A fire alarm screening model, method and system

Abstract

The present application relates to the technical field of fire monitoring, and particularly relates to a fire alarm screening model, method and system. The present application adopts a screening model comprising multiple judgment modules to identify real alarms and false alarms; each judgment module generates a judgment label of the current stage based on the associated spatial position information of the detector and the output label of the previous stage judgment module, and the judgment label of the last stage is the judgment label output by the screening model. The present application greatly reduces the false alarm rate while ensuring the detection sensitivity.

Description

Technical Field

[0001] This invention relates to the field of fire monitoring technology, and in particular to a fire alarm screening model, method and system. Background Technology

[0002] Smoke detectors are a crucial means of fire monitoring. Existing fire alarm systems can accurately monitor abnormal environmental conditions, detect fires in their early stages, and trigger fire control systems to extinguish them in their initial stages, reducing property damage and casualties. However, in reality, false alarms frequently occur due to environmental interference, equipment malfunctions, and other reasons, leading to the common occurrence of "crying wolf" scenarios.

[0003] To address the issue of false alarms from detectors, scholars have conducted extensive research on how to reduce the false alarm rate, primarily focusing on detection technologies and algorithms. For example, multi-fusion detection technology integrates different types of fire detectors into a collaborative detection system using Internet of Things (IoT) technology, enabling information exchange and performance complementarity between detectors.

[0004] Currently, dual-sensor detection devices integrating temperature and smoke detection functions are widely used. The redundant information provided by each detector helps enhance overall reliability, while complementary information can compensate for the limitations and uncertainties in the detection capabilities of single-type detectors, thereby effectively improving the overall performance and judgment accuracy of fire detection systems. However, due to the need to integrate multiple sensors, the hardware requirements are relatively large, and it is currently mostly used in research and not widely adopted.

[0005] Research on fire detection algorithms has led scholars to continuously develop efficient, high-precision algorithms adaptable to various complex environments. For example, in 2011, Sermanet et al. first applied deep learning methods to traffic sign detection, achieving significant results. The fusion of object detection and deep learning technologies can effectively address the problems of high false negative and false positive rates, low detection accuracy and speed, and significant spatial limitations inherent in traditional fire detection algorithms. Advanced convolutional neural networks for image detection include CNN, Fast-RCNN, and the YOLO series of algorithms. For instance, Chen Yanhong established the CNN-9 and CNN-17 models, achieving high accuracy in parallel smoke and flame detection. XU et al. proposed a lightweight YOLOv5 model, which can improve the accuracy and detection speed in complex fire scenarios.

[0006] The continuously updated judgment algorithm not only helps to achieve fast and accurate early fire warning, but also helps to reduce false alarms and missed alarms. However, current domestic and international research on fire detection, whether it is the innovation of detection principles, the development of recognition criteria, or the upgrade of basic components and discrimination algorithms, has not fundamentally solved the essential contradiction between sensitivity and false alarm rate. Currently, fire detection still relies on the single-point threshold alarm mode, lacking the联动 analysis of the spatio-temporal relationship between detector alarms. Summary of the Invention

[0007] To overcome the problem in the above-mentioned prior art that the detection of false fire alarms cannot solve the essential contradiction between sensitivity and false alarm rate, the present invention proposes a fire alarm screening model to explore the alarm linkage relationship between smoke detectors under the conditions of real and false alarms respectively, and through the mode of hierarchical alarm, predict the possibility of fire in real time and achieve false alarm recognition.

[0008] A fire alarm screening method proposed by the present invention uses a screening model including multiple judgment modules to identify real alarms and false alarms; each judgment module generates the judgment label of this level based on the associated detector spatial position information and the output label of the previous judgment module, and the judgment label of the last level is the judgment label output by the screening model; When any detector is triggered, the input data of the screening model is constructed based on the specified number of detectors that are triggered first in the monitoring area where the detector is located, and then the judgment label of this monitoring area is generated through the screening model.

[0009] Preferably, the detector spatial position information associated with the judgment module includes the characteristic data of each detector group associated with the judgment module; Let the characteristic data of the detector group composed of detector j and detector k be denoted as x(j,k), which includes the micro-element area of detector j, the micro-element type of detector j, the spatial distance, spatial relationship, time interval, floor difference, loop difference, and connectivity between the two detectors, etc.

[0010] Preferably, the set of detector groups associated with the nth judgment module is {Z(i,n + 1); 1 ≤ i ≤ n}, and Z(i,n + 1) represents the detector group composed of detector i and detector n + 1.

[0011] Preferably, the calculation rule of the connectivity α between two detectors is as follows: When two detectors are in the same micro-element, the connectivity α takes the connectivity coefficient corresponding to this micro-element; When two detectors are in different micro-elements, the connectivity α takes the cumulative product of the connectivity coefficients of all micro-elements passed through on the connectivity path between the two detectors; The connectivity coefficients of each micro-element type are set values.

[0012] Preferably, the connectivity coefficient is proportional to the openness corresponding to the micro-element type, and the connectivity coefficient of the room is greater than that of the corridor and the shaft.

[0013] Preferably, the screening model includes N-1 level judgment modules. This method specifically includes the following steps: St1. First, obtain the information of the N detectors that are triggered first in the monitoring area and construct a sample X(1)=x(1,2). Then, let the first-level judgment module generate a label Y(1) based on X(1). St2, iterate through h=2, 3……N-1, and let the h-th level judgment module generate the label Y(h) based on X(h)=x(1,h+1)||x(2,h+1)||…||x(h,h+1)||Y(h-1); output the label Y(N-1) as the final prediction result; Y(h-1) is the output label of the h-1 level judgment module.

[0014] The present invention proposes a model training method for the fire alarm screening method, which first obtains a training dataset with real labels, and then extracts a portion of the samples from the training dataset to construct a sample set C(1) for training the first-level judgment module. Then, iterating through h=2, 3...N-1, the judgment model at each level is trained as follows: Extract a portion of samples from the training dataset and label them Y(h-1) using the trained h-1 level judgment module; construct a sample set C(h) by combining the samples and the label Y(h-1), and train the h-level judgment module on the sample set C(h); Once the (N-1)th level judgment module has been trained, the selection model is fixed; N is the number of judgment modules.

[0015] The model training method as described in claim 7 is characterized by comprising the following steps: S1. Obtain the training dataset {[x(j,k), 1≤j≤N-1, 2≤j≤N]; y}, where y is the true label; x(j,k) is the feature data of the detector group consisting of detector i and detector k; the training dataset is obtained by oversampling the original dataset using SMOTE. S2. Extract a portion of the samples from the training dataset to construct a sample set C(1)={X(1),y(1)}, and train the first-level judgment module on the sample set C(1); X(1) includes the feature data of each detector group associated with the first-level judgment module; S3. Extract a portion of samples from the training dataset to construct a learning set {X(h-1),X(h);y(h)}, with h initially set to 2; X(h-1) and X(h) are the feature data of each detector group associated with the h-1 and h-th level judgment modules, respectively. S4. Let the h-1 level judgment module label the learning set with the label Y(h-1) based on the input X(h-1); S5. Combine the labeled label Y(h-1) and the learning set {X(h-1),X(h);y(h)} to construct the sample set C(h)={X(h),y(h)}, and train the h-th level judgment module on the sample set C(h); S6. Determine whether h is greater than or equal to N-1; If not, update h to h+1 and then return to step S2; If yes, then the screening model is fixed.

[0016] The loss function CB_FocalLoss proposed in this invention for the training process of the second-level judgment module to the (N-1)th judgment module is:

[0017]

[0018] in, Effective weights; β is the number of positive samples, and β is the equilibrium hyperparameter; It is a balancing factor; Modulation factor; This indicates the probability that the current judgment module predicts a real fire alarm. This represents the focus loss of the current judgment module.

[0019] The present invention proposes a fire alarm screening system, which includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the fire alarm screening method.

[0020] The advantages of this invention are: (1) The fire alarm screening model proposed in this invention achieves effective identification and decision fusion of complex alarm information through a hierarchical and cascaded architecture design; through the continuous updating of alarm information, the accuracy of the model in judging fire false alarms is continuously improved, and the accuracy of false alarm identification is gradually improved by continuously iterating alarm information.

[0021] (2) This invention achieves progressively refined analysis and verification of fire alarms through a multi-level (cascaded) judgment model. Each level integrates spatial context information and previous judgment results, mimicking the thought process of human experts in comprehensive judgment, thereby effectively identifying false alarms caused by environmental interference (such as dust and steam) and improving system reliability. This invention is the first to use the spatial positional relationship between detectors as the core judgment basis, enabling the model to understand the diffusion path and correlation of smoke and heat generated by fire in the building structure, upgrading the judgment basis from isolated "point" alarms to a correlated "network" situation assessment, with a more solid theoretical foundation. This invention also provides rich and multi-dimensional feature data, characterizing the correlation strength and pattern between alarm events of two detectors from multiple perspectives such as physical space, electrical wiring, and time series, providing a sufficient information foundation for the model to learn complex discrimination rules.

[0022] (3) This invention introduces the concept of connectivity in spatial relationships, transforming the concept of "connectivity" in architecture and fire protection engineering into calculable mathematical model parameters, enabling the model to "understand" the essential differences in the ease of smoke diffusion in different structural spaces (such as enclosed rooms, narrow corridors, and vertical shafts). Based on the fact that open spaces are more prone to smoke interconnection, this invention sets up connectivity calculation rules, making the model's assessment of alarm signal correlation more scientific and accurate, and more applicable to modern buildings with complex structures.

[0023] (4) The model input of the present invention only depends on the information of the first batch of detectors that are triggered. It does not need to wait for the alarm of the whole area or a fixed time window. It can quickly start the analysis and output the results in the early stage of the fire, so as to buy valuable time for subsequent emergency linkage and ensure the sensitivity of fire alarm detection.

[0024] (5) The present invention clarifies the input data of each level of judgment module, ensuring that as the judgment level increases, the information of the newly triggered detector can be cross-compared with all the triggered detectors, gradually constructing a complete local fire situation correlation map, so that the judgment field expands synchronously with the development of the event.

[0025] (6) To address the problem of unbalanced sample distribution in fire datasets, this invention designs the SMOTE equalization method and the CB_FocalLoss loss function. Through sample generation technology and the combination of focus loss and class balance method, the common class imbalance problem in alarm data is effectively alleviated, and the recognition accuracy of minority class samples is improved.

[0026] (7) By using the method of the present invention, real fire alarms and false alarms can be identified, so that relevant fire protection measures can be set to be implemented automatically only when a real fire alarm is detected, thus achieving a dual guarantee of fire detection sensitivity and reliability. Attached Figure Description

[0027] Figure 1 Flow chart of the training method for the screening model proposed by the present invention; Figure 2 Flow chart of a fire alarm screening method proposed by the present invention; Figure 3 Schematic diagram of detector layout and spatial distance; Figure 4 Confusion matrix of the prediction effect of the first-level judgment module in the embodiment; Figure 5 Confusion matrix of the prediction effect of the second-level judgment module in the embodiment; Figure 6 Model performance display in the embodiment. Specific implementation manners

[0028] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] A fire alarm screening model proposed by the present invention identifies real fire alarms and false fire alarms based on features such as the spatial position relationship and alarm time difference of the first N detectors triggered in the monitoring area.

[0030] The screening model includes N - 1 levels of judgment modules. Each level of judgment module generates a judgment label for this level based on the associated detector spatial position information and the output label of the previous level of judgment module. The judgment label of the last level is the judgment label output by the screening model.

[0031] The nth level of judgment module generates a judgment label based on the feature data of each detector group Z(i, n + 1) in the detector group set {Z(i, n + 1); 1 ≤ i ≤ n} and the judgment label Y(n - 1) output by the (n - 1)th level of judgment module. The judgment label includes real fire alarms and false fire alarms. When n = 1, Y(n - 1) is initialized as an empty set. The detectors in the detector group set are all triggered detectors.

[0032] Let the feature data of the detector group composed of detector j and detector k be denoted as x(j, k), where 1 ≤ j ≤ N - 1, 2 ≤ j ≤ N, and j < k. x(j, k) may specifically include: the micro - element area and micro - element type of detector j, as well as the spatial distance, spatial relationship, time interval, floor difference, loop difference, and connectivity between detectors j and k.

[0033] The micro - element refers to the overall chamber where the detector is located. The micro - element types include corridors, stairwells, rooms, etc.; Spatial relationships include: 0 indicates that the two detectors are in the same unit, 1 indicates that they are in adjacent units, 2 indicates that they are in the same layer but not adjacent units, 3 indicates that they are in adjacent layers, and 4 indicates that they are in non-adjacent layers. Spatial distance refers to the path distance between two detectors, for example Figure 3 In the above, detectors 2 and 3 are in the same micro-element, and the spatial distance d between them is the spatial coordinate distance; detectors 6 and 10 are in different micro-elements, and the path needs to pass through the connecting gate, so the spatial distance d is the sum of the distance d1 from detector 6 to the gate and the distance d2 from detector 10 to the gate. The time interval represents the time difference between the alarms of two detectors. Floor difference refers to the difference in floor level between the two detectors; The loop difference represents the difference in fire protection code loops between the two detectors; Connectivity indicates the degree of affinity between two detectors. The formula for calculating connectivity is: C = α × r / d Where r is the average of the protection radii of the two detectors, and the protection radius of the detector is the effective detection radius of the detector, which can be set to 7 meters; d is the spatial distance between the two detectors, and α is the connectivity coefficient.

[0034] If two detectors are located in the same micro-element space, then α is taken as the connectivity coefficient corresponding to that micro-element type. The connectivity coefficient is directly proportional to the openness corresponding to the micro-element type. For example, the larger the longest side length / vertical area of ​​the micro-element type, the larger the connectivity coefficient. For example, the longest side length of a corridor is its length, and the vertical area is the area of ​​the vertical surfaces at both ends; the longest side length of a shaft is its depth, and the vertical area is the area of ​​the horizontal surface. Specifically, the connectivity coefficients for rooms, corridors, and shafts can be set as follows: α r =0.7, α c =0.6, α s =0.6; If two detectors are located in different infinitesimal spaces, then α is the cumulative product of the connectivity coefficients of all infinitesimal elements traversed on the connecting path of the two detectors. α can be calculated according to the rules in the table below: Table 1: Connectivity Calculation Rules

[0035] Thus, the input of the first-level judgment module is denoted as X(1)=x(1,2), and the output label is denoted as Y(1); The input to the second-level judgment module is denoted as X(2)=x(1,3)||x(2,3)||Y(1), and the output label is denoted as Y(2); The input of the h-th (2≤h≤N-1) level judgment module is denoted as X(h), and the output label is denoted as Y(h); X(h)=x(1,h+1)||x(2,h+1)||…||x(h,h+1)||Y(h-1); Y(h-1) is the output label of the (h-1)th level judgment module.

[0036] x(1,h+1) represents the characteristic data of the detector group consisting of detector 1 and detector h+1, x(2,h+1) represents the characteristic data of the detector group consisting of detector 2 and detector h+1, x(h,h+1) represents the characteristic data of the detector group consisting of detector h and detector h+1; x(1,3) represents the characteristic data of the detector group consisting of detector 1 and detector 3, and x(2,3) represents the characteristic data of the detector group consisting of detector 2 and detector 3.

[0037] Reference Figure 1 The training method for the screening model proposed in this embodiment includes the following steps: S1. Obtain the training dataset {[x(j,k), 1≤j≤N-1, 2≤j≤N]; y}, where y is the true label; S2. Extract some samples from the training dataset to construct a sample set C(1)={X(1),y(1)}, and train the first-level judgment module on the sample set C(1); S3. Extract a portion of samples from the training dataset to construct a learning set {X(h-1),X(h);y(h)}, with h initially set to 2; S4. Let the h-1 level judgment module label the learning set with the label Y(h-1) based on the input X(h-1); S5. Combine the labeled label Y(h-1) and the learning set {X(h-1),X(h);y(h)} to construct the sample set C(h)={X(h),y(h)}, and train the h-th level judgment module on the sample set C(h); S6. Determine whether h is greater than or equal to N-1; If not, update h to h+1 and then return to step S2; If yes, then the screening model is fixed.

[0038] Reference Figure 2 The fire alarm screening method proposed in this embodiment has the following steps: St1. First, obtain the information of the N detectors that are triggered first in the monitoring area and construct sample X(1). Then, let the first-level judgment module generate label Y(1) based on X(1). St2, iterate through h=2, 3……N-1, and let the h-th level judgment module generate label Y(h) based on X(h); output label Y(N-1) as the final prediction result. Specific implementation examples: In this embodiment, N=3; the first-level judgment module and the second-level judgment module have the same structure, both using a 1D-CNN+DNN network.

[0040] The 1D-CNN+DNN network consists of a sequentially connected input layer, 1D convolutional layer, max pooling layer, adaptive average pooling layer, fully connected layer, and output layer.

[0041] Input layer: Receives input data; 1D convolutional layer: uses 32 convolutional kernels of size 1×3, padded with 1, and the activation function is ReLU; Max pooling layer: The pooling kernel size is 1×2, used for dimensionality reduction and feature selection; Adaptive average pooling layer: Adjusts the feature maps to a fixed length of 16; Fully connected network: contains two hidden layers (64 and 32 neurons), each followed by a ReLU activation function and a Dropout layer (dropout rate 0.2); Output layer: 1 neuron, using the Sigmoid activation function to output the initial classification probability.

[0042] Specifically, the input layer of the first-level judgment module receives the standardized 1-2 alarm point information difference data; the input layer of the second-level judgment module receives the 2-3 alarm information difference data, the 1-3 alarm information difference data, and the predicted label from the first-level judgment module; the second-level judgment module performs deep feature extraction and final classification.

[0043] The learning rate for this network model is set to 0.001 during training, and the Adam optimizer is used.

[0044] In this embodiment, the historical data database contains 1175 real fire incidents and 11236 highly suspected false alarms (actually false alarms, but the existing model judges them as real alarms) from the fire cloud platform. The original dataset is standardized. The standardized original dataset {[x(j,k), 1≤j≤N-1, 2≤j≤N;y]} is randomly shuffled and divided into 80% training dataset and 20% validation dataset. The training dataset contains 940 real alarms and 8989 false alarms. The number of real alarms is increased to 8989 using SMOTE, thus forming a training set with a total data volume of 8989+8989. The training set is further divided into a first training set {x(1,2);y} and a second training set {x(1,3),x(2,3);y}.

[0045] In this embodiment, the first-level judgment module is first trained on the first training set until convergence, and the loss function is the focal loss. Then, the first-level judgment module is used to perform first-level labeling on the second learning set. The second-level judgment module is trained on the second learning set {x(1,3),x(2,3),Y(1);y} after first-level labeling until convergence. The loss function is the composite loss function CB_FocalLoss. (3) (4) in, To improve the effectiveness of the weighting, higher loss weights are assigned to classes with fewer samples, effectively mitigating the class imbalance problem. The number of positive samples (real fire alarms) is denoted by β, and β is a balancing hyperparameter (value 0.99). This represents the weighting.

[0046] It is a balance factor (value 0.25). This design allows the model to maintain attention to hard-to-classify samples while balancing the differences in the number of different classes, thus improving classification performance on imbalanced datasets. This indicates the probability that the second-level judgment module predicts a real fire alarm. This is the focus loss for the second-level decision module.

[0047] γ is a modulation factor used to reduce the loss contribution of easily classified samples, allowing the model to focus more on difficult-to-classify samples; the specific value is γ = 2.

[0048] Both the Level 1 and Level 2 decision modules are binary classification models, outputting the probabilities of the two categories. The category with the highest probability is used as the label. Therefore, You can extract it directly from the model output.

[0049] In this embodiment, after the first-level judgment module and the second-level judgment module are trained, they form a screening model, and the performance of the screening model is tested on the validation set.

[0050] like Figure 4 , Figure 5 and Figure 6 As shown, in this embodiment, the annotation results of the first-level judgment module and the annotation results of the second-level judgment module are statistically analyzed respectively.

[0051] Figure 6 The data shows the true alarm recognition rate, true alarm recognition accuracy, and F1 score for the two-level judgment on the validation set. The true alarm recognition rate represents the proportion of predicted true fire alarms out of the total number of true fire alarms; that is, it measures the proportion of positive samples correctly predicted by the model. The true alarm recognition accuracy represents the proportion of fire alarms judged by the model to be true fire alarms; it measures the proportion of positive samples predicted by the model to be actually positive samples.

[0052] from Figure 6 It can be seen that the two-level judgment can gradually improve the accuracy of real alarm recognition while ensuring a real alarm recognition rate of over 98%. The second-level model (level 2 judgment module) maintains a balance in the real alarm recognition rate compared to the first-level model (level 1 judgment module), which may be due to the fact that a very small number of samples are difficult to distinguish and the number of edge cases processed by the second level is small. The second-level model improves the fire alarm recognition accuracy and F1 score, especially the real alarm recognition accuracy, which is improved by 9.34%, indicating the effectiveness of the second-level judgment model. Similarly, by continuously updating the alarm point information (such as the fourth alarm point information), the fire alarm recognition accuracy can be gradually improved while ensuring that real alarms are identified, thereby effectively reducing false alarms and reducing the workload of manual alarm verification.

[0053] In this embodiment, the screening model proposed in this invention is compared with the prior art, and the results are shown in Table 2.

[0054] Table 2: Real Police Recognition Rate and Accuracy under Different Models

[0055] The existing model in Table 2 triggers alarms based on the alarm time interval and spatial location of the alarm detectors; for example, if two or more alarms are triggered within the same structural unit (such as the same floor) within a set time interval (15 minutes), then a fire is determined to have occurred.

[0056] As shown in Table 2, the model of this invention outperforms LSTM, SVM, and residual DNN in fire alarm recognition accuracy by 8.14%, 8.26%, and 15.87%, respectively, indicating that the model structure of this invention performs better in false alarm identification. The model of this invention performs best in both true and false alarm identification, demonstrating higher accuracy in judging true and false fire alarms.

[0057] In this embodiment, an ablation experiment was further conducted. Table 3: Ablation Experiment Statistics

[0058] Ablation Model 1: This refers to the absence of SMOTE equalization processing on historical data; Ablation Model 2: The loss function used in the training process of the Level 2 judgment model is Focal Loss.

[0059] As shown in Table 3, the dataset after SMOTE equalization significantly improved in three metrics: real police identification rate, identification accuracy, and F1 score. This indicates that data equalization can effectively improve the model's accuracy in identifying highly suspected real police officers.

[0060] Compared to the standard loss function, CB_FocalLoss shows a significant improvement in the real police detection rate, but a decrease in the detection accuracy and F1 score. This indicates that the model in this invention pays more attention to minority samples, especially showing a significant improvement in the ability to identify minority class samples. This demonstrates that by combining a dual optimization method of class balancing weights and a difficult-to-easy sample focusing mechanism, the negative impact of class imbalance on model performance can be effectively mitigated. Considering the overall goal of avoiding false positives and reducing false positives, the CB_FocalLoss loss function has a significant advantage in identifying highly suspected real police cases.

[0061] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0062] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0063] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A fire alarm screening method, characterized in that, A screening model that includes multiple judgment modules is used to identify real alarms and false alarms. Each judgment module generates a judgment label based on the associated detector spatial location information and the output label of the previous judgment module. The final judgment label is the judgment label output by the screening model. When any detector is triggered, the input data for the screening model is constructed based on the specified number of detectors that were first triggered in the monitoring area where the detector is located. Then, the screening model is used to generate a judgment label for the monitoring area.

2. The fire alarm screening method as described in claim 1, characterized in that, The spatial location information of the detectors associated with the judgment module includes the characteristic data of each detector group associated with the judgment module; Let the characteristic data of the detector group consisting of detector j and detector k be denoted as x(j,k), which includes the infinitesimal area of ​​detector j, the infinitesimal type of detector j, and multiple factors such as the spatial distance between the two detectors, spatial relationship, time interval, floor difference, loop difference, and connectivity.

3. The fire alarm screening method as described in claim 2, characterized in that, The set of detector groups associated with the nth-level judgment module is {Z(i,n+1); 1≤i≤n}, where Z(i,n+1) represents a detector group consisting of detector i and detector n+1.

4. The fire alarm screening method as described in claim 2, characterized in that, The rule for calculating the connectivity α between two detectors is as follows: When two detectors are in the same micro-element, the connectivity α is taken as the connectivity coefficient corresponding to that micro-element. When two detectors are located in different infinitesimal elements, the connectivity α is the cumulative product of the connectivity coefficients of all infinitesimal elements traversed on the connecting path of the two detectors. The connectivity coefficients for each micro-element type are set values.

5. The fire alarm screening method as described in claim 4, characterized in that, The connectivity coefficient is directly proportional to the openness corresponding to the micro-element type, and the connectivity coefficient of a room is greater than that of a corridor and a shaft.

6. The fire alarm screening method as described in claim 2, characterized in that, Let the screening model include N-1 levels of judgment modules. The method specifically includes the following steps: St1. First, obtain the information of the N detectors that are triggered first in the monitoring area and construct a sample X(1)=x(1,2). Then, let the first-level judgment module generate a label Y(1) based on X(1). St2, iterate through h=2, 3……N-1, and let the h-th level judgment module generate the label Y(h) based on X(h)=x(1,h+1)||x(2,h+1)||…||x(h,h+1)||Y(h-1); output the label Y(N-1) as the final prediction result; Y(h-1) is the output label of the h-1 level judgment module.

7. A model training method for the fire alarm screening method as described in any one of claims 1-6, characterized in that, First, obtain the training dataset with real labels, and then extract some samples from the training dataset to construct a sample set C(1) for training the first-level judgment module; Then, iterating through h=2, 3...N-1, the judgment model at each level is trained as follows: Extract a portion of samples from the training dataset and label them Y(h-1) using the trained h-1 level judgment module; construct a sample set C(h) by combining the samples and the label Y(h-1), and train the h-level judgment module on the sample set C(h); Once the (N-1)th level judgment module has been trained, the selection model is fixed; N is the number of judgment modules.

8. The model training method as described in claim 7, characterized in that, Specifically, the following steps are included: S1. Obtain the training dataset {[x(j,k), 1≤j≤N-1, 2≤j≤N]; y}, where y is the true label; x(j,k) is the feature data of the detector group consisting of detector i and detector k; the training dataset is obtained by oversampling the original dataset using SMOTE. S2. Extract a portion of the samples from the training dataset to construct a sample set C(1)={X(1),y(1)}, and train the first-level judgment module on the sample set C(1); X(1) includes the feature data of each detector group associated with the first-level judgment module; S3. Extract a portion of samples from the training dataset to construct a learning set {X(h-1),X(h);y(h)}, with h initially set to 2; X(h-1) and X(h) are the feature data of each detector group associated with the h-1 and h-th level judgment modules, respectively. S4. Let the h-1 level judgment module label the learning set with the label Y(h-1) based on the input X(h-1); S5. Combine the labeled label Y(h-1) and the learning set {X(h-1),X(h);y(h)} to construct the sample set C(h)={X(h),y(h)}, and train the h-th level judgment module on the sample set C(h); S6. Determine whether h is greater than or equal to N-1; If not, update h to h+1 and then return to step S2; If yes, then the screening model is fixed.

9. The model training method as described in claim 7, characterized in that, The loss function CB_FocalLoss used in the training process of the second-level judgment module to the (N-1)th judgment module is: in, Effective weights; β is the number of positive samples, and β is the equilibrium hyperparameter; It is a balancing factor; Modulation factor; This indicates the probability that the current judgment module predicts a real fire alarm. This represents the focus loss of the current judgment module.

10. A fire alarm screening system, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the fire alarm screening method as described in claim 8 or 9.

Images