A multi-party alarm system based on cloud-edge collaboration

Abstract

The application discloses a kind of multi-party alarm system based on cloud edge cooperation, specifically related to edge computing technical field, including multi-modal sensor state reliability calibration, structured cross-modal feature vector generation, edge high-confidence event determination, optimal alarm disposal team list generation, multi-party linkage alarm instruction generation.The application fuses acoustic and thermal imaging visual data, and generates reliability scalar by real-time evaluation of signal quality, can dynamically adjust the weight of data source in decision-making, enhance the accuracy and robustness of event identification.Event determination is completed completely on the edge side, eliminates the delay of network transmission and cloud secondary check, ensures the immediacy of alarm.After alarm triggering, the system intelligently matches according to event demand and personnel dynamic capability vector, selects the optimal personnel combination, and then generates instructions for the selected different personnel, converts the alarm event into a closed-loop, automated multi-party linkage emergency response process.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a multi-party alarm system based on cloud-edge collaboration. Background Technology

[0002] Edge computing technology encompasses a distributed computing paradigm that pushes computation and data storage to the logical edge of the network. The core of this technology lies in reducing latency, saving network bandwidth, and enhancing system response speed. Its systematic process begins with data sources such as IoT devices or terminal sensors. This data undergoes initial processing and analysis at edge nodes close to its origin. These edge nodes can be smart gateways or dedicated servers, uploading only the processed results, metadata, or data requiring in-depth analysis to a centralized cloud data center. The cloud then handles large-scale data analysis, model training, and long-term storage tasks, and distributes updated models or instructions to the edge nodes, forming a collaborative architecture where the cloud handles training and management, and the edge handles execution and inference.

[0003] One such multi-party alarm system based on cloud-edge collaboration refers to a specific solution for layered processing of audio data to achieve multi-user alarms. This technical solution addresses real-time analysis of audio signals, preliminary event assessment, and alarm information distribution and confirmation. Specifically, a pre-built acoustic model runs on the microphone or a nearby edge computing device. This device captures the raw audio waveform, extracts its Mel-frequency cepstral coefficients, and inputs them into the model for comparison with pre-defined abnormal sound event features. When the model's classification confidence exceeds a predetermined threshold, a preliminary alarm event is generated locally. Subsequently, the edge device packages the type, timestamp, and corresponding audio feature vector of this preliminary alarm event and sends it to the cloud server. Upon receiving the data, the cloud server calls a larger-scale, more accurate deep learning model to perform secondary verification and analysis of the audio feature vector. Simultaneously, based on the microphone's unique identifier and event type, it retrieves multiple recipients to be notified from a pre-defined association matrix. After secondary verification confirms a genuine alarm, the final alarm information is pushed to all designated recipient terminals.

[0004] The shortcomings of existing technologies lie in their reliance on a single acoustic signal for judgment, making them prone to misjudgment or missed judgment in industrial scenarios with complex background noise. The system lacks an evaluation mechanism for the sensor's own operational status; when the microphone's signal is distorted due to interference or malfunction, the system cannot perceive the unreliability of the data source, affecting decision-making accuracy. Its alarm process has inherent delays, with final confirmation relying on secondary analysis in the cloud. The time spent on data traveling between the cloud and the edge is fatal for emergencies requiring instantaneous response. Furthermore, its alarm distribution mechanism is rigid, pushing information to fixed recipients based solely on a preset matrix, failing to intelligently allocate resources based on real-time personnel status, resulting in inefficient allocation of emergency resources and prolonged emergency response time. Summary of the Invention

[0005] The main objective of this invention is to provide a multi-party alarm system based on cloud-edge collaboration, which can effectively solve the problems mentioned in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A cloud-edge collaborative multi-party alarm system includes the following steps: S1. Multimodal sensor state reliability calibration: Acquire microphone audio and thermal imaging video, calculate audio signal-to-noise ratio, clipping distortion, and video sharpness and saturation, and obtain the multimodal reliability scalar set through mapping matrix transformation; S2. Structured cross-modal feature vector generation: Based on the multimodal confidence scalar set, extract audio and video feature vectors, and connect the confidence scalars with the corresponding feature vectors in sequence to generate structured confidence feature vectors. S3. High-confidence event determination at the edge: Based on the structured confidence feature vector input to the edge device network, obtain the pipeline event confidence vector, determine whether the leakage and howling dimension exceeds the threshold, and if it does, encapsulate the timestamp, location and event identifier to establish a high-confidence edge alarm event. S4. Optimal Alarm Handling Team List Generation: Based on the retrieval demand vector of high-confidence edge alarm events, match the capability vector of each receiver, filter the optimal user combination, and obtain the optimal alarm handling team list. S5. Generation of multi-party linkage alarm commands: Call the optimal alarm handling team list and high-confidence edge alarm events, configure emergency shut-off valve control permissions, navigation paths and other tasks for team members and encapsulate them to generate multi-party linkage alarm commands.

[0007] Preferably, the multimodal confidence scalar set includes audio confidence scalar and image confidence scalar; the structured cross-modal feature vector specifically includes audio confidence data segment, audio feature data segment, image confidence data segment, and image feature data segment; the high-confidence edge alarm event includes event type identifier, event occurrence timestamp, event spatial coordinates, and event association feature vector; the optimal alarm handling team list includes pre-selected user identifier, comprehensive matching score, and predetermined response role; and the multi-party linkage alarm instruction specifically refers to remote control token, navigation path data, and security protocol instruction.

[0008] Preferably, step S1 includes: Audio signals from MEMS microphone arrays deployed in a specific area are collected, the ratio of audio signal power to background noise power is calculated in parallel, and the proportion of sample points with amplitude exceeding the preset upper limit in the audio signal is counted to obtain the audio signal quality index. The signal-to-noise ratio and clipping distortion in the audio signal quality indicators are called, and a preset two-dimensional mapping matrix is ​​retrieved. The corresponding relationship in the matrix is ​​used to perform table lookup and transformation to obtain the audio reliability scalar. The video signal from the infrared thermal imaging sensor is acquired, and the sharpness and brightness saturation of the video signal are converted into image reliability scalars in parallel. The audio reliability scalar is called, and the two are integrated to generate a multimodal reliability scalar set.

[0009] Preferably, S2 includes: Based on the multimodal confidence scalar set, the original audio signal contained therein is called, and a short-time Fourier transform is performed on the signal to obtain the time spectrum. The Mel-scale triangular filter bank is then used to filter the time spectrum, and a discrete cosine transform is performed on the logarithmic energy of each channel after filtering to obtain the Mel frequency cepstral coefficient vector. Based on the multimodal confidence scalar set, the original video signal contained therein is called. For the sequential image frames in the video signal, matrix convolution operation with preset convolution kernel is performed frame by frame. Max pooling operation is performed on the output after convolution operation to downsample. The processing results of multiple frames are concatenated to generate convolution feature map vector. The audio and image reliability scalars from the multimodal reliability scalar set are called, along with the Mel frequency cepstral coefficient vector and the convolutional feature map vector. The audio reliability scalar, Mel frequency cepstral coefficient vector, image reliability scalar, and convolutional feature map vector are concatenated end-to-end according to a preset dimensional order to establish a structured reliability feature vector.

[0010] Preferably, S3 includes: For the structured confidence feature vector, it is input into the network structure inside the edge computing device. In this structure, matrix multiplication is performed with the preset weight matrix, and the result is processed by a nonlinear activation function. The processed data is then passed sequentially in the stacked structure to obtain an event classification confidence vector for the state of the high-pressure steam pipeline. Call the event classification confidence vector, and set the event trigger threshold according to the receiver operation characteristic curve. Determine whether the dimension value corresponding to the leakage howling sound in the vector exceeds the event trigger threshold. Convert the judgment result to a Boolean value and generate an alarm trigger status identifier. Based on the alarm trigger status identifier, if it indicates a triggered state, the current timestamp and location coordinates are retrieved from the device and encapsulated together with the preset event type identifier and structured confidence feature vector to establish a high-confidence edge alarm event.

[0011] Preferably, S4 includes: Based on the event type identifier in the high-confidence edge alarm event, the event demand vector is retrieved from the cloud knowledge base, and the response capability vector uploaded by each receiving terminal, which consists of dynamic location parameters, activity status scalar, historical processing efficiency parameters and skill weight parameters, is called. A weighted summation operation is performed on the response capability vector and the event demand vector of each receiving party to obtain the matching degree of each receiving party. Based on the matching degree of each recipient and the number of people required for the event, all recipients that meet the number of people conditions are selected and user groups are generated. For each user group, the sum of the skill weight parameters of its members is calculated, and the comprehensive geographical distance between it and the event location is calculated based on the dynamic location parameters to establish a candidate handling team plan. The candidate handling team plan is invoked, and all user combinations in the plan list are iteratively compared. The sorting and selection operations are performed based on the dual screening criteria of maximizing the sum of skill weights and minimizing the comprehensive geographical distance, to determine the unique user combination that meets the conditions and obtain the optimal alarm handling team list.

[0012] Preferably, S5 includes: The system calls up the optimal alarm handling team list and high-confidence edge alarm events, retrieves the preset permission matrix for the central control room operators in the list, and generates a remote operation authorization credential associated with the emergency shut-off valve based on the operator's user ID and the event type identifier. Based on the optimal alarm handling team list and high-confidence edge alarm events, a navigation path is generated for the on-site inspectors in the list by calling the map service interface, and the wearing instructions for related personal protective equipment are retrieved from the device safety protocol library. The two are then integrated into a set of on-site handling procedures. The system invokes remote operation authorization credentials and on-site handling procedures, extracts core event information from high-confidence edge alarm events, sets alarm levels and distribution channels, encapsulates all the above information items, and generates multi-party linkage alarm commands.

[0013] Compared with the prior art, the present invention has the following beneficial effects: By fusing acoustic and thermal imaging visual data and generating a reliability scalar through real-time signal quality evaluation, the weight of data sources in decision-making can be dynamically adjusted, enhancing the accuracy and robustness of event identification. Event determination is completed entirely at the edge, eliminating latency from network transmission and secondary verification in the cloud, ensuring the immediacy of alarms. After an alarm is triggered, the system intelligently matches the event requirements with the dynamic capability vectors of personnel, selecting the optimal combination of personnel in terms of location, skills, and status. The system then generates specific, executable instructions for the selected personnel in different roles, such as remote control permissions and navigation paths, transforming the alarm event into a closed-loop, automated, multi-party collaborative emergency response process. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0015] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0016] Example 1, as Figure 1 As shown, a multi-party alarm system based on cloud-edge collaboration includes the following steps: S1. Multimodal sensor state reliability calibration: Acquire audio signals from a MEMS microphone array deployed in a specific area and video signals from an infrared thermal imaging sensor. Calculate the ratio of audio signal power to background noise power in parallel to obtain the signal-to-noise ratio (SNR). Calculate the proportion of audio signal amplitude exceeding a preset upper limit to the total number of sample points to obtain clipping distortion. Call a preset two-dimensional mapping matrix to convert the SNR and clipping distortion into audio reliability scalars. Convert the video signal's sharpness and brightness saturation into image reliability scalars to obtain a multimodal reliability scalar set. S2. Structured Cross-Modal Feature Vector Generation: Based on the multimodal confidence scalar set, short-time Fourier transform and Mel-scale filtering are performed on the audio signal to obtain the Mel frequency cepstral coefficient vector. Convolution and pooling operations are performed on the video signal to extract the convolution feature map vector. The audio confidence scalar, audio Mel frequency cepstral coefficient vector, image confidence scalar, and image convolution feature map vector are called and concatenated end to end according to the preset dimensional order to generate a structured confidence feature vector. S3. High-Confidence Event Judgment at the Edge: For the structured confidence feature vector, it is input into a network structure composed of matrix multiplication and nonlinear activation function stacked inside the edge computing device to obtain an event classification confidence vector for the status of high-pressure steam pipeline. It is determined whether the dimension value corresponding to the leakage whistling sound in the vector exceeds the event trigger threshold set according to the receiver operation feature curve. If it exceeds, the current timestamp and location coordinates are retrieved from inside the device and encapsulated together with the event type identifier and the structured confidence feature vector to establish a high-confidence edge alarm event. S4. Generation of the optimal alarm handling team list: Based on the event type identifier in the high-confidence edge alarm event, retrieve the event demand vector from the cloud knowledge base, and call the response capability vector uploaded by each receiving terminal, which consists of dynamic location parameters, activity status scalar, historical handling efficiency parameters and skill weight parameters. Perform a weighted summation operation to obtain the matching degree of each receiving party, filter out the user combination that meets the number of people requirement, maximizes the sum of skill weights and minimizes geographical distance, and obtain the optimal alarm handling team list; S5. Multi-party linkage alarm command generation: Invoke the optimal alarm handling team list and high-confidence edge alarm events, retrieve the permission matrix for the central control room operators in the list, generate a temporary remote control token associated with the emergency shut-off valve based on their user identifier and event type identifier, generate a navigation path for the field inspector by calling the map service interface, and retrieve the wearing instructions associated with personal protective equipment from the equipment safety protocol library. Encapsulate the alarm level, distribution channel and core event information to generate multi-party linkage alarm commands.

[0017] The multimodal reliability scalar set includes audio reliability scalar and image reliability scalar. The structured cross-modal feature vectors specifically include audio reliability data segments, audio feature data segments, image reliability data segments, and image feature data segments. High-confidence edge alarm events include event type identifiers, event occurrence timestamps, event spatial coordinates, and event-related feature vectors. The optimal alarm handling team list includes pre-selected user identifiers, comprehensive matching scores, and pre-defined response roles. Multi-party linkage alarm instructions specifically refer to remote control tokens, navigation path data, and security protocol instructions.

[0018] Step S1 includes: Audio signals from MEMS microphone arrays deployed in a specific area are collected, the ratio of audio signal power to background noise power is calculated in parallel, and the proportion of sample points with amplitude exceeding the preset upper limit in the audio signal is counted to obtain the audio signal quality index. Furthermore, based on the calculation of the signal power and background noise power within a specific time window, for example, if the signal power is 0.8W and the background noise power is 0.1W, the signal-to-noise ratio is calculated using the formula: The signal-to-noise ratio is 9dB. The number of sample points whose signal amplitude exceeds 95% of the upper limit of the 16-bit sampling range (31133) within the time window is counted. If the total number of sample points is 48000 and the number of sample points exceeding the upper limit is 480, then the clipping distortion is 1%, and the audio signal quality index is obtained.

[0019] The signal-to-noise ratio and clipping distortion in the audio signal quality indicators are called, and a preset two-dimensional mapping matrix is ​​retrieved. The corresponding relationship in the matrix is ​​used to perform table lookup and transformation to obtain the audio reliability scalar. For example, a combination with a signal-to-noise ratio range of 5dB to 10dB and a clipping distortion of less than 1.5% corresponds to a reliability scalar of 0.6. Based on this correspondence, a lookup table and conversion are performed to obtain the audio reliability scalar.

[0020] The video signal from the infrared thermal imaging sensor is acquired, and the sharpness and brightness saturation of the video signal are converted into image reliability scalars in parallel. The audio reliability scalar is called, and the two are integrated to generate a multimodal reliability scalar set.

[0021] For example, the sharpness value is calculated to be 0.8 using the Laplacian operator, and the brightness and saturation value is calculated to be 0.7 using the HSV color space. A weighted summation operation is performed based on preset weight coefficients: 0.6 + 0.8 + 0.4 + 0.7 to obtain the image reliability scalar value of 0.76. The audio reliability scalar value of 0.6 is called, and the two are integrated into a two-dimensional array to generate a multimodal reliability scalar set.

[0022] S2 includes: Based on the multimodal confidence scalar set, the original audio signal contained therein is called, and a short-time Fourier transform is performed on the signal to obtain the time spectrum. The Mel-scale triangular filter bank is then used to filter the time spectrum, and a discrete cosine transform is performed on the logarithmic energy of each channel after filtering to obtain the Mel frequency cepstral coefficient vector. Furthermore, based on the multimodal confidence scalar set, the original audio signal contained therein is called, and the frame length is set to 25 milliseconds and the frame shift to 10 milliseconds to perform a frame-by-frame windowing operation on the signal. A short-time Fourier transform is performed on each frame to obtain the time spectrum. The time spectrum is then passed through a Mel-scale filter bank consisting of 40 triangular filters. The logarithmic energy of the output of each filter is calculated. Finally, a discrete cosine transform is performed on these 40 logarithmic energy values ​​and the first 13 coefficients are taken to obtain the Mel frequency cepstral coefficient vector.

[0023] Based on the multimodal confidence scalar set, the original video signal contained therein is called. For the sequential image frames in the video signal, matrix convolution operation with preset convolution kernel is performed frame by frame. Max pooling operation is performed on the output after convolution operation to downsample. The processing results of multiple frames are concatenated to generate convolution feature map vector. For a sequence of 128x128 pixel image frames in the signal, a matrix convolution operation with a preset 3x3 kernel is performed frame by frame. Then, a 2x2 window max pooling operation is performed on the 126x126 output matrix after convolution to downsample it, resulting in a 63x63 feature map. The feature maps of three consecutive frames are flattened and concatenated to generate a convolutional feature map vector.

[0024] The audio and image reliability scalars from the multimodal reliability scalar set are called, along with the Mel frequency cepstral coefficient vector and the convolutional feature map vector. The audio reliability scalar, Mel frequency cepstral coefficient vector, image reliability scalar, and convolutional feature map vector are concatenated end-to-end according to a preset dimensional order to establish a structured reliability feature vector.

[0025] Furthermore, the audio reliability scalar 0.6 and the image reliability scalar 0.76 from the multimodal reliability scalar set are invoked, along with the Mel frequency cepstral coefficient vector and the convolutional feature map vector. According to the preset dimensional order, the audio reliability scalar 0.6 is used as the first element, followed by the 13-dimensional Mel frequency cepstral coefficient vector, then the image reliability scalar 0.76, and finally the convolutional feature map vector. The four elements are then concatenated end-to-end to establish a structured reliability feature vector.

[0026] S3 includes: For the structured confidence feature vector, it is input into the network structure inside the edge computing device. In this structure, matrix multiplication is performed with the preset weight matrix, and the result is processed by a nonlinear activation function. The processed data is then passed sequentially in the stacked structure to obtain an event classification confidence vector for the state of the high-pressure steam pipeline. For the structured confidence feature vector, it is input into a network structure containing two fully connected layers. The vector is first multiplied by the weight matrix of the first layer. The result is processed by the modified linear unit activation function. Then the processed result is input into the second layer and multiplied by the weight matrix of the second layer. Finally, the output vector is normalized by the Softmax function to obtain an event classification confidence vector for the state of high-pressure steam pipeline.

[0027] Call the event classification confidence vector, and set the event trigger threshold according to the receiver operation characteristic curve. Determine whether the dimension value corresponding to the leakage howling sound in the vector exceeds the event trigger threshold. Convert the judgment result to a Boolean value and generate an alarm trigger status identifier. The event classification confidence vector is invoked, and based on the preset receiver operation characteristic curve, the event trigger threshold is set to 0.85 while ensuring a 95% recall rate. It is then determined whether the value of the second dimension representing the leaking howling sound in the vector, 0.9, exceeds the event trigger threshold of 0.85. Since 0.9 is greater than 0.85, this determination result is converted to a Boolean value to generate an alarm trigger status identifier.

[0028] Based on the alarm trigger status identifier, if it indicates a triggered state, the current timestamp and location coordinates are retrieved from the device and encapsulated together with the preset event type identifier and structured confidence feature vector to establish a high-confidence edge alarm event.

[0029] Based on the alarm trigger status identifier, since its current value is true, the timestamp 20251007013112 and the location coordinates of pipe No. 3 in area B are called from inside the edge device, and together with the preset event type identifier L-02 and the structured confidence feature vector, they are encapsulated into a JSON format data object to establish a high confidence edge alarm event.

[0030] S4 includes: Based on the event type identifier in the high-confidence edge alarm event, the event demand vector is retrieved from the cloud knowledge base, and the response capability vector uploaded by each receiving terminal, which consists of dynamic location parameters, activity status scalar, historical processing efficiency parameters and skill weight parameters, is called. A weighted summation operation is performed on the response capability vector and the event demand vector of each receiving party to obtain the matching degree of each receiving party. Based on the event type identifier L-02 in the high-confidence edge alarm event, the corresponding event requirement vector is retrieved from the cloud knowledge base. Its value is [pipeline maintenance skill weight 0.7, safety protocol familiarity weight 0.2, emergency equipment operation weight 0.1]. The response capability vector uploaded by receiver A is then retrieved, with corresponding skill weight parameters [0.9, 0.8, 0.6]. A weighted summation operation is performed on the two, resulting in 0.7 + 0.9 + 0.2 + 0.8 + 0.1. 0.6 gives receiver A a matching score of 0.85. The matching scores of other receivers are calculated in the same way to obtain the matching degree of each receiver.

[0031] Based on the matching degree of each recipient and the number of people required for the event, all recipients that meet the number of people conditions are selected and user groups are generated. For each user group, the sum of the skill weight parameters of its members is calculated, and the comprehensive geographical distance between it and the event location is calculated based on the dynamic location parameters to establish a candidate handling team plan. Furthermore, considering the requirement of two people for event L-02, receivers A, B, C, and D with a matching degree higher than 0.7 were selected, generating six user groups. For the group (A,C), the sum of the skill weight parameters of its members was calculated. For example, the sum of A's weights was 2.3, C's was 2.1, and the sum of the group's weights was 4.4. Based on receiver A's dynamic location parameters (10,50) and the event location (15,55), the Euclidean distance was calculated to be 7.07. The distances of other members within the group were calculated in the same way and averaged to establish a candidate handling group plan.

[0032] The candidate handling team plan is invoked, and all user combinations in the plan list are iteratively compared. The sorting and selection operations are performed based on the dual screening criteria of maximizing the sum of skill weights and minimizing the comprehensive geographical distance, to determine the unique user combination that meets the conditions and obtain the optimal alarm handling team list.

[0033] Furthermore, the groups are first sorted in descending order based on the sum of their skill weights. If the sum of the weights of the combination (A,C) is 4.4, it is selected directly. If the sum of the weights of the combination (B,D) is also 4.4, the comprehensive geographical distance between the two groups is compared. The group with the distance value of 7.07, which is lower than the sum of the weights of the combination (B,D) (8.12), is selected to obtain the optimal alarm handling team list.

[0034] S5 includes: The system calls up the optimal alarm handling team list and high-confidence edge alarm events, retrieves the preset permission matrix for the central control room operators in the list, and generates a remote operation authorization credential associated with the emergency shut-off valve based on the operator's user ID and the event type identifier. Furthermore, the optimal alarm handling team list and high-confidence edge alarm events are invoked. For the central control room operator with user identifier C-01 in the list, the preset permission matrix is ​​retrieved. Based on the operator identifier C-01 and the event type identifier L-02, the operation permission is retrieved to remotely close the No. 3 emergency shut-off valve. This permission is encrypted and combined with the user identifier, device number and a dynamic timestamp with a validity period of 5 minutes to generate a remote operation authorization certificate associated with the emergency shut-off valve.

[0035] Based on the optimal alarm handling team list and high-confidence edge alarm events, a navigation path is generated for the on-site inspectors in the list by calling the map service interface, and the wearing instructions for related personal protective equipment are retrieved from the device safety protocol library. The two are then integrated into a set of on-site handling procedures. Furthermore, the map service interface is called for the on-site inspector identified as F-05 in the list. The current coordinates (10,50) and the event location coordinates (15,55) are input to generate a navigation path. The wearing instructions associated with the L-02 event are retrieved from the equipment safety protocol library, namely wearing heat-insulating protective clothing and safety helmet. The navigation path data and the wearing instruction text are integrated to obtain the on-site handling procedure set.

[0036] The system invokes remote operation authorization credentials and on-site handling procedures, extracts core event information from high-confidence edge alarm events, sets alarm levels and distribution channels, encapsulates all the above information items, and generates multi-party linkage alarm commands.

[0037] Furthermore, the core information of the high-confidence edge alarm event is extracted, namely the event type L-02 and the location of pipe No. 3 in area B. The alarm level is set to the highest level according to the event type, and the distribution channel is set to the central control room terminal and the on-site inspector mobile application. All the above information items are encapsulated into a JSON data packet to generate a multi-party linkage alarm command.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-party alarm system based on cloud-edge collaboration, characterized in that, Includes the following steps: S1. Multimodal sensor state reliability calibration: Acquire microphone audio and thermal imaging video, calculate audio signal-to-noise ratio, clipping distortion, and video sharpness and saturation, and obtain the multimodal reliability scalar set through mapping matrix transformation; S2. Structured cross-modal feature vector generation: Based on the multimodal confidence scalar set, extract audio and video feature vectors, and connect the confidence scalars with the corresponding feature vectors in sequence to generate structured confidence feature vectors. S3. High-confidence event determination at the edge: Based on the structured confidence feature vector input to the edge device network, obtain the pipeline event confidence vector, determine whether the leakage and howling dimension exceeds the threshold, and if it does, encapsulate the timestamp, location and event identifier to establish a high-confidence edge alarm event. S4. Optimal Alarm Handling Team List Generation: Based on the retrieval demand vector of high-confidence edge alarm events, match the capability vector of each receiver, filter the optimal user combination, and obtain the optimal alarm handling team list. S5. Generation of multi-party linkage alarm commands: Call the optimal alarm handling team list and high-confidence edge alarm events, configure emergency shut-off valve control permissions, navigation paths and other tasks for team members and encapsulate them to generate multi-party linkage alarm commands.

2. The multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: The multimodal confidence scalar set includes audio confidence scalar and image confidence scalar. The structured cross-modal feature vector specifically includes audio confidence data segment, audio feature data segment, image confidence data segment, and image feature data segment. The high-confidence edge alarm event includes event type identifier, event occurrence timestamp, event spatial coordinates, and event association feature vector. The optimal alarm handling team list includes pre-selected user identifier, comprehensive matching score, and predetermined response role. The multi-party linkage alarm command specifically refers to remote control token, navigation path data, and security protocol command.

3. The multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: Step S1 includes: Audio signals from MEMS microphone arrays deployed in a specific area are collected, the ratio of audio signal power to background noise power is calculated in parallel, and the proportion of sample points with amplitude exceeding the preset upper limit in the audio signal is counted to obtain the audio signal quality index. The signal-to-noise ratio and clipping distortion in the audio signal quality indicators are called, and a preset two-dimensional mapping matrix is ​​retrieved. The table is looked up and transformed according to the correspondence in the matrix to obtain the audio reliability scalar. The video signal from the infrared thermal imaging sensor is acquired, and the sharpness and brightness saturation of the video signal are converted into image reliability scalars in parallel. The audio reliability scalar is called, and the two are integrated to generate a multimodal reliability scalar set.

4. A multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: S2 includes: Based on the multimodal confidence scalar set, the original audio signal contained therein is called, and a short-time Fourier transform is performed on the signal to obtain the time spectrum. The Mel-scale triangular filter bank is then used to filter the time spectrum, and a discrete cosine transform is performed on the logarithmic energy of each channel after filtering to obtain the Mel frequency cepstral coefficient vector. Based on the multimodal confidence scalar set, the original video signal contained therein is called. For the sequential image frames in the video signal, matrix convolution operation with preset convolution kernel is performed frame by frame. Max pooling operation is performed on the output after convolution operation to downsample. The processing results of multiple frames are concatenated to generate convolution feature map vector. The audio and image reliability scalars from the multimodal reliability scalar set are called, along with the Mel frequency cepstral coefficient vector and the convolutional feature map vector. The audio reliability scalar, Mel frequency cepstral coefficient vector, image reliability scalar, and convolutional feature map vector are concatenated end-to-end according to a preset dimensional order to establish a structured reliability feature vector.

5. A multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: S3 includes: For the structured confidence feature vector, it is input into the network structure inside the edge computing device. In this structure, matrix multiplication is performed with the preset weight matrix, and the result is processed by a nonlinear activation function. The processed data is then passed sequentially in the stacked structure to obtain an event classification confidence vector for the state of the high-pressure steam pipeline. Call the event classification confidence vector, and set the event trigger threshold according to the receiver operation characteristic curve. Determine whether the dimension value corresponding to the leakage howling sound in the vector exceeds the event trigger threshold. Convert the judgment result to a Boolean value and generate an alarm trigger status identifier. Based on the alarm trigger status identifier, if it indicates a triggered state, the current timestamp and location coordinates are retrieved from the device and encapsulated together with the preset event type identifier and structured confidence feature vector to establish a high-confidence edge alarm event.

6. A multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: S4 includes: Based on the event type identifier in the high-confidence edge alarm event, the event demand vector is retrieved from the cloud knowledge base, and the response capability vector uploaded by each receiving terminal, which consists of dynamic location parameters, activity status scalar, historical processing efficiency parameters and skill weight parameters, is called. A weighted summation operation is performed on the response capability vector and the event demand vector of each receiving party to obtain the matching degree of each receiving party. Based on the matching degree of each recipient and the number of people required for the event, all recipients that meet the number of people conditions are selected and user groups are generated. For each user group, the sum of the skill weight parameters of its members is calculated, and the comprehensive geographical distance between it and the event location is calculated based on the dynamic location parameters to establish a candidate handling team plan. The candidate handling team plan is invoked, and all user combinations in the plan list are iteratively compared. The sorting and selection operations are performed based on the dual screening criteria of maximizing the sum of skill weights and minimizing the comprehensive geographical distance, to determine the unique user combination that meets the conditions and obtain the optimal alarm handling team list.

7. A multi-party alarm system based on cloud-edge collaboration according to claim 1, characterized in that: S5 includes: The system calls up the optimal alarm handling team list and high-confidence edge alarm events, retrieves the preset permission matrix for the central control room operators in the list, and generates a remote operation authorization credential associated with the emergency shut-off valve based on the operator's user ID and the event type identifier. Based on the optimal alarm handling team list and high-confidence edge alarm events, a navigation path is generated for the on-site inspectors in the list by calling the map service interface, and the wearing instructions for related personal protective equipment are retrieved from the device safety protocol library. The two are then integrated into a set of on-site handling procedures. The system invokes remote operation authorization credentials and on-site handling procedures, extracts core event information from high-confidence edge alarm events, sets alarm levels and distribution channels, encapsulates all the above information items, and generates multi-party linkage alarm commands.

Images