simulation experiments can be carried out for these situations.

(2) Offline training of the PNN network model, as shown in Figure 3. Obtain simulated fire site data through gas sensors and temperature sensors. Select gas sensor response value S _i , gas sensor change rate ΔS _i and temperature response value T _i as corresponding features for preprocessing. Using array normalization can achieve good results, and the calculation formula is:

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; /</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 7</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$

Among them, Xi _,j represents the jth eigenvalue in the i-th measurement, Represents the transformed value. normalized data Perform PCA principal component analysis. The contribution rate of the first three principal components reaches 90%, which can be used to fit the original data. A new sample set X={ _xi,n |n=1,2,3} is formed by using the first three principal components, ie feature parameters, and applied to the training of the new PNN network model. When the output is expected to meet the requirements, the training is terminated, and the formed PNN network model is input to the monitoring host.

(3) The real-time fire identification algorithm in the present invention, as shown in Figure 2, the gas sensor collects _O2 , CO, _CO2 and other signals of the monitored site through the sampling tube, and inputs the trained PNN network model after data preprocessing , to predict the fire occurrence probability U _i . If the probability of fire occurrence is greater than the given threshold U _d1 =0.4, start recording the warning time t(i), otherwise, input the data into the online training program of the PNN network model. Make a second judgment, if the probability of fire occurrence is greater than or equal to the given threshold value U _d2 =0.6, then directly turn on the sound and light alarm device and connect the exhaust gas treatment device; otherwise, judge whether the early warning time t(i) is greater than or equal to the given valve value Value t _d =15: Yes, turn on the sound and light alarm device, connect the exhaust gas treatment device, and input the data into the online training program of the PNN network model; if not, input the data into the online training program of the PNN network model.

The warning time is defined as follows:

$\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\end{array}\right.$

u(·) represents a unit step signal, i∈[0,T). T=20 is the cycle time, when i=T, reset the early warning time. Taking p=5 means that if four consecutive U _i are less than or equal to U _d1 , the warning time t(i) will be reset.

(4) Online training of the PNN network model, as shown in Figure 4. After being judged by the fire monitoring main program, the monitoring site data is sent to the online training program. When judging whether the system is in a steady state, set the threshold θ=0.01, when:

‖X(i+1)-X(i)‖<θ

It is considered that the system has entered a steady state, so the online training is stopped and the PNN network model remains unchanged. When the system is not in a steady state, the output of the fire monitoring main program is used as the expected output of the sample, and is added to the training sample set of the PNN network model for online training until the expected output of the test sample meets the conditions. The obtained new PNN network model is rewritten into the monitoring host for subsequent monitoring.

In a word, the present invention utilizes the electronic nose technology to deal with the gas concentration change during the fire, the alarm time is shorter, and the rate of false alarm and false alarm is reduced; the PNN network model is adopted, and steady-state judgment is introduced to realize real-time monitoring and online update; the introduction of early warning The time judgment mechanism is more intelligent and reliable. The invention can be applied to fire detection and alarm in confined spaces such as power distribution cabinets, machine rooms, material warehouses, aircraft cargo compartments and aerospace cabins.

The description of the present invention does not describe in detail the part that belongs to the known technology of those skilled in the art.

The above is only a specific implementation mode in the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technology can understand the conceivable transformation or replacement within the technical scope disclosed in the present invention. All should be covered within the scope of the present invention, therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (4)

1. A confined space fire detection and alarm system based on electronic nose technology, characterized in that it comprises: a sampling tube, a filtering device, three gas sensors, a temperature sensor, a sensor power supply unit, a sensor control unit, a signal acquisition unit, and a signal conditioning unit , monitoring host, sound and light alarm device, flow meter, vacuum pump, three-way solenoid valve and exhaust gas treatment device; The sensor power supply unit provides power for the three gas sensors and temperature sensors, and its output terminals are connected to the sensors; the sensor control unit controls the opening and closing of the three gas sensors and temperature sensors, and its output terminals are connected to the sensors; the three sensors are respectively _O2 sensors , CO sensor and _CO sensor; the signal acquisition unit is used to collect the data of the three gas sensors and the temperature sensor, its input end is connected to the sensor, and the output end is connected to the signal conditioning unit; the signal conditioning unit amplifies the collected signal, Filtering, its input end is connected to the signal acquisition unit, and its output end is connected to the monitoring host; the monitoring host is equipped with a fire monitoring main program and a PNN neural network online training algorithm, which can identify fires in real time and give an alarm, online update PNN neural network and storage data, its output end is connected to the sound and light alarm device and the three-way solenoid valve; The sampling tube collects and transmits on-site gas data in a confined space, one end of which is placed in the confined space, and the other end is connected to a filter device; the filter device filters water vapor and smoke particles in the gas to reduce the impact on the sensor, one end of which is connected to the sampling tube, and the other end Connect the gas analysis chamber through the pipeline; the gas analysis chamber is used to place three gas sensors and temperature sensors, the other end of which is connected to the flowmeter through the pipeline; the flowmeter controls the flow rate in the pipeline, and the other end is connected to the vacuum pump through the pipeline; the vacuum pump pumps air , to provide power for system sampling, the other end of which is connected to the three-way solenoid valve through the pipeline; the three-way solenoid valve controls the flow direction of the exhaust gas, and the other end is connected to the exhaust gas treatment device through the pipeline; the exhaust gas treatment device can process CO and CO ₂ in the sampled gas Harmful gas, prevent pollution; The realization process of the monitoring host: the O ₂ , CO, CO ₂ and temperature signals of the monitored site are input into the trained PNN neural network model after data preprocessing to predict the probability of fire occurrence and judge the early warning time. If it is judged that a fire occurs, Then turn on the sound and light alarm device, connect the exhaust gas treatment device, and input the data into the PNN network model online training program that has been trained; if it is judged that there is no fire, then continue monitoring, and input the data into the PNN network model online training program; Before the online training, make a steady state judgment. If it is judged to be in a steady state, return to the main program of fire monitoring; if not, add the sample to the training of the PNN neural network model to obtain a new PNN network model and write it into the main program of fire monitoring ; The trained PNN network model adopts an offline training method, and the offline training process is as follows: firstly, the hidden fire hazards in the confined space are investigated, and a simulated fire experiment is carried out; the simulated fire scene data is obtained through the gas sensor and the temperature sensor, and the selected The gas sensor response value S _i , the gas sensor change rate △S _i and the temperature response value T _i are preprocessed as corresponding features. The data preprocessing method adopts the array normalization method, and the calculation formula is: ${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; /</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$ Among them, X _{i, j} represents the jth eigenvalue in the i-th measurement, and X _i ' _{, j} represents the transformed value; Perform PCA principal component analysis on the normalized data X _i ' _{, j} , and find N (N<j) orthogonal characteristic variables, that is, principal components, through dimensionality reduction, so that they can reflect the main components of the data X _i ' _{, j} feature, to compress the scale of the original data matrix; when the contribution rate of the current N principal components reaches 90%, it is considered that the N principal components can reflect the main features of X _i ' _,j , and are used to fit the original data; use the former N Principal components, namely feature parameters, form a new sample set X={ _xi,n |n=1,2,...,N}, and apply it to the training of the new PNN network model; when the output expectation meets the requirements, terminate Training, write the formed PNN network model into the monitoring host. 2. The confined space fire detection and alarm system based on electronic nose technology according to claim 1, characterized in that: the online training program of the PNN network model is implemented as: judging whether the system is in a steady state; setting the threshold θ , when satisfying: ||X(i+1)-X(i)||<θ It is considered that the system has entered a steady state, so as to stop online training and maintain the PNN network model unchanged; when the system is not in a steady state, the output of the fire monitoring main program is used as the expected output of the sample, and added to the training sample set of the PNN network model Perform online training until the expected output of the test sample meets the conditions; the obtained new PNN network model is rewritten into the monitoring host for subsequent monitoring. 3. The confined space fire detection and alarm system based on electronic nose technology according to claim 1, characterized in that: the early warning time t(i) is defined as follows: $\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>& {\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}\end{array}\right.$ u(·) represents the unit step signal, i∈[0,T), T is the cycle time, when i=T, reset the early warning time t(i), U _i is the probability of fire occurrence. 4. A fire detection and alarm method for confined spaces based on electronic nose technology, characterized in that the implementation steps are as follows: (1) First conduct fire accident investigation in confined spaces, find out the main fire risks, and conduct simulation experiments for these main risks; (2) According to the data obtained from the simulation experiment, the PNN network model is trained offline; After normalizing the data, perform PCA principal component analysis and apply it to the training of the new PNN network model; when the output expectation meets the requirements, the training is terminated, and the formed PNN network model is input into the monitoring host; (3) Write the trained PNN network model into the monitoring host; Place the system sampling pipe in the confined space to be monitored, connect the three-way solenoid valve to the air, turn on the sensor power supply unit and control unit, turn on the vacuum pump to start pumping air, and start the monitoring host, and the data collected from the monitoring site can be obtained through the PNN network model. Occurrence probability, combined with early warning time judgment mechanism for real-time fire identification; (4) Carry out online training to the PNN network model by utilizing the real-time data obtained from the monitoring site; First, judge whether the system is in a steady state; when the system is in a steady state, stop online training and keep the PNN network model unchanged; when the system is not in a steady state, add data to the training sample set of the PNN network model for online After training, a new PNN network model is obtained, and rewritten into the monitoring host for subsequent monitoring.