An intelligent alarm method and system

Abstract

This invention discloses an intelligent alarm method and system, relating to the field of intelligent security technology. It includes the following steps: S1, acquiring multimodal sensing data in real time from different types of sensors deployed within the target monitoring area. The sensing data includes at least two of the following: video data, audio data, infrared thermal imaging data, millimeter-wave radar data, and vibration data; and performing standardized preprocessing on the sensing data, including heterogeneous data normalization, timestamp alignment, and noise suppression, to obtain standardized sensing data in a unified format. This invention, by acquiring multimodal sensing data in real time and performing standardized preprocessing, eliminates reliance on single sensors, removes sensing blind spots, and solves the problem of incomplete sensing in existing systems. Combined with preset environmental context information, a multimodal fusion algorithm is used to extract behavioral features of the target object, overcoming the limitations of traditional single-data judgment, improving the comprehensive understanding of events, and reducing the false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, specifically to an intelligent alarm method and system. Background Technology

[0002] With the deepening development of smart city and IoT technologies, the demand for intelligent alarm systems in the fields of public safety and personal protection is increasing. As a core component of the security system, intelligent alarm technology is widely used in homes, communities, transportation, and industrial scenarios. Its core objective is to achieve rapid perception, accurate identification, and timely response to abnormal events. Current alarm systems typically rely on single sensors or preset rule triggering mechanisms, which are difficult to adapt to complex and ever-changing real-world environments, resulting in high false alarm rates, delayed responses, or frequent missed alarms, severely restricting the reliability and practicality of security systems.

[0003] Among these, intelligent alarm methods focus on improving the accuracy and robustness of alarms through multi-source information fusion and adaptive decision-making mechanisms. This direction aims to break through the limitations of traditional threshold triggering or simple logical judgments, and introduce capabilities such as environmental context awareness, behavioral pattern recognition, and dynamic risk assessment to effectively identify real threats.

[0004] Existing technologies for implementing intelligent alarms suffer from the following problems: First, the alarm triggering mechanism is too rigid, unable to dynamically adjust according to environmental changes or user habits, leading to frequent false alarms in complex scenarios. Second, it lacks the ability to collaboratively analyze multimodal sensing data, such as video, audio, infrared, and vibration, making it difficult to build a comprehensive event understanding model. Third, the system response strategy is simplistic, typically providing only fixed-level alarm outputs, unable to tiered handle events based on severity or link with other security devices. Finally, the user interaction experience is weak, failing to provide explainable alarm evidence or convenient feedback and correction mechanisms, reducing the system's credibility and usability. These problems are particularly prominent in application scenarios such as home security, monitoring of elderly people living alone, and unattended locations, urgently requiring a new alarm method and system with environmental adaptability, multi-source fusion sensing, and intelligent decision-making capabilities. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent alarm method and system that solves the problems of rigid alarm triggering mechanisms compared to existing technologies, which cannot be dynamically adjusted according to environmental changes or user habits, resulting in high false alarm rates, delayed responses, single response strategies, and weak user interaction experience in complex scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent alarm method, comprising: S1. Real-time acquisition of multimodal sensing data from different types of sensors deployed in the target monitoring area, including at least two of the following: video data, audio data, infrared thermal imaging data, millimeter-wave radar data, and vibration data; and standardization preprocessing of the sensing data, including heterogeneous data normalization, timestamp alignment, and noise suppression, to obtain standardized sensing data in a unified format. S2. Based on standardized perception data and combined with preset environmental context information obtained from historical records or external sources, the behavioral features of the target object are extracted through a multimodal fusion algorithm. The behavioral features include specific postures, movement trajectories, abnormal sound patterns, intrusion identification features, or changes in physiological signs. S3. Based on behavioral characteristics and environmental context information, a pre-trained intelligent decision-making model is used to conduct continuous dynamic risk situation assessment in order to generate a risk level corresponding to the severity of the event. S4. Based on the risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, hierarchical alarm decisions are made to generate alarm instructions that include alarm level, response type and handling measures; the alarm strategy can be dynamically adjusted according to environmental changes and historical event feedback. S5. Based on the alarm command, execute the corresponding hierarchical alarm response, including at least one of local audible and visual alarms, remote notifications, linkage with security equipment, and activation of emergency plans; and collect user feedback on alarm results to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

[0007] Furthermore, step S1 specifically includes: different types of sensors including high-definition cameras, voiceprint recognition microphone arrays, infrared thermal imagers, vibration sensors, and environmental parameter sensors; standardized preprocessing also includes target detection and tracking of video data, abnormal sound source localization and classification of audio data, and abnormal body temperature detection of infrared thermal image data, in order to generate preliminary event clues.

[0008] Furthermore, step S2 specifically includes: preset environmental context information including the geographical location of the monitoring area, current time, weather conditions, historical pedestrian traffic data, identity information of people in the area, and predefined normal behavior patterns; the multimodal fusion algorithm adopts a deep fusion network based on an attention mechanism to weightedly fuse feature vectors from different modalities in order to capture the correlation between modalities; behavioral features also include the interaction patterns between the target object and the environment.

[0009] Furthermore, step S3 specifically includes: the intelligent decision-making model is trained based on massive historical event data, expert-annotated data, and simulated attack and defense exercise data. The model can calculate the potential threat index of an event based on the duration, intensity, frequency, spatial range, and matching degree with the environmental context of the behavioral characteristics; the risk level is subdivided into five levels, including: safe, attention, warning, severe, and emergency, with each level corresponding to different response levels and handling priorities.

[0010] Furthermore, step S4 specifically includes: the adaptive alarm strategy dynamically updates the strategy parameters after receiving user feedback or observing the development of an event through a reinforcement learning mechanism; the hierarchical alarm decision also considers the alarm timeliness requirements, prioritizing the triggering of responses for high-risk events and simultaneously triggering the associated contingency plan processes; the alarm instruction also includes the location of the event, timestamp, and relevant image or video evidence.

[0011] Furthermore, the step of executing the corresponding graded alarm response according to the alarm command includes: when the alarm level is warning or above, issuing a high-intensity alarm through the local sound and light alarm device and sending a remote notification containing event details to the designated management personnel through an encrypted channel; when the alarm level is serious or emergency, in addition to the above response, the system is also linked to the intelligent access control system to lock down the area, launch drones for on-site patrols or directly report the alarm information to the public safety management platform, and automatically dispatch on-site personnel.

[0012] Furthermore, the collection of user feedback information on alarm results to continuously optimize the intelligent decision-making model and adaptive alarm strategy includes: user feedback information on alarm results includes confirmation of alarm, marking of false alarms or missed alarms, suggestions for correcting alarm levels, and evaluation of response efficiency; the system uses the feedback information as labeled data or reward signals to retrain the intelligent decision-making model and adjust the parameters of the adaptive alarm strategy to improve the model's adaptability in complex and variable scenarios.

[0013] Furthermore, the preset environmental context information also includes: seasonal personnel flow patterns, day and night light variation patterns, equipment operation status data, and a database of historical abnormal events in the specific monitoring area established through initial configuration and system self-learning; the personnel flow patterns are used to identify movement during abnormal periods or on abnormal paths, and the equipment operation status data is used to eliminate false alarms caused by equipment failures.

[0014] Furthermore, the dynamic risk situation assessment also includes: when multiple independent or related threat events occur in close proximity in time or space, conducting collaborative risk analysis to assess the overall risk level and potential cascading effects; when the confidence level of the assessment result is lower than a preset threshold, the system automatically triggers secondary cross-validation of multimodal data, or requests human experts to make auxiliary judgments, or generates manual review prompts and pushes them to the regulatory terminal, and receives the manual review results returned by the regulatory terminal, so as to reduce the misjudgment rate and improve the reliability of decision-making.

[0015] The present invention also provides an intelligent alarm system, applied to any of the above-mentioned intelligent alarm methods, comprising: The multi-source sensing module is used to acquire multimodal sensing data from different types of sensors deployed in the target monitoring area in real time, and to perform standardized preprocessing on the sensing data to obtain standardized sensing data in a unified format. The feature extraction and fusion module is used to extract the behavioral features of the target object based on standardized perceptual data and combined with preset environmental context information obtained from historical records or external sources, through a multimodal fusion algorithm. The risk situation assessment module is used to continuously and dynamically assess risk situation based on behavioral characteristics and environmental context information, using a pre-trained intelligent decision-making model to generate a risk level corresponding to the severity of the event. The alarm decision module is used to make hierarchical alarm decisions based on risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, so as to generate alarm instructions containing alarm level, response type and handling measures. The alarm response and feedback optimization module is used to execute corresponding hierarchical alarm responses according to alarm commands and collect user feedback information on alarm results in order to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention eliminates reliance on single sensors and removes blind spots by acquiring multimodal sensing data in real time and performing standardized preprocessing, thus solving the problem of incomplete perception in existing systems. Combined with pre-set environmental context information, a multimodal fusion algorithm is used to extract behavioral features of target objects, overcoming the limitations of traditional single-data judgment, improving the comprehensive understanding of events, and reducing false alarm rates. A pre-trained intelligent decision-making model is used to dynamically assess risk levels, coupled with an adaptive alarm strategy based on reinforcement learning or expert rule bases, to achieve tiered alarm decision-making, solving the problems of response lag and strategy rigidity. Tiered alarm responses are executed and linked with security equipment, avoiding single responses and improving the effectiveness of handling. User feedback is collected to continuously optimize the model and strategy, enhancing system adaptability and user trust, and improving the overall reliability and practicality of the intelligent alarm system to meet the needs of various security scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides an intelligent alarm method, comprising: S1. Real-time acquisition of multimodal sensing data from different types of sensors deployed in the target monitoring area, including at least two of the following: video data, audio data, infrared thermal imaging data, millimeter-wave radar data, and vibration data; and standardization preprocessing of the sensing data, including heterogeneous data normalization, timestamp alignment, and noise suppression, to obtain standardized sensing data in a unified format. S2. Based on standardized perception data and combined with preset environmental context information obtained from historical records or external sources, the behavioral features of the target object are extracted through a multimodal fusion algorithm. The behavioral features include specific postures, movement trajectories, abnormal sound patterns, intrusion identification features, or changes in physiological signs. S3. Based on behavioral characteristics and environmental context information, a pre-trained intelligent decision-making model is used to conduct continuous dynamic risk situation assessment in order to generate a risk level corresponding to the severity of the event. S4. Based on the risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, hierarchical alarm decisions are made to generate alarm instructions that include alarm level, response type and handling measures; the alarm strategy can be dynamically adjusted according to environmental changes and historical event feedback. S5. Based on the alarm command, execute the corresponding hierarchical alarm response, including at least one of local audible and visual alarms, remote notifications, linkage with security equipment, and activation of emergency plans; and collect user feedback on alarm results to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

[0020] Specifically, this intelligent alarm method is applied in home security scenarios to address the problems of existing systems, such as reliance on single sensors, high false alarm rates, delayed response, and rigid strategies. First, multimodal sensing data collection is implemented. High-definition cameras are deployed at the entrance of residential courtyards to collect video data of people and vehicles. Voiceprint recognition microphone arrays are installed at the edges of doors and windows to capture audio data such as door opening and closing sounds and abnormal impact sounds. Infrared thermal imagers are placed in the corners of living rooms and bedrooms to collect human thermal image data. Vibration sensors are installed near door and window locks to monitor vibration data generated by lock picking and impacts. Through the collaborative coverage of multiple types of sensors, the blind spots of a single sensor are completely eliminated. Then, standardized preprocessing is performed, normalizing the pixel brightness values ​​of video data, the decibel values ​​of audio data, and the amplitude values ​​of vibration data to the [0,1] range, eliminating analysis obstacles caused by differences in heterogeneous data formats. Timestamp alignment technology is used to synchronize the acquisition times of different sensors to the millisecond level, avoiding event timing errors caused by data transmission delays. A Gaussian filtering algorithm is used to filter environmental noise in the audio and light and shadow interference in the video, ensuring data validity.

[0021] Next, the system combines pre-defined environmental context information, including family daily activity patterns such as no one being home during the day on weekdays and family members mostly resting in the bedroom after 10 PM, as well as predefined normal behavioral patterns such as the voiceprint characteristics of family members opening doors and their daily walking trajectories. Using a multimodal fusion algorithm, the standardized data is analyzed for correlation, focusing on uncovering the logical relationships between different modalities. For example, when non-family member silhouettes appear in the video data and a metallic friction sound inconsistent with a normal door-opening voiceprint is heard in the synchronized audio data, the system can accurately extract the combined behavioral features of "abnormal personnel movement and abnormal friction sound," overcoming the limitations of traditional systems that rely solely on single data points for judgment.

[0022] A pre-trained intelligent decision-making model is used for dynamic risk assessment. This model is trained on massive amounts of historical home security event data, including real intrusion cases, false alarm scenarios, and expert-annotated abnormal behavior data. It can dynamically adjust its judgment logic based on environmental context. For example, if the model detects the above-mentioned combined behavioral characteristics when no one is home during the weekday, it will comprehensively assess the severity of the event and generate a corresponding risk level, avoiding the problem of traditional fixed rules failing to adapt to environmental changes. Based on the risk level and adaptive alarm strategy, a tiered decision-making process is implemented. The adaptive alarm strategy is continuously optimized through reinforcement learning. If there have been previous false alarms due to video movement caused by wind blowing curtains, the system will adjust the strategy parameters based on user feedback, no longer classifying video movement without synchronized audio anomalies as high risk. Finally, a tiered alarm response is executed. High-risk events trigger local audio and visual alarms, remote encrypted notifications, and smart lock linkage locking. User feedback is collected and used for model and strategy optimization, effectively reducing false alarm and missed alarm rates and improving the accuracy and timeliness of alarm response.

[0023] In this embodiment, step S1 specifically includes: different types of sensors including high-definition cameras, voiceprint recognition microphone arrays, infrared thermal imagers, vibration sensors, and environmental parameter sensors; standardized preprocessing also includes target detection and tracking of video data, abnormal sound source localization and classification of audio data, and abnormal body temperature detection of infrared thermal image data, in order to generate preliminary event clues.

[0024] Specifically, in the office building security scenario, the specific operations of step S1 involve deploying different types of sensors according to the security needs of the office building's public and key areas. High-definition cameras are installed at the lobby entrance, elevator cars, and corridors on each floor to capture real-time video dynamics of personnel flow and object movement. Voiceprint recognition microphone arrays are distributed near glass curtain walls, equipment room doors, and fire escape doors to accurately collect audio information around doors and windows. Infrared thermal imagers are deployed in office areas and power distribution rooms to simultaneously monitor personnel body temperature and equipment operating temperature. Vibration sensors are installed on equipment room cabinet doors, fire escape door locks, and elevator shaft walls to detect vibrations caused by abnormal impacts or prying. Environmental parameter sensors are fixed on the ceilings of each floor to collect indoor temperature, humidity, smoke concentration, and air quality data.

[0025] In the standardization preprocessing stage, in addition to normalizing heterogeneous data, aligning timestamps, and suppressing noise, specific feature extraction is carried out for different types of perceived data. For video data, the YOLO object detection algorithm is used to identify target categories such as people, tables, chairs, and equipment, and the Kalman filter is used to track the target movement path. For example, if people are detected moving in the computer room area during non-working hours and their trajectories are irregular, they are immediately marked as targets of interest. For audio data, Mel-frequency cepstral coefficients are used to extract features, and the support vector machine algorithm is used to locate and classify abnormal sound sources, accurately distinguishing normal sounds such as air conditioner operation and conversations from abnormal sounds such as glass breaking and equipment malfunctions. If an abnormal sound is detected near the computer room and located around the power distribution cabinet, an audio anomaly clue is directly generated. For infrared thermal imaging data, a threshold segmentation algorithm is used to detect abnormal body temperature. If the body temperature of people in the office area exceeds the normal range, or the temperature of equipment in the computer room is significantly higher than the normal operating value, preliminary event clues are generated.

[0026] In this embodiment, step S2 specifically includes: preset environmental context information including the geographical location of the monitoring area, current time, weather conditions, historical pedestrian traffic data, identity information of people in the area, and predefined normal behavior patterns; the multimodal fusion algorithm adopts a deep fusion network based on an attention mechanism to weightedly fuse feature vectors from different modalities in order to capture the correlation between modalities; behavioral features also include the interaction patterns between the target object and the environment.

[0027] Specifically, in the community security scenario, the operation of step S2 involves pre-setting environmental context information that covers multi-dimensional static and dynamic data of the community. This includes the geographical location of the community in the urban-rural fringe area, the current time (2:00 AM), real-time rainfall conditions, historical pedestrian flow data showing minimal activity in the community during this time period, a database of resident facial and vehicle information built through prior data entry, and predefined normal behavior patterns such as residents only moving around their own buildings at night and vehicles needing to register at the access control system before entering the community. Furthermore, this pre-set information also includes seasonal population flow patterns in the community formed by the system through initial configuration and long-term self-learning, such as large numbers of people leaving during the Spring Festival, the community mainly populated by the elderly and children during weekdays, day-night light variation patterns such as only main roads and building entrances being lit at night while other areas are dark, real-time operational status data of various monitoring devices such as whether cameras are functioning properly and whether the access control system is malfunctioning, and a historical database of abnormal events recording past cases of outsiders climbing over walls during the early morning hours.

[0028] The multimodal fusion algorithm is constructed based on a deep fusion network using an existing attention mechanism. The attention mechanism is a mature technology and will not be elaborated here. The specific fusion process is implemented through formula (1): ; The weight coefficients of each modal feature satisfy the following: , in the formula, The fused feature vector represents the final output. The number of modes participating in the fusion is represented here by three modes: video, audio, and infrared. , Representing the Weight coefficients of modal features Representing the The feature vectors of each modality after standardization. The weight coefficients... Data is acquired through network training, using historical multimodal anomaly event data as samples. The goal is to improve the accuracy of fused features in identifying abnormal behavior, and the data is continuously adjusted. Until the model converges.

[0029] In practical applications, if the video data captures the outline of people moving within the fenced area... The audio data collected included sound characteristics of clothing rubbing and climbing over walls. The infrared data detected the live thermal image features of the moving target. The system will use the weights obtained during training. The fusion features are calculated using the above formula. This fusion feature not only includes independent information from each modality, but also strengthens the correlation between "person movement, climbing sounds, and live thermal images". At the same time, it can combine the interaction patterns between the target object and the environment, such as the way people contact the wall and whether the movement path deviates from the normal activity range of the community, to accurately extract abnormal behavior features.

[0030] In this embodiment, step S3 specifically includes: the intelligent decision-making model is trained based on massive historical event data, expert-annotated data, and simulated attack and defense exercise data. The model can calculate the potential threat index of an event based on the duration, intensity, frequency, spatial range, and matching degree with the environmental context of the behavioral characteristics; the risk level is subdivided into five levels, including: safe, attention, warning, severe, and emergency, with each level corresponding to different response levels and handling priorities.

[0031] Specifically, in the security scenario of the industrial park, the operation of step S3 is implemented. The training data of the intelligent decision model includes historical event data of the industrial park over the past five years, including real cases such as equipment failure alarms, unauthorized entry of personnel into restricted areas, and intrusion by outsiders; event feature data jointly annotated by park security experts and equipment maintenance experts; and test data generated through simulated attack and defense exercises, such as scenario data simulating outsiders climbing over the factory wall and unauthorized operation of production equipment. The model training process adopts the existing random forest algorithm framework. The algorithm framework will not be elaborated here. The potential threat index of the event is calculated by formula (2): In the formula, The potential threat index represents the event. The weight coefficients represent the duration, intensity, frequency, and spatial range of the behavioral characteristics, respectively. Each weight coefficient is obtained by fitting the training data. The fitting objective is to make the calculated threat index highly correlated with the severity of the event labeled by the experts. This represents the normalized duration of the behavioral feature, which converts the original duration to the [0,1] interval. This represents the normalized intensity of behavioral characteristics, such as the result of normalizing the original intensity data of equipment operation force and personnel movement speed. This represents the normalized frequency of behavioral characteristics, such as the normalized result of the number of times personnel travel to and from restricted areas. This represents the normalized spatial range of behavioral characteristics, such as the normalized result of the restricted area covered by personnel activity zones.

[0032] The above normalization uses min-max linear normalization to map the original data to the [0,1] interval.

[0033] During dynamic risk situation assessment, if human behavior characteristics are detected in a certain area, the system will first obtain the original duration, intensity, frequency, and spatial range data of the behavior, and then normalize them to obtain... Then substitute the values ​​into the above formula to calculate the threat index. For example, during non-working hours in the production workshop, if a person is detected entering and continuously touching the control buttons of the production equipment, the original duration of this behavior is 12 minutes, which is then normalized. The intensity is manifested in the normalized state of the forced operation equipment. The frequency is the normalized value after continuous operation. The spatial range covers the core control area of ​​the equipment after normalization. Combined with the weight coefficients obtained during training The threat index was calculated. .

[0034] Risk levels are divided into five categories: safe, alert, warning, severe, and urgent. The system determines the risk level based on the threat index. The range of determination levels, such as For safety, For attention, For warning, Serious It was an emergency. In the above cases... Once an emergency level is identified, the highest response level and priority are assigned to ensure that the system can match appropriate follow-up actions based on the severity of the incident, avoiding over-response that wastes resources or under-response that leads to increased risk.

[0035] In this embodiment, step S4 specifically includes: the adaptive alarm strategy dynamically updates the strategy parameters after receiving user feedback or observing the development of an event through a reinforcement learning mechanism; the hierarchical alarm decision also considers the alarm timeliness requirements, prioritizing the triggering of responses for high-risk events and simultaneously triggering the associated contingency plan processes; the alarm instruction also includes the location of the event, timestamp, and related image or video evidence.

[0036] Specifically, in the mall security scenario, step S4 is implemented. The adaptive alarm strategy is built based on the existing reinforcement learning mechanism. The reinforcement learning framework is an existing technology and will not be elaborated here. The strategy optimization process uses user feedback and event development results as reward signals to dynamically update the strategy parameters. For example, previously, during peak weekend hours in the mall, the system judged the behavior of a large number of people gathering in front of a store in a short period of time as abnormal and triggered an alarm. After on-site verification by mall security personnel, it was found that the situation was normal shopping traffic caused by promotional activities. After the staff reported this result to the system, the system treated it as a negative reward signal and adjusted the judgment threshold of "people gathering" behavior in the strategy. Subsequently, short-term people gathering in specific promotional areas on weekends will no longer be judged as abnormal.

[0037] During the tiered alarm decision-making process, the system prioritizes alarm timeliness. For high-risk events, such as detecting smoke and flame outlines in a store within the mall, the system immediately triggers the response process, simultaneously activating the mall's fire emergency plan. This includes cutting off power to the area and activating smoke extraction equipment to prevent risk spread due to response delays. The generated alarm commands contain rich information, clearly defining the alarm level, response type, and handling measures. They also indicate the specific store location where the smoke and flames appeared, such as store number 3 in the food area on B1 floor, the precise timestamp of the event (e.g., 202X-202X-202X-2023:15), and images and short video evidence from on-site cameras. This information is simultaneously transmitted to the mall's fire control room and security command center via an encrypted network, providing staff with complete evidence for quickly locating the event and understanding the situation, significantly improving response efficiency. Furthermore, the alarm strategy dynamically adjusts according to changes in the mall's environment at different times. For example, during peak holiday periods, the alarm threshold for pedestrian traffic is appropriately relaxed, while vigilance is increased during weekday non-business hours, ensuring the strategy adapts to the complex and ever-changing operational environment of the mall.

[0038] In this embodiment, according to the alarm command, the corresponding graded alarm response is executed, including: when the alarm level is warning or above, a high-intensity alarm is issued through the local sound and light alarm device, and a remote notification containing event details is sent to the designated management personnel through an encrypted channel; when the alarm level is serious or emergency, in addition to the above response, the intelligent access control system is linked to lock down the area, a drone is launched for on-site inspection, or the alarm information is directly reported to the public safety management platform, and on-site personnel are automatically dispatched, or a dispatch command is sent to the terminal device of the on-site personnel, or an alarm information requesting dispatch is sent to the emergency management platform.

[0039] Specifically, in residential community security scenarios, a tiered alarm response is implemented. When the alarm level is determined to be a warning, for example, if the system detects through video data that an unfamiliar person is loitering for an extended period at the entrance of a building unit, and simultaneously captures audio data indicating that the person is repeatedly attempting to touch the unit door's combination lock, the local audio-visual alarm device installed at the entrance of that building unit will immediately activate, emitting a high-intensity audio-visual signal. This serves both to alert residents inside the building and to deter the unfamiliar person. Simultaneously, the system sends a remote notification to the community security personnel's terminals via an encrypted channel. The notification includes details of the unfamiliar person's physical characteristics, such as height approximately 1.75 meters, wearing a dark coat, and having loitered for at least 15 minutes, facilitating rapid on-site verification by security personnel.

[0040] When the alarm level is determined to be serious or urgent, if security personnel confirm through real-time footage transmitted by the system that an unknown person has forcibly broken into the building and entered the building, in addition to triggering the aforementioned local audible and visual alarms and remote notifications, the system will also link with the community's intelligent access control system to automatically block all elevator floor buttons in the building, prohibiting elevators from stopping in the building, and simultaneously closing the fire escape doors within the building to prevent people from escaping; activate the community's drone patrol equipment, controlling the drone to fly to the window of the corresponding floor in the building, to capture real-time images of indoor personnel activities and transmit the images back to the security command center; and report the alarm information directly to the city's public safety management platform through a dedicated interface. After receiving the information, the platform will automatically match nearby security patrol personnel, generate a dispatch order, and arrange personnel to go to the scene to handle the situation in the shortest possible time.

[0041] In this embodiment, user feedback on alarm results is collected to continuously optimize the intelligent decision-making model and adaptive alarm strategy. This includes user feedback on alarm results, such as confirmation of alarms, marking of false alarms or missed alarms, suggestions for correcting alarm levels, and evaluation of response efficiency. The system uses the feedback information as labeled data or reward signals to retrain the intelligent decision-making model and adjust the parameters of the adaptive alarm strategy to improve the model's adaptability in complex and variable scenarios.

[0042] Specifically, in factory security scenarios, user feedback optimization is implemented. Users participating in the feedback include factory security management personnel, workshop supervisors, and equipment maintenance personnel. When the system triggers an alarm, relevant users submit feedback information to the system based on on-site verification results. If an alarm indicates an abnormal overheating of a machine tool in the workshop due to a heat dissipation failure, and the equipment maintenance personnel confirm the alarm's accuracy after on-site inspection, the system will report "Alarm valid." If an alarm caused by vibration from the normal operation of a large machine in the workshop is mistakenly identified as an abnormal impact, the maintenance personnel will mark the alarm as a "false alarm" and provide an explanation for the false alarm. If an alarm concerning minor equipment noise in the workshop is classified as a warning by the system, but the workshop supervisor, based on production experience, believes the noise does not currently pose a warning risk, a correction suggestion of "reducing the alarm level to alert" will be made. If a high-risk alarm, such as a fire signal in the workshop's raw material warehouse, results in a long arrival time for security personnel due to route planning issues, the system's response efficiency will be evaluated as "slow response, route recommendation needs optimization."

[0043] After integrating these feedback messages, the system will classify and process them as follows: For confirmation feedback such as "alarm valid" and "false alarm", they will be added to the training dataset of the intelligent decision-making model as labeled data to retrain the model and adjust the model's judgment criteria for features such as equipment temperature and vibration; For suggestion feedback such as "correct alarm level" and "evaluate response efficiency", they will be used as reward signals or adjustment basis to modify the parameters of the adaptive alarm strategy. For example, for false alarm feedback of normal equipment vibration, the abnormal judgment threshold of vibration data in the strategy will be increased.

[0044] In this embodiment, the preset environmental context information also includes: seasonal personnel flow patterns, day and night light variation patterns, equipment operation status data, and a database of historical abnormal events in the specific monitoring area established through initial configuration and system self-learning; the personnel flow patterns are used to identify movement during abnormal periods or on abnormal paths, and the equipment operation status data are used to eliminate false alarms caused by equipment failures.

[0045] Specifically, in campus security scenarios, additional preset environmental context information is applied. The system enters basic environmental data of the campus through initial configuration, including the functional division of various areas of the campus such as teaching area, living area, sports area, and main passage location. Combined with the self-learning function during long-term operation, the environmental context system is continuously updated and improved. The seasonal personnel flow pattern includes the surge in campus traffic caused by students returning to school at the start of the semester and the sharp decrease in traffic caused by only a small number of faculty and staff remaining on campus during winter and summer vacations. The diurnal light variation pattern records the characteristics of sufficient natural light in teaching areas during the day and weaker light in areas such as sports fields at night, where only the main roads and dormitories are lit. The equipment operation status data collects the operating parameters of all surveillance cameras, access control systems, and fire-fighting equipment on campus in real time, such as whether the cameras have blurry images or obstructed lenses, whether the access control system can properly recognize student campus cards and faculty and staff work permits, and whether the fire alarms are in normal standby mode. The historical abnormal event database stores past campus safety incidents, such as cases of unregistered outsiders entering teaching buildings during winter and summer vacations, students climbing over the campus wall to go out, and minor fires caused by laboratory equipment malfunctions. Each case includes detailed information such as the time, location, and behavioral characteristics involved in the incident.

[0046] In actual security monitoring, this additional environmental context information works in conjunction with multimodal perception data. For example, during winter and summer vacations, if the system detects movement of people in the teaching area through video data, it will combine seasonal personnel flow patterns to determine the abnormality of personnel activity during that period. If personnel movement is detected in the sports field at night, it will combine the day and night light change patterns, call infrared thermal image data, and calculate the fusion feature using Formula 1 to further confirm whether the target is a living person, avoiding false alarms caused by insufficient light. If the image captured by the camera at the entrance of a teaching building is abnormally blurry, the system will retrieve the equipment operation status data to confirm whether the camera lens is blocked or the equipment is malfunctioning. If it is determined to be an equipment problem, the data from that camera will be temporarily blocked, and only data from other sensors will be used for analysis, effectively eliminating false alarms caused by equipment failure.

[0047] In this embodiment, the dynamic risk situation assessment also includes: when multiple independent or related threat events occur in close time or space, conducting collaborative risk analysis to assess the overall risk level and potential cascading effects; when the confidence level of the assessment result is lower than a preset threshold, the system automatically triggers secondary cross-validation of multimodal data, or requests human experts to make auxiliary judgments, or generates human review prompts and pushes them to the regulatory terminal, and receives the human review results returned by the regulatory terminal, so as to reduce the misjudgment rate and improve the reliability of decision-making.

[0048] Specifically, in the security scenarios of commercial complexes, an additional operation for dynamic risk situation assessment is implemented. When the system detects multiple independent or related threat events within a similar time or space—for example, smoke signals appear in a restaurant on the first floor of the complex, while people are running and shouting in the clothing area on the third floor—the system will initiate a collaborative risk analysis process. This process is based on existing event correlation analysis algorithms, which will not be elaborated on here. By mining the potential logical relationships between events, it determines the possible causal relationship between the smoke on the first floor and the running on the third floor. That is, the smoke on the first floor may trigger panic among people on the third floor, leading to running behavior. The system will comprehensively assess the risk level of the combined effect of the two events, avoiding the underestimation of risk caused by assessing only a single event. At the same time, it predicts possible cascading effects, such as stampedes and the spread of panic to other floors, providing a basis for formulating a comprehensive response strategy.

[0049] When the system's confidence level in assessing an event falls below a preset threshold—for example, detecting a suspicious package in the underground parking garage of a complex, where video data is insufficient to clearly identify the package's appearance due to low light, and infrared data is also insufficient to determine if there is an abnormal temperature inside the package—the system will automatically trigger a secondary cross-validation of multimodal data. During the validation process, the system will utilize millimeter-wave radar data from the underground parking garage to detect the presence of suspicious components such as metal or liquid within the package's internal structure. Simultaneously, it will combine this data with parking garage audio data to detect any abnormal sounds, such as the ticking of a timer. Formula 1 is used to fuse and calculate the data from the four modalities: video, infrared, millimeter-wave radar, and audio. If the confidence level of the assessment result after the secondary validation still does not reach the preset threshold, the system will send a request to the complex's security experts via its internal communication module, pushing all modal data and preliminary analysis results to the expert's terminal for manual assistance in making a judgment.

[0050] Please see Figure 2 The present invention also provides an intelligent alarm system, applied to any of the above-mentioned intelligent alarm methods, comprising: The multi-source sensing module is used to acquire multimodal sensing data from different types of sensors deployed in the target monitoring area in real time, and to perform standardized preprocessing on the sensing data to obtain standardized sensing data in a unified format. The feature extraction and fusion module is used to extract the behavioral features of the target object based on standardized perceptual data and combined with preset environmental context information obtained from historical records or external sources, through a multimodal fusion algorithm. The risk situation assessment module is used to continuously and dynamically assess risk situation based on behavioral characteristics and environmental context information, using a pre-trained intelligent decision-making model to generate a risk level corresponding to the severity of the event. The alarm decision module is used to make hierarchical alarm decisions based on risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, so as to generate alarm instructions containing alarm level, response type and handling measures. The alarm response and feedback optimization module is used to execute corresponding hierarchical alarm responses according to alarm commands and collect user feedback information on alarm results in order to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

[0051] Specifically, in a smart park scenario, this intelligent alarm system is deployed with a multi-source sensing module serving as the core of the system's data input. Various sensors are deployed at key locations based on the park's security needs, including high-definition cameras and license plate recognition cameras at park entrances and exits, voiceprint recognition microphone arrays along main roads, infrared thermal imagers and vibration sensors in building server rooms, and environmental parameter sensors in public areas. The module collects multi-modal sensing data in real time, including video, audio, infrared, and vibration data. The module's internal preprocessing unit standardizes the data, performing heterogeneous data normalization, timestamp alignment, and noise suppression, converting raw data of different formats into standardized sensing data in a unified format, providing high-quality data support for subsequent modules.

[0052] After receiving standardized perception data, the feature extraction and fusion module first retrieves preset environmental context information from the park management system, including the functional division of each area of ​​the park, historical traffic data, personnel and vehicle authorization information, equipment operating status, etc. Then, based on the deep fusion network with attention mechanism, it performs feature fusion on multimodal data through Formula 1 to extract the behavioral features of the target object, such as identifying features such as unauthorized vehicles entering the park restricted area or abnormal personnel activities in the equipment room. This process can effectively capture the correlation between different modal data and improve the accuracy of feature extraction.

[0053] The risk situation assessment module calls a pre-trained intelligent decision-making model, which is trained based on past security incident data, expert-annotated data, and simulation exercise data in the park. The model calculates the potential threat index of the event using Formula 2, and assesses the severity of the event by combining environmental context information, generating a corresponding risk level. For example, the behavior of unauthorized personnel operating equipment in the computer room is assessed as a serious risk level.

[0054] The alarm decision module generates alarm commands based on the risk level and the adaptive alarm strategy. The adaptive alarm strategy is continuously optimized through a reinforcement learning mechanism and can dynamically adjust the decision logic according to changes in the park environment. The alarm command clearly includes the alarm level, response type, handling measures, event location, timestamp, and relevant evidence information.

[0055] The alarm response and feedback optimization module executes response operations based on alarm commands. For example, severe risk events will trigger local audible and visual alarms, link access control to lock down areas, and send dispatch instructions to security personnel. At the same time, the module collects user feedback on alarm results in real time and uses it to optimize intelligent decision-making model parameters and adaptive alarm strategies, enabling continuous iteration and upgrading of the system and comprehensively improving the intelligence level and reliability of park security.

[0056] In summary, this invention eliminates reliance on single sensors and removes blind spots by acquiring multimodal sensing data in real time and performing standardized preprocessing, thus solving the problem of incomplete perception in existing systems. By combining pre-set environmental context information with a multimodal fusion algorithm to extract behavioral features of target objects, it overcomes the limitations of traditional single-data judgment, improves the comprehensive understanding of events, and reduces false alarm rates. A pre-trained intelligent decision-making model dynamically assesses risk levels, coupled with an adaptive alarm strategy based on reinforcement learning or expert rule bases, to achieve tiered alarm decision-making, solving the problems of response lag and strategy rigidity. Tiered alarm responses are executed and linked with security equipment to avoid single responses and improve the effectiveness of handling. User feedback is collected to continuously optimize the model and strategy, enhancing system adaptability and user trust, and overall improving the reliability and practicality of the intelligent alarm system to meet the needs of various security scenarios.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent alarm method, characterized in that, include: S1. Real-time acquisition of multimodal sensing data from different types of sensors deployed in the target monitoring area, including at least two of the following: video data, audio data, infrared thermal imaging data, millimeter-wave radar data, and vibration data; and standardization preprocessing of the sensing data, including heterogeneous data normalization, timestamp alignment, and noise suppression, to obtain standardized sensing data in a unified format. S2. Based on standardized perception data and combined with preset environmental context information obtained from historical records or external sources, the behavioral features of the target object are extracted through a multimodal fusion algorithm. The behavioral features include specific postures, movement trajectories, abnormal sound patterns, intrusion identification features, or changes in physiological signs. S3. Based on behavioral characteristics and environmental context information, a pre-trained intelligent decision-making model is used to conduct continuous dynamic risk situation assessment in order to generate a risk level corresponding to the severity of the event. S4. Based on the risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, hierarchical alarm decisions are made to generate alarm instructions that include alarm level, response type and handling measures; the alarm strategy can be dynamically adjusted according to environmental changes and historical event feedback. S5. Based on the alarm command, execute the corresponding hierarchical alarm response, including at least one of local audible and visual alarms, remote notifications, linkage with security equipment, and activation of emergency plans; and collect user feedback on alarm results to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

2. The intelligent alarm method according to claim 1, characterized in that, Step S1 specifically includes: different types of sensors including high-definition cameras, voiceprint recognition microphone arrays, infrared thermal imagers, vibration sensors, and environmental parameter sensors; standardized preprocessing also includes target detection and tracking of video data, abnormal sound source localization and classification of audio data, and abnormal body temperature detection of infrared thermal image data, in order to generate preliminary event clues.

3. The intelligent alarm method according to claim 1, characterized in that, Step S2 specifically includes: preset environmental context information including the geographical location of the monitoring area, current time, weather conditions, historical pedestrian flow data, identity information of people in the area, and predefined normal behavior patterns; the multimodal fusion algorithm adopts a deep fusion network based on an attention mechanism to weightedly fuse feature vectors from different modalities in order to capture the correlation between modalities; behavioral features also include the interaction patterns between the target object and the environment.

4. The intelligent alarm method according to claim 1, characterized in that, Step S3 specifically includes: the intelligent decision-making model is trained based on massive historical event data, expert-annotated data, and simulated attack and defense exercise data. The model can calculate the potential threat index of an event based on the duration, intensity, frequency, spatial range, and matching degree with the environmental context of the behavioral characteristics. The risk level is subdivided into five levels, including: safe, attention, warning, severe, and emergency. Each level corresponds to a different response level and handling priority.

5. The intelligent alarm method according to claim 1, characterized in that, Step S4 specifically includes: the adaptive alarm strategy dynamically updates the strategy parameters after receiving user feedback or observing the development of an event through a reinforcement learning mechanism; the hierarchical alarm decision also considers the alarm timeliness requirements, prioritizing the triggering of responses for high-risk events and simultaneously triggering the associated contingency plan processes; the alarm command also includes the location of the event, timestamp, and relevant image or video evidence.

6. The intelligent alarm method according to claim 1, characterized in that, The process of executing corresponding graded alarm responses based on alarm commands includes: when the alarm level is warning or above, issuing a high-intensity alarm through local audible and visual alarm devices and sending a remote notification containing event details to designated management personnel through an encrypted channel; when the alarm level is serious or urgent, in addition to the above responses, linking the intelligent access control system to lock down the area, launching drones for on-site patrols, or directly reporting the alarm information to the public safety management platform, and automatically dispatching on-site personnel.

7. The intelligent alarm method according to claim 1, characterized in that, The process of collecting user feedback on alarm results to continuously optimize the intelligent decision-making model and adaptive alarm strategy includes: user feedback on alarm results including confirmation of alarm, marking of false alarms or missed alarms, suggestions for correcting alarm levels, and evaluation of response efficiency; the system uses the feedback information as labeled data or reward signals to retrain the intelligent decision-making model and adjust the parameters of the adaptive alarm strategy to improve the model's adaptability in complex and variable scenarios.

8. The intelligent alarm method according to claim 3, characterized in that, The preset environmental context information also includes: seasonal personnel flow patterns, day and night light variation patterns, equipment operation status data, and a database of historical abnormal events in the specific monitoring area established through initial configuration and system self-learning; the personnel flow patterns are used to identify movement during abnormal periods or on abnormal paths, and the equipment operation status data is used to eliminate false alarms caused by equipment failures.

9. The intelligent alarm method according to claim 1, characterized in that, The dynamic risk situation assessment also includes: when multiple independent or related threat events occur in close time or space, conducting collaborative risk analysis to assess the overall risk level and potential cascading effects; when the confidence level of the assessment results is lower than a preset threshold, the system automatically triggers secondary cross-validation of multimodal data or requests human experts to make auxiliary judgments to reduce the misjudgment rate and improve the reliability of decision-making.

10. An intelligent alarm system, applied to any one of the intelligent alarm methods of claims 1-9, characterized in that, include: The multi-source sensing module is used to acquire multimodal sensing data from different types of sensors deployed in the target monitoring area in real time, and to perform standardized preprocessing on the sensing data to obtain standardized sensing data in a unified format. The feature extraction and fusion module is used to extract the behavioral features of the target object based on standardized perceptual data and combined with preset environmental context information obtained from historical records or external sources, through a multimodal fusion algorithm. The risk situation assessment module is used to continuously and dynamically assess risk situation based on behavioral characteristics and environmental context information, using a pre-trained intelligent decision-making model to generate a risk level corresponding to the severity of the event. The alarm decision module is used to make hierarchical alarm decisions based on risk level and the adaptive alarm strategy formed by the system through reinforcement learning or expert rule base, so as to generate alarm instructions containing alarm level, response type and handling measures. The alarm response and feedback optimization module is used to execute corresponding hierarchical alarm responses according to alarm commands and collect user feedback information on alarm results in order to continuously optimize the intelligent decision-making model and adaptive alarm strategy.

Images