Intelligent linkage control method and system of automatic fire-fighting facilities
The fire protection system, which integrates cross-modal feature fusion and dynamic parameter reconstruction, solves the problems of inaccurate fire situation extraction and insufficient adaptability in traditional fire linkage control, and achieves early warning sensitivity and efficient response throughout the entire life cycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 韦天常
- Filing Date
- 2026-03-25
- Publication Date
- 2026-06-05
Smart Images
Figure CN122157458A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent fire protection and automated control technology, and in particular to an intelligent linkage control method and system for automatic fire protection facilities. Background Technology
[0002] With the increasing complexity of modern building structures and spaces, the difficulty of early fire prevention and emergency response has also significantly increased. Automated fire suppression systems, as core infrastructure for protecting life and property, play a crucial role in fire early warning and disaster relief. In conventional fire emergency response procedures, the system primarily relies on physical detection terminals distributed throughout various building zones to continuously sense abnormal environmental fluctuations. Upon confirming a fire, it automatically triggers low-level fire suppression equipment such as smoke exhaust fans, fireproof roller shutters, and fire extinguishing devices based on pre-entered rule matrices within the system. This aims to effectively block the spread of fire in its early stages and buy valuable time for safe evacuation.
[0003] However, existing fire-fighting linkage control mechanisms have gradually revealed significant limitations in dealing with complex and ever-changing disaster scenarios. On the one hand, traditional systems rely heavily on single and fixed structured sensor signals, lacking the ability to deeply analyze massive amounts of unstructured audio-visual data at fire scenes. When faced with alarm voices accompanied by strong background noise and complex dialects, as well as on-site images containing a large amount of irrelevant environmental interference, they struggle to accurately and efficiently extract core fire elements, leading to delays and misjudgments in early warning. On the other hand, existing equipment often employs static and rigid control rules, failing to adapt to the long-term dynamic evolution of the building's physical environment and functional requirements, easily resulting in a severe decline in control effectiveness. Summary of the Invention
[0004] Therefore, it is necessary to provide an intelligent linkage control method and system for automatic fire-fighting facilities that can accurately integrate multimodal early warning information from complex sites and realize the dynamic evolution of control logic under absolutely safe physical boundary constraints, in order to address the above-mentioned technical problems.
[0005] In a first aspect, this application provides an intelligent linkage control method for automatic fire-fighting facilities, including:
[0006] S1. Acquire the original fire alarm voice signal and on-site context information; perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence; filter candidate keywords in the preliminary text data sequence to obtain certain keywords; perform entity relationship binding processing based on certain keywords and on-site context information to generate a text feature sequence.
[0007] S2. Acquire the fire scene image stream, remove irrelevant visual background from the fire scene image stream based on the text feature sequence to obtain the visual features to be fused, and perform feature mapping on the text feature sequence and the visual features to be fused to obtain high-order cross-modal fusion features; perform feature dimensionality reduction on the high-order cross-modal fusion features to generate a structured control alarm data stream;
[0008] S3. Based on the structured control alarm data stream, solve the optimization problem of the physical safety boundary that satisfies the control barrier function, generate a legal safety action sequence, and broadcast the legal safety action sequence; wherein, the legal safety action sequence is used to drive the underlying fire-fighting equipment to perform linkage fire extinguishing actions, and the control barrier function is used to define the set of safety states that the system must satisfy to ensure that the system always stays within the set of safety states;
[0009] S4. Fit the probability distribution of the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; calculate the distribution divergence of the actual observation distribution to quantify the decay offset of the currently used linkage control parameter matrix; when the decay offset exceeds the decay offset threshold, reconstruct the parameters to generate a candidate strategy matrix; generate an updated linkage control parameter matrix based on the candidate strategy matrix; wherein, the updated linkage control parameter matrix is used to replace the currently used linkage control parameter matrix to take over the subsequent fire linkage control.
[0010] Secondly, this application also provides an intelligent linkage control system for automatic fire-fighting facilities, used to implement the method described in the first aspect, the system comprising:
[0011] The acoustic probability analysis module is used to acquire the original fire alarm voice signal and the on-site context state information, perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence; filter the candidate keywords in the preliminary text data sequence to obtain the certain keywords; and perform entity relationship binding processing based on the certain keywords and the on-site context state information to generate a text feature sequence.
[0012] The cross-modal fusion module is used to acquire fire scene image streams, remove irrelevant visual backgrounds from the fire scene image streams based on text feature sequences to obtain visual features to be fused, and perform feature mapping on the text feature sequences and visual features to be fused to obtain high-order cross-modal fusion features; feature dimensionality reduction is performed on the high-order cross-modal fusion features to generate a structured control alarm data stream;
[0013] The safety linkage decision module is used to solve the optimization problem of physical safety boundary that satisfies the control barrier function based on the structured control alarm data stream, generate a legal safety action sequence, and broadcast the legal safety action sequence;
[0014] The linkage logic evolution module is used to fit the probability distribution of the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; to calculate the distribution divergence of the actual observation distribution and quantify the decay offset of the currently used linkage control parameter matrix; to reconstruct the parameters when the decay offset exceeds the decay offset threshold and generate a candidate strategy matrix; and to generate an updated linkage control parameter matrix based on the candidate strategy matrix.
[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an intelligent linkage control method for automatic fire-fighting facilities as described in the first aspect.
[0016] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an intelligent linkage control method for automatic fire-fighting facilities as described in the first aspect.
[0017] The aforementioned intelligent linkage control method and system for automatic fire-fighting facilities, through cross-modal feature fusion of alarm voice and on-site image streams, can broaden the system's perception dimension in the early stages of a fire, transforming unstructured audiovisual information into structured control alarms and improving the system's early warning sensitivity. By calculating the distribution divergence of feedback time-series data to quantify decay offset and reconstructing parameters when boundaries are exceeded, adaptive monitoring and updating of the underlying linkage logic can be achieved. Simultaneously, by combining the physical safety boundaries limited by control barrier functions, a safe and gradual evolution of control strategies can be achieved while maintaining compliant operation. This can alleviate the logic drift problem caused by traditional static rules, thereby improving the comprehensive response efficiency of the fire protection system throughout its entire lifecycle. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating an intelligent linkage control method for automatic fire-fighting facilities according to an embodiment of the present invention.
[0020] Figure 2 This is a schematic diagram of the structure of an intelligent linkage control system for automatic fire-fighting facilities according to one embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0022] In one embodiment, such as Figure 1 As shown, an intelligent linkage control method for automatic fire-fighting facilities is provided. This embodiment illustrates the application of this method to a control system. It is understood that this method can also be applied to a server, and further to a system including both a control system and a server, and is implemented through the interaction between the control system and the server. In this embodiment, the method includes the following steps:
[0023] S1. Acquire the original fire alarm voice signal and on-site context information, perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence; filter the candidate keywords in the preliminary text data sequence to obtain the confirmed keywords; perform entity relationship binding processing based on the confirmed keywords and on-site context information to generate a text feature sequence.
[0024] Specifically, the original fire alarm voice signal can be fire reporting data represented as audio electrical signals acquired by voice acquisition equipment. On-site contextual state information can be variable values indicating the physical environment state. Candidate keywords can be phrases with potential fire indication significance extracted from text data sequences. Confirmed keywords can be words determined to have indicative significance after probabilistic inference.
[0025] Optionally, the control system can acquire audio signal streams and environmental variables from the physical environment, and apply Fast Fourier Transform to convert time-domain sound waves into two-dimensional spectrogram signals in the frequency domain. The control system can map acoustic emission probabilities using a deep neural network and, combined with the state transition calculation of a Hidden Markov Model, extract candidate keywords from the initial text data sequence. The control system can then combine a language database and, based on Bayesian inference, calculate the posterior probability of each candidate keyword under given frequency domain features. The control system can select the word corresponding to the extreme value of the posterior probability score as the confidence keyword based on the maximum a posteriori probability rule. Finally, the control system can input the confidence keywords and environmental state information into a Long Short-Term Memory network, establish spatial and temporal dependencies between words according to dependency syntax rules, and generate a text feature sequence.
[0026] S2. Acquire the fire scene image stream, remove irrelevant visual background from the fire scene image stream based on the text feature sequence to obtain the visual features to be fused, and perform feature mapping on the text feature sequence and the visual features to be fused to obtain high-order cross-modal fusion features; perform feature dimensionality reduction on the high-order cross-modal fusion features to generate a structured control alarm data stream.
[0027] Specifically, the irrelevant visual background can be a pixel region within a fire scene image stream that does not contain burning or changes in environmental features. Higher-order cross-modal fusion features can be matrix representations that simultaneously include semantic information and visual spatial distribution.
[0028] Optionally, the control system can segment the fire scene image stream into independent tiles and extract spatial feature vectors from these tiles. The control system can calculate the cosine similarity between the tile spatial feature vectors and the text feature sequence, and then calculate the variance. Based on the variance values, the control system can retain the top-ranked independent tiles as visual features to be fused. The control system can use the text feature sequence as query input and the visual features to be fused as key-value inputs, calculating the image-text correlation through an asymmetric attention mechanism matrix and generating high-order cross-modal fusion features. The control system can input these high-order cross-modal fusion features into a retrieval enhancement model, performing a multi-dimensional nearest neighbor search in a static benchmark knowledge base composed of national fire protection standards and building topology. The control system can match context variables in a dynamic knowledge base including environmental sensor data, combining the retrieval results from the static benchmark knowledge base and the dynamic knowledge base to output a structured control alarm data stream with parameter fields.
[0029] S3. Based on the structured control alarm data stream, solve the optimization problem of the physical safety boundary that satisfies the control barrier function, generate a legal safety action sequence, and broadcast the legal safety action sequence; wherein, the legal safety action sequence is used to drive the underlying fire-fighting equipment to perform linkage fire extinguishing actions, and the control barrier function is used to define the set of safety states that the system must satisfy to ensure that the system always stays within the set of safety states.
[0030] Specifically, the control barrier function can be a mathematical function used to restrict the operating state within a specified area. The physical safety boundary can be an operational limit threshold determined by the equipment's operating specifications. The legal safety action sequence can be a set of control instructions that satisfy the system's physical constraints.
[0031] Optionally, the control system can determine the necessary safety state set conditions based on the control obstacle function and continuously differentiable boundary function defined in the specifications. The control system can input the structured control alarm data stream into a reinforcement learning network to obtain the initial linkage actions. The control system can construct a quadratic programming optimization calculation model with Lie derivative constraints, calculating the action variable values that minimize the objective function without violating the monotonically increasing extended K-type function constraints. The control system can map the initial linkage actions to generate a legal safety action sequence. The control system can then distribute the legal safety action sequence to the underlying fire control terminal via a communication bus, driving the fans and pumps to execute the corresponding linkage fire extinguishing actions.
[0032] S4. Fit the probability distribution of the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; calculate the distribution divergence of the actual observation distribution to quantify the decay offset of the currently used linkage control parameter matrix; when the decay offset exceeds the decay offset threshold, reconstruct the parameters to generate a candidate strategy matrix; generate an updated linkage control parameter matrix based on the candidate strategy matrix; wherein, the updated linkage control parameter matrix is used to replace the currently used linkage control parameter matrix to take over the subsequent fire linkage control.
[0033] Specifically, feedback time-series data can be a set of environmental response physical indicators recorded over time after the device performs an action. Actual observed distribution can be a statistical representation of the current operating performance of the control system. Distribution divergence can be an indicator that quantifies the difference between two probability distributions. Fading offset can be a numerical representation of the current performance of the control system deviating from the initial design. Fading offset threshold can be the maximum allowable fading limit of the control system. Candidate strategy matrix can be a set of control parameters that have not yet been taken over by the control system. Linkage control parameter matrix can be an array of parameters executing commands in the current control system.
[0034] Optionally, the control system can calculate the actual observed distribution of the feedback time-series data. The control system can retrieve a set benchmark distribution and quantify the divergence between the actual observed distribution and the benchmark distribution using the relative entropy calculation formula. The control system can calculate the divergence-standardized residuals generated at each time point. The control system can calculate the decay offset based on the adaptive cumulative sum algorithm and the exponentially weighted least squares regression method. The control system can generate a candidate policy matrix when the decay offset exceeds the decay offset threshold. The control system can construct a Gaussian process surrogate model with the candidate policy matrix as input. The control system can collect sparse feedback test data and update the mean and variance of the Gaussian process according to the Bayesian posterior formula. The control system can calculate the upper confidence bound acquisition function to obtain the expected return and cognitive uncertainty variance. The control system can generate an updated linkage control parameter matrix and take over fire control when the expected return is greater than the original policy return and the convergence condition is met.
[0035] In one embodiment of the intelligent linkage control method for automatic fire-fighting facilities described above, cross-modal feature fusion of alarm voice and on-site image streams broadens the system's perception dimension in the early stages of a fire, transforming unstructured audiovisual information into structured control alarms and improving the system's early warning sensitivity. By calculating the distribution divergence of feedback time-series data to quantify decay offset and reconstructing parameters when boundaries are exceeded, adaptive monitoring and updating of the underlying linkage logic can be achieved. Simultaneously, by combining the physical safety boundaries limited by control barrier functions, a safe and gradual evolution of control strategies can be achieved while maintaining compliant operation. This alleviates the logic drift problem caused by traditional static rules, thereby improving the comprehensive response efficiency of the fire protection system throughout its entire lifecycle.
[0036] In one embodiment, S1 may include:
[0037] S11. The original fire alarm voice signal is converted into the frequency domain by fast Fourier transform to obtain the spectrogram signal. The spectrogram signal is then acoustically analyzed based on the underlying acoustic model to obtain the preliminary text data sequence.
[0038] Specifically, the spectrogram signal can be an acoustic representation matrix carrying frequency and time parameters. The underlying acoustic model can be a probabilistic computation architecture including hidden Markov models and deep neural networks. The initial text data sequence can be a collection of characters generated based on acoustic feature maps.
[0039] Optionally, the control system can calculate the frequency domain parameters of the original fire alarm voice signal based on the Fast Fourier Transform algorithm, and generate a spectrogram signal based on the frequency domain parameters. The control system can configure the node weight parameters of the underlying acoustic model, input the spectrogram signal into the underlying acoustic model, and activate the deep neural network in the underlying acoustic model to calculate the emission probability between homophones of the voice features. The control system can calculate the transition probability between voice states using a Hidden Markov Model in the underlying acoustic model. The control system can combine the emission probability and the state transition probability to establish an acoustic coherence model, decode the acoustic coherence model using the Viterbi decoding algorithm, and output a preliminary text data sequence.
[0040] S12. Extract candidate keywords from the preliminary text data sequence; based on the Gaussian mixture model, use Bayesian inference to perform probability inference on the candidate keywords to obtain the posterior probability corresponding to the candidate keywords; based on the posterior probability, filter the candidate keywords according to the maximum a posteriori probability rule to obtain the confirmed keywords.
[0041] For example, the formula for calculating the posterior probability can be:
[0042]
[0043] in, Indicates the signal in a given spectrogram Under the conditions, the first Candidate keywords The posterior probability of occurrence; For the first One candidate keyword; For spectrogram signals; Indicates the assumption that the first The likelihood probability of generating a spectrogram signal when a candidate keyword occurs; Indicates the first The prior probabilities of each candidate keyword in a pre-built language database; Represents spectrogram signals The normalized marginal probability of occurrence.
[0044] Specifically, a Gaussian mixture model can be a probabilistic model composed of multiple Gaussian distribution functions. Bayesian inference is a method for calculating the probability of an event occurring based on prior probability and likelihood probability.
[0045] Optionally, the control system can remove stop words without business indication function from the initial text data sequence and extract candidate keywords with business indication function. The control system can construct a Gaussian mixture model to model and calculate the acoustic distribution of the spectrogram signal. The control system can query and extract the prior probability of each candidate keyword from the language database and call the Gaussian mixture model to calculate the likelihood probability of generating the current spectrogram signal assuming the occurrence of the candidate keyword. The control system can calculate the marginal probability of the current spectrogram signal occurring. The control system can calculate the product of the prior probability and the likelihood probability based on the Bayesian inference formula and divide the product result by the marginal probability. The control system can calculate the posterior probability corresponding to each candidate keyword and sort the candidate keywords in descending order according to the posterior probability values. The control system can extract the word at the top of the sorted list based on the maximum a posteriori probability rule and set the extracted word as the sure keyword.
[0046] S13. Based on the confirmed keywords and the on-site context information, the spatiotemporal dependency entity relationship between words is bound through the Long Short-Term Memory Network to generate a text feature sequence.
[0047] Specifically, a long short-term memory network can be a recurrent neural network architecture with memory units.
[0048] Optionally, the control system can convert the certainty keywords and on-site context state information into word vectors, and input the word vectors into a Long Short-Term Memory (LSTM) network. The control system can control the retention or discarding of historical state information through a forget gate. The control system can calculate the write value of the current input information based on the input gate. The control system can update the state of the core memory units of the LTM network and calculate the hidden layer feature matrix through the output gate. The control system can match the association constraints between the certainty keywords and on-site context state information based on dependency syntax rules. The control system can calculate the spatiotemporal dependency entity relations containing physical location attributes and event occurrence attributes, and concatenate the word vectors according to the spatiotemporal dependency entity relations to generate a text feature sequence reflecting semantic associations.
[0049] In an embodiment of the intelligent linkage control method for automatic fire-fighting facilities described above, frequency domain conversion and acoustic analysis of the original fire alarm voice signal enable the extraction and mapping of sound wave features. By applying Gaussian mixture models and Bayesian inference to calculate the posterior probability of candidate keywords and performing extreme value screening, probability correction and data filtering for dialect pronunciation deviations and background noise can be achieved. By binding spatiotemporal dependency entity relationships through long short-term memory networks, discrete alarm words can be transformed into logically related text feature sequences. Combining these methods improves the system's noise resistance in extracting early warning information in emergency scenarios and increases the success rate of converting unstructured voice data into usable fire-fighting clues.
[0050] In one embodiment, S2 may include:
[0051] S21. Extract visual features from the fire scene image stream to obtain tile space feature vectors; calculate the cosine similarity of the tile space feature vectors based on the text feature sequence to obtain a similarity vector, and calculate the similarity variance based on the similarity vector; perform saliency screening on the tile space feature vectors based on the similarity variance to obtain the visual features to be fused.
[0052] Specifically, visual features can be matrix representations characterizing the local and global pixel distribution of an image. Patch spatial feature vectors can be dimensionality-reduced mathematical expressions generated after dividing the image into grid regions. The visual features to be fused can be local image vectors retained for subsequent modality alignment calculations.
[0053] For example, the formula for calculating the similarity variance can be:
[0054]
[0055] in, Indicates the first The similarity variance of each image slice patch; This represents the index of each semantic dimension extracted from the text feature sequence; This represents the total number of text segments contained in the text feature sequence; Indicates the first The spatial feature vector of the i-th map patch and the i-th text feature sequence Cosine similarity values between text segmentation features; Indicates the first Each tile and all The arithmetic mean of the cosine similarity values between the text segmentation features.
[0056] Optionally, the control system can extract features from the fire scene image stream using a visual feature encoding network, segmenting the image into multiple tile blocks to generate tile spatial feature vectors reflecting spatial characteristics. The control system can calculate the cosine similarity between each tile spatial feature vector and the text feature sequence based on tensor multiplication. The control system can combine the cosine similarity values of different text feature sequences corresponding to the same tile into a similarity vector. The control system can calculate the sum of squared deviations of each value from the mean based on the similarity vector to obtain the similarity variance. The control system can sort the tile spatial feature vectors according to the similarity variance values, extract the top-ranked tile spatial feature vectors, and set the extracted tile spatial feature vectors as the visual features to be fused.
[0057] S22. Based on the dynamic attention computing model, asymmetric attention feature mapping is performed on the text feature sequence and the visual features to be fused in the unified text-image semantic latent space to obtain high-order cross-modal fusion features.
[0058] For example, the formula for calculating high-order cross-modal fusion features can be:
[0059]
[0060] in, This represents a high-order cross-modal fusion feature; For text feature sequences; Visual features to be fused; For query matrix; The key matrix; It is a value matrix; Scaling factor This indicates the corresponding key. Feature dimensions; This is the activation function.
[0061] Specifically, a dynamic attention computation model can be a computational architecture that assigns specific feature weights based on a query matrix, key matrix, and value matrix. A unified text-image semantic latent space can be a mathematical multidimensional space that accommodates comparisons of features from different modalities. Asymmetric attention feature mapping can be a feature transformation operation that extracts information from one modality as the query subject from another. High-order cross-modal fusion features can be multidimensional tensor data that contains location concepts and visual representations.
[0062] Optionally, the control system can set dimensional parameters for a unified text-image semantic latent space. The control system can apply a query matrix structure to the text feature sequence and a key-value matrix structure to the visual features to be fused. The control system can calculate the query feature matrix corresponding to the text feature sequence based on the dimensional parameters, and can calculate the key feature matrix and value feature matrix corresponding to the visual features to be fused based on the dimensional parameters. The control system can perform a matrix inner product operation of the transpose of the query feature matrix and the key feature matrix, and can introduce a scaling factor to numerically scale the matrix inner product result. The control system can normalize the scaled values using an activation function to obtain the attention weight matrix. The control system can perform multiplication operations on the attention weight matrix and the value feature matrix to generate a high-order cross-modal fusion feature containing relevance indicators.
[0063] S23. Based on the dual-track knowledge base in the retrieval-enhanced large model, feature dimensionality reduction is performed on the high-order cross-modal fusion features through multi-dimensional nearest neighbor search to generate a structured control alarm data stream. The dual-track knowledge base includes a static baseline knowledge base and a real-time dynamic knowledge base. The static baseline knowledge base is used to provide long-term stable rule constraints, including building topology and physical equipment coordinate matrices. The real-time dynamic knowledge base is used to provide transient environmental context, including environmental sensing time-series data and personnel flow heatmaps.
[0064] Specifically, a retrieval-enhanced generative model can be a deep learning architecture that combines external database information to perform text generation. Multidimensional nearest neighbor search can be a query action that finds the minimum distance value in a high-dimensional vector space. Structured control alarm data streams can be instruction messages with standardized formats that can be parsed by hardware. A static benchmark knowledge base can be a database storing building topology and physical equipment coordinate matrices. A real-time dynamic knowledge base can be a database storing environmental sensor time-series data and heatmaps of personnel flow. Transient environmental context can be real-time measured variables reflecting the state of the physical space.
[0065] Optionally, the control system can input high-order cross-modal fusion features for retrieval enhancement to generate a large model. It can query and retrieve building topology and physical equipment coordinate matrices from a static benchmark knowledge base, and query and retrieve environmental sensor time-series data and personnel flow heatmaps from a real-time dynamic knowledge base. The control system can calculate the spatial distance between the high-order cross-modal fusion features and the embedded vectors in the dual-track knowledge base. The control system can perform feature dimensionality reduction on the high-order cross-modal fusion features based on multi-dimensional nearest neighbor search matching feature information, combined with long-term stable rule constraints and transient environmental context. The control system can eliminate duplicate alarm data in the information dimension and compile and output a structured control alarm data stream according to a formatted template.
[0066] In the embodiment of the intelligent linkage control method for automatic fire-fighting facilities described above, by calculating the similarity variance of the feature vectors in the map space and performing saliency screening, irrelevant visual backgrounds in the fire scene image stream can be removed, improving the targeting of visual feature extraction. Based on a dynamic attention calculation model, asymmetric attention feature mapping enables feature alignment calculation of textual and visual features within a unified text-image semantic latent space, generating high-order cross-modal fusion features reflecting multi-dimensional parameters of the scene. By introducing a retrieval-enhanced generation model and a dual-track knowledge base to perform multi-dimensional nearest neighbor search, combined with static building topology and real-time environmental sensing data, dimensionality reduction calculation and alarm redundancy removal of high-order cross-modal fusion features can be achieved. These steps enable the standardized conversion of unstructured alarm data into structured control alarm data streams, improving the efficiency of fire protection systems in parsing and executing complex environmental disaster information.
[0067] In one embodiment, S3 may include:
[0068] S31. Based on a security reinforcement learning network, perform reinforcement learning inference on the structured control alarm data stream to obtain the original linkage action sequence.
[0069] Specifically, a security reinforcement learning network can be an artificial intelligence model that combines a reward function maximization mechanism with a physical security constraint mechanism to perform optimization calculations. The original sequence of linked actions can be a set of initial device start-stop instructions directly generated by the neural network without being filtered or intercepted by the underlying physical security constraints.
[0070] Optionally, the control system loads a pre-trained safety reinforcement learning network parameter matrix into memory. The control system can extract event type attributes and spatial coordinate attributes from the structured control alarm data stream, converting them into tensor formats required by the input layer of the safety reinforcement learning network. The control system can compute the tensors and drive the hidden layer nodes to perform reinforcement learning inference. The control system can calculate negative reward values related to the current fire spread area and positive reward values related to the smoke exhaust clearance height, based on a set reward function. The control system can output a state-action value matrix based on maximizing the cumulative reward function, and match corresponding physical device start and stop commands based on the state-action value matrix. The control system can combine multiple physical device start and stop commands to generate an original linkage action sequence.
[0071] S32. Obtain the set physical operating boundary, set the safety state set of the physical operating boundary based on the continuously differentiable function, and obtain the physical safety boundary of the control obstacle function.
[0072] Specifically, the physical operating boundary can be the limit operating state parameters determined by national building fire protection design standards and the mechanical properties of the underlying electromechanical equipment. The safety state set setting can be a computational operation based on mathematical formulas to delineate the set of legal parameters that allow the system to exist in a multi-dimensional state space. The control barrier function can be a mathematical barrier model used to force the system's state trajectory to always remain within a specified safety set. The physical safety boundary can be a virtual barrier limit formed by the representation of the mathematical properties of the control barrier function in the system's state space.
[0073] Optionally, the control system can construct physical operating boundary parameters based on building fire protection design data and the factory-calibrated limit speed and pressure threshold data of electromechanical equipment. The control system can substitute these physical operating boundary parameters into a continuously differentiable function to calculate and construct the safety state set setting conditions within a multi-dimensional state space. Based on the safety state set setting conditions, the control system can define the constraint equations of the control obstacle function and calculate the gradient vector and Lie derivative values of the control obstacle function. Based on the Lie derivative values, the control system can establish the physical safety boundary limit parameters in the state space, generating the physical safety boundary.
[0074] S33. Based on the physical safety boundary, perform secondary planning optimization on the original linkage action sequence to generate a legal and safe action sequence, and broadcast the legal and safe action sequence.
[0075] For example, the formula for calculating a legal and safe action sequence can be:
[0076]
[0077]
[0078] in, This is a legal and safe sequence of actions; This represents the operation of finding the variable values that minimize the objective function. This represents the candidate control inputs during the optimization process. Indicates the set of allowed actions; This is the original sequence of linked actions; This represents the constraints that must be satisfied to solve a quadratic programming problem. Represents the transpose of the gradient of a continuously differentiable function; This is the current state; This represents the intrinsic state evolution term in the system dynamics function that is unaffected by the control input; This represents the control gain term in the system dynamics function that is affected by the control input; This represents a monotonically increasing extended K-class function corresponding to a continuously differentiable function.
[0079] Specifically, quadratic programming optimization can be a process of finding variables that minimize the quadratic objective function within constraints. Candidate control inputs can be command variables to be verified during the optimization process. Inherent state evolution terms can be internal state evolution values in the system dynamics equations that are unaffected by external control signals. Control gain terms can be state evolution weights in the system dynamics equations driven by external control signals. Monotonically increasing extended K-type functions can be mathematical constraint functions used to adjust the braking buffer rate when the control system approaches a dangerous boundary.
[0080] Optionally, the control system can establish a quadratic programming objective function aimed at finding the candidate control input with the minimum spatial distance from the original linkage action sequence. The control system can extract the intrinsic state evolution term and control gain term from the system dynamics function. The control system can calculate the gradient transpose matrix of continuously differentiable functions and the constraint values of monotonically increasing extended K-type functions. The control system can construct Lie derivative inequality constraints including the gradient transpose matrix, intrinsic state evolution term, control gain term, and monotonically increasing extended K-type functions. Under the Lie derivative inequality constraints, the control system can perform quadratic programming optimization, outputting the variable values that minimize the quadratic objective function and generating a legal safety action sequence. The control system can perform mathematical projection correction on out-of-bounds instructions in the original linkage action sequence, encapsulating the legal safety action sequence into communication data packets. The control system can broadcast the encapsulated communication data packets via the controller area network bus, driving the underlying fire relays and programmable logic controllers to execute corresponding equipment start-up and shutdown actions.
[0081] In one embodiment of the intelligent linkage control method for automatic fire-fighting facilities described above, reinforcement learning inference of structured control alarm data streams based on a security reinforcement learning network enables dynamic generation and calculation of underlying physical linkage strategies based on environmental context. By extracting physical operational boundary parameters and substituting them into continuously differentiable functions to calculate control obstacle functions, quantifiable settings of physical safety boundaries in the system state space can be achieved. Establishing Lie derivative inequality constraints based on physical safety boundaries and performing quadratic programming optimization can realize the mathematical projection of dangerous actions that exceed boundaries onto compliant and safe action surfaces. Applying this combination of methods can improve the compliance of commands output by the artificial intelligence model in the fire-fighting field and enhance the hardware operational safety of the system when performing strategy optimization under complex disaster conditions.
[0082] In one embodiment, S4 may include:
[0083] S41. Fit the probability distribution of the feedback time series data to obtain the actual observed distribution of system performance. Calculate the distribution offset of the actual observed distribution of system performance based on the set system benchmark distribution to obtain the divergence standardized residual.
[0084] Specifically, probability distribution fitting can be a mathematical calculation process that derives a probability density function based on discrete data. The actual observed distribution of system performance can be a probabilistic model reflecting the system's operating state, generated based on collected data. The system baseline distribution can be a reference probability model initially set by the system, reflecting a predetermined operating state.
[0085] For example, the formula for calculating the divergence-standardized residual can be:
[0086]
[0087]
[0088] in, The relative entropy of the actual observed distribution with respect to the system's baseline distribution is called the Kullback-Leibler divergence. This represents the actual observed distribution; This serves as the system's baseline distribution. Represents a single discrete state variable in the state space; Represents the set of state spaces; This represents the corresponding state variable. The probability of it occurring in the actual observed distribution; This represents the corresponding state variable. The probability of it appearing in the system's baseline distribution;
[0089] The Jensen-Shannon divergence is a measure of the smooth divergence between the actual observed distribution and the system's baseline distribution. This represents the average mixed distribution of the actual observed distribution and the system baseline distribution; It represents the relative entropy of the actual observed distribution relative to the average mixed distribution; It represents the relative entropy of the system's baseline distribution relative to the average mixed distribution.
[0090] Optionally, the control system can calculate the statistical characteristics of the feedback time-series data based on a probability distribution fitting algorithm to generate the actual observed distribution of system performance. The control system can then use the relative entropy divergence formula or the Jensen-Shannon divergence formula to perform integration by substituting the actual observed distribution of system performance and the system baseline distribution into the divergence formula to obtain the distribution offset value. Finally, the control system can process the distribution offset value using a residual normalization formula to output the divergence-standardized residual.
[0091] S42. Based on the adaptive accumulation and monitoring algorithm, the divergence standardized residual is sequentially estimated in time to quantify the decay offset of the underlying linkage logic, and a drift alarm signal is generated when the decay offset exceeds the decay offset threshold. The adaptive accumulation and monitoring algorithm is used to sequentially monitor small and continuous systematic drift under the condition of random environmental interference, and the underlying linkage logic is used to determine the benchmark response rules for the current control of the underlying fire protection equipment to perform linkage actions.
[0092] Specifically, temporal sequential estimation can be a mathematical operation that estimates parameters one by one according to the time sequence of data generation. Low-level linkage logic can be the baseline response rule that determines the linkage action executed by the current control of the low-level fire-fighting equipment. Systemic drift can be a continuous deviation of system performance parameters over a time span.
[0093] Optionally, the control system can estimate the drift parameter in the divergence standardized residual based on the exponentially weighted least squares regression method. The control system can substitute the drift parameter into the iterative formula of the adaptive cumulative sum monitoring algorithm to perform time-series sequential estimation calculations under the condition of eliminating random disturbance variables. The control system can quantitatively calculate the decay offset of the underlying linkage logic, compare the magnitude of the decay offset with the decay offset threshold, and generate a drift alarm signal when the decay offset exceeds the decay offset threshold.
[0094] S43. Based on the drift alarm signal, reconstruct the control parameters of the underlying linkage logic to obtain the candidate strategy matrix.
[0095] Optionally, the control system can extract environmental state parameters and feedback indicators attached to the drift alarm signal. The control system can stop using the underlying linkage logic that caused the systematic drift. The control system can input the environmental state parameters into the reinforcement learning network to calculate and generate the control parameters for the underlying linkage logic. The control system can matrix-arrange and reconstruct the control parameters to generate a candidate policy matrix.
[0096] S44. Acquire sparse feedback test data under test conditions without taking over the actual physical bus, and perform posterior prediction on the candidate policy matrix using the sparse feedback test data to obtain the expected return and cognitive uncertainty variance of the new control policy; wherein, the new control policy is used to provide candidate response rules to replace the underlying linkage logic after detecting performance degradation of the underlying linkage logic.
[0097] Specifically, sparse feedback test data can be environmental response parameters collected using a small number of samples during the test. Expected return can be a quantitative measure of the performance improvement predicted after the new control strategy is implemented. Cognitive uncertainty variance can be a numerical indicator reflecting the system's lack of confidence in predicting the effectiveness of the new control strategy.
[0098] For example, the formula for calculating expected return can be:
[0099]
[0100] in, This represents the candidate parameters corresponding to the candidate policy matrix. Expected returns; Represents the input vector and candidate parameters The transpose of the kernel function covariance matrix between them; Represents the input vector The inverse of its own kernel function covariance matrix; This represents the output observations of the sparse feedback test data; This represents the input vector. The prior mean function.
[0101] For example, the formula for calculating the variance of cognitive uncertainty can be:
[0102]
[0103] in, This represents the candidate parameters corresponding to the candidate policy matrix. and candidate parameters The variance of cognitive uncertainty; Candidate parameters With candidate parameters The kernel function covariance matrix between them; Represents the input vector With candidate parameters The kernel function covariance matrix between them.
[0104] Optionally, the control system can block the command transmission port of the actual physical bus, establishing a test state that does not take over the actual physical bus. The control system can collect sparse feedback test data in fire-fighting test mode to construct a Gaussian process model with a candidate policy matrix as input. The control system can update the probability distribution of the Gaussian process model using the sparse feedback test data. The control system can perform posterior prediction calculations on the candidate policy matrix using the sparse feedback test data, outputting the expected return of the new control policy. The control system can output the cognitive uncertainty variance, reflecting the cognitive boundary of the Gaussian process model.
[0105] S45. Based on the Bayesian confidence bound maximization evaluation algorithm, the expected return and cognitive uncertainty variance are jointly evaluated and verified. When the expected return meets the convergence condition and is greater than the original logical return, the corresponding candidate policy matrix is determined as the target parameter, and the updated linkage control parameter matrix is generated based on the target parameter.
[0106] Specifically, the Bayesian upper confidence bound maximization evaluation algorithm is an optimization algorithm that simultaneously maximizes the expected value and the uncertainty bound during the optimization process. Joint evaluation and verification can be a computational process that comprehensively compares multiple performance indicators to make a decision. The convergence condition can be a mathematical state where the cognitive uncertainty variance shrinks to near zero or below a set lower bound. The original logical gain can be the performance quantification value of the baseline response rule before systematic drift occurs.
[0107] Optionally, the control system can jointly evaluate and verify the expected return and the variance of cognitive uncertainty, read the set convergence conditions, and determine whether the expected return meets the convergence conditions. The control system can compare the numerical values of the expected return and the original logical return. When the expected return meets the convergence conditions and is greater than the original logical return, a parameter confirmation operation is performed, and the corresponding candidate strategy matrix is determined as the target parameter. The control system can perform format conversion calculations based on the target parameter. The control system can generate an updated linkage control parameter matrix based on the target parameter. The control system can replace the currently used linkage control parameter matrix with the updated linkage control parameter matrix and take over subsequent fire linkage control operations.
[0108] In one embodiment of the intelligent linkage control method for automatic fire protection facilities described above, by calculating the divergence between the actual observed distribution of system performance and the system baseline distribution, and introducing an adaptive accumulation and monitoring algorithm, the quantitative monitoring of minor systematic drift of the system under complex environmental disturbances can be achieved. By extracting sparse feedback test data under test conditions to perform posterior prediction of the candidate strategy matrix, and using a Bayesian upper confidence bound maximization evaluation algorithm for joint evaluation, the safe reconstruction and gradual takeover of the underlying linkage logic parameters can be achieved. This process endows the system's underlying rules with an unsupervised adaptive evolution mechanism, which can improve the continuous response performance of the fire protection system to changes in the building environment throughout its entire lifecycle.
[0109] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0110] Based on the same inventive concept, this application also provides a system for implementing the intelligent linkage control method for an automatic fire-fighting facility as described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the intelligent linkage control system for automatic fire-fighting facilities provided below can be found in the above-described limitations of the intelligent linkage control method for an automatic fire-fighting facility, and will not be repeated here.
[0111] In one exemplary embodiment, such as Figure 2 As shown, an intelligent linkage control system 100 for automatic fire protection facilities is provided to implement the methods in the above-described method embodiments. This system may include:
[0112] The acoustic probability analysis module 101 is used to acquire the original fire alarm voice signal and the on-site context state information, perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence, filter the candidate keywords in the preliminary text data sequence to obtain the confidence keywords, and perform entity relationship binding processing based on the confidence keywords and the on-site context state information to generate a text feature sequence.
[0113] The cross-modal fusion module 102 is used to acquire fire scene image streams, remove irrelevant visual backgrounds from the fire scene image streams based on text feature sequences to obtain visual features to be fused, and perform feature mapping on the text feature sequences and visual features to be fused to obtain high-order cross-modal fusion features; perform feature dimensionality reduction on the high-order cross-modal fusion features to generate a structured control alarm data stream.
[0114] The safety linkage decision module 103 is used to solve the optimization problem of physical safety boundary that satisfies the control barrier function based on the structured control alarm data stream, generate a legal safety action sequence, and broadcast the legal safety action sequence.
[0115] The linkage logic evolution module 104 is used to fit the probability distribution of the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; to calculate the distribution divergence of the actual observation distribution and quantify the decay offset of the currently used linkage control parameter matrix; to reconstruct the parameters when the decay offset exceeds the decay offset threshold and generate a candidate strategy matrix; and to generate an updated linkage control parameter matrix based on the candidate strategy matrix.
[0116] In one embodiment, the acoustic probability analysis module 101 may include:
[0117] The data sequence unit is used to perform frequency domain conversion of the original fire alarm voice signal using fast Fourier transform to obtain a spectrogram signal. Based on the underlying acoustic model, the spectrogram signal is acoustically analyzed to obtain a preliminary text data sequence.
[0118] The keyword unit is used to extract candidate keywords from the initial text data sequence. Based on the Gaussian mixture model, Bayesian inference is used to perform probabilistic inference on the candidate keywords to obtain the posterior probability of the candidate keywords. Based on the posterior probability, the candidate keywords are screened according to the maximum a posteriori probability rule to obtain the confirmed keywords.
[0119] The feature sequence unit is used to generate text feature sequences by binding the spatiotemporal dependency entity relationships between words through a long short-term memory network based on the confirmed keywords and the on-site context state information.
[0120] In one embodiment, the cross-modal fusion module 102 may include:
[0121] The visual feature fusion unit is used to extract visual features from the fire scene image stream to obtain a tile space feature vector; based on the text feature sequence, it calculates the cosine similarity of the tile space feature vector to obtain a similarity vector, and calculates the similarity variance based on the similarity vector; based on the similarity variance, it performs saliency screening on the tile space feature vector to obtain the visual features to be fused.
[0122] The high-order fusion feature unit is used to perform asymmetric attention feature mapping on text feature sequences and visual features to be fused within a unified text-image semantic latent space based on a dynamic attention computation model, thereby obtaining high-order cross-modal fusion features.
[0123] The control alarm unit is used to perform feature dimensionality reduction of high-order cross-modal fusion features by multi-dimensional nearest neighbor search based on the dual-track knowledge base in the large model of retrieval enhancement, and generate a structured control alarm data stream.
[0124] In one embodiment, the security linkage decision module 103 may include:
[0125] The original linkage action unit is used to perform reinforcement learning inference on the structured control alarm data stream based on the security reinforcement learning network to obtain the original linkage action sequence.
[0126] The safety boundary unit is used to obtain the set physical operating boundary. Based on the continuously differentiable function, the safety state set of the physical operating boundary is set to obtain the physical safety boundary of the control barrier function.
[0127] The safety action and broadcast unit is used to perform secondary planning optimization on the original linkage action sequence based on the physical safety boundary, generate a legal safety action sequence, and broadcast the legal safety action sequence.
[0128] In one embodiment, the linkage logic evolution module 104 may include:
[0129] The distribution offset unit is used to fit the probability distribution of the feedback time series data to obtain the actual observed distribution of system performance. Based on the set system benchmark distribution, the distribution offset of the actual observed distribution of system performance is calculated to obtain the divergence standardized residual.
[0130] The drift alarm unit is used to perform time-sequential estimation of the divergence standardized residual based on an adaptive accumulation and monitoring algorithm, quantize the decay offset of the underlying linkage logic, and generate a drift alarm signal when the decay offset exceeds the decay offset threshold.
[0131] The control parameter reconstruction unit is used to reconstruct the control parameters of the underlying linkage logic based on the drift alarm signal to obtain the candidate strategy matrix.
[0132] The revenue and variance unit is used to acquire sparse feedback test data in a test state without taking over the actual physical bus. The sparse feedback test data is used to perform posterior prediction on the candidate policy matrix to obtain the expected revenue and cognitive uncertainty variance of the new control policy.
[0133] The matrix update unit is used to jointly evaluate and verify the expected return and the variance of cognitive uncertainty based on the Bayesian confidence bound maximization evaluation algorithm. When the expected return meets the convergence condition and is greater than the original logical return, the corresponding candidate policy matrix is determined as the target parameter, and the updated linkage control parameter matrix is generated based on the target parameter.
[0134] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent linkage control method for an automatic fire-fighting facility as described above.
[0135] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0136] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0137] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.
Claims
1. A method for intelligent linkage control of automatic fire-fighting facilities, characterized in that, The method includes: S1. Acquire the original fire alarm voice signal and the on-site context state information; perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence; filter the candidate keywords in the preliminary text data sequence to obtain certain keywords; perform entity relationship binding processing based on the certain keywords and the on-site context state information to generate a text feature sequence. S2. Acquire a fire scene image stream, remove irrelevant visual background from the fire scene image stream based on the text feature sequence to obtain visual features to be fused, and perform feature mapping on the text feature sequence and the visual features to be fused to obtain high-order cross-modal fusion features; perform feature dimensionality reduction on the high-order cross-modal fusion features to generate a structured control alarm data stream; S3. Based on the structured control alarm data stream, solve the optimization problem of the physical safety boundary that satisfies the control barrier function, generate a legal safety action sequence, and broadcast the legal safety action sequence; wherein, the legal safety action sequence is used to drive the underlying fire-fighting equipment to perform linkage fire-fighting actions, and the control barrier function is used to define the set of safety states that the system must satisfy to ensure that the system always stays within the set of safety states; S4. Fit the probability distribution of the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; calculate the distribution divergence of the actual observation distribution to quantify the decay offset of the currently used linkage control parameter matrix; when the decay offset exceeds the decay offset threshold, reconstruct the parameters to generate a candidate strategy matrix; generate an updated linkage control parameter matrix based on the candidate strategy matrix; wherein, the updated linkage control parameter matrix is used to replace the currently used linkage control parameter matrix to take over subsequent fire linkage control.
2. The method according to claim 1, characterized in that, S1 includes: S11. Perform a fast Fourier transform on the original fire alarm voice signal to obtain a spectrogram signal. Perform acoustic analysis on the spectrogram signal based on the underlying acoustic model to obtain the preliminary text data sequence. S12. Extract candidate keywords from the preliminary text data sequence; based on a Gaussian mixture model, use Bayesian inference to perform probability inference on the candidate keywords to obtain the posterior probability corresponding to the candidate keywords; based on the posterior probability, filter the candidate keywords according to the maximum a posteriori probability rule to obtain the confirmed keywords; wherein, the formula for calculating the posterior probability is: in, This indicates that, given the spectrogram signal Under the conditions, the first The candidate keywords The posterior probability of occurrence; For the first The candidate keywords; The spectrogram signal; Indicates the assumption that the first The likelihood probability of generating the spectrogram signal when the candidate keywords occur; Indicates the first The prior probabilities of the candidate keywords in a pre-set language database; Represents the spectrogram signal The normalized factor marginal probability of occurrence; S13. Based on the confirmed keywords and the on-site context state information, the text feature sequence is generated by binding the spatiotemporal dependency entity relationships between words through a long short-term memory network.
3. The method according to claim 1, characterized in that, S2 includes: S21. Extract the visual features of the fire scene image stream to obtain a tile space feature vector; calculate the cosine similarity of the tile space feature vector based on the text feature sequence to obtain a similarity vector, and calculate the similarity variance based on the similarity vector; perform saliency screening on the tile space feature vector based on the similarity variance to obtain the visual features to be fused. S22. Based on a dynamic attention calculation model, asymmetric attention feature mapping is performed on the text feature sequence and the visual features to be fused within a unified text-image semantic latent space to obtain the higher-order cross-modal fusion feature; wherein, the calculation formula of the higher-order cross-modal fusion feature is: in, This refers to the higher-order cross-modal fusion feature; The text feature sequence; The visual features to be fused; For query matrix; The key matrix; It is a value matrix; Scaling factor This indicates the corresponding key. Feature dimensions; For activation functions; S23. Based on the dual-track knowledge base in the retrieval-enhanced generation model, perform multi-dimensional nearest neighbor search feature dimensionality reduction on the high-order cross-modal fusion features to generate the structured control alarm data stream; wherein, the dual-track knowledge base includes a static benchmark knowledge base and a real-time dynamic knowledge base, the static benchmark knowledge base is used to provide long-term stable rule constraints including building topology and physical equipment coordinate matrices, and the real-time dynamic knowledge base is used to provide transient environmental context including environmental sensing time-series data and personnel flow heatmaps.
4. The method according to claim 1, characterized in that, The S3 includes: S31. Based on a security reinforcement learning network, perform reinforcement learning inference on the structured control alarm data stream to obtain the original linkage action sequence; S32. Obtain the set physical operating boundary, and set a safety state set for the physical operating boundary based on a continuously differentiable function to obtain the physical safety boundary of the control barrier function; S33. Based on the physical security boundary, perform a quadratic programming optimization solution on the original linkage action sequence to generate the legal safe action sequence, and broadcast the legal safe action sequence; wherein, the calculation formula for the legal safe action sequence is: in, This refers to the legal and safe action sequence; This represents the operation of finding the variable values that minimize the objective function. This represents the candidate control inputs during the optimization process. Indicates the set of allowed actions; This refers to the original sequence of linked actions; This represents the constraints that must be satisfied to solve a quadratic programming problem. This represents the transpose of the gradient of the continuously differentiable function; This is the current state; This represents the intrinsic state evolution term in the system dynamics function that is unaffected by the control input; This represents the control gain term in the system dynamics function that is affected by the control input; This represents the monotonically increasing extended K-class function corresponding to the continuously differentiable function.
5. The method according to claim 1, characterized in that, The S4 includes: S41. Fit the probability distribution of the feedback time series data to obtain the actual observed distribution of system performance. Calculate the distribution offset of the actual observed distribution of system performance based on the set system benchmark distribution to obtain the divergence standardized residual. S42. Based on the adaptive accumulation and monitoring algorithm, the divergence standardized residual is sequentially estimated in time to quantify the decay offset of the underlying linkage logic, and a drift alarm signal is generated when the decay offset exceeds the decay offset threshold; wherein, the adaptive accumulation and monitoring algorithm is used to sequentially monitor small and continuous systematic drift under the condition of random environmental interference, and the underlying linkage logic is used to determine the benchmark response rule for the current control of the underlying fire protection equipment to perform linkage actions. S43. Based on the drift alarm signal, reconstruct the control parameters of the underlying linkage logic to obtain a candidate strategy matrix; S44. Acquire sparse feedback test data under test conditions without taking over the actual physical bus, and perform posterior prediction on the candidate strategy matrix using the sparse feedback test data to obtain the expected return and cognitive uncertainty variance of the new control strategy; wherein, the new control strategy is used to provide candidate response rules to replace the underlying linkage logic after detecting performance degradation of the underlying linkage logic. S45. Based on the Bayesian confidence bound maximization evaluation algorithm, the expected return and the cognitive uncertainty variance are jointly evaluated and verified. When the expected return meets the convergence condition and is greater than the original logical return, the corresponding candidate policy matrix is determined as the target parameter, and the updated linkage control parameter matrix is generated based on the target parameter.
6. An intelligent linkage control system for automatic fire-fighting facilities, used to implement the method according to any one of claims 1 to 5, characterized in that, The system includes: The acoustic probability analysis module is used to acquire the original fire alarm voice signal and the on-site context state information, perform acoustic analysis on the original fire alarm voice signal to obtain a preliminary text data sequence; filter the candidate keywords in the preliminary text data sequence to obtain certain keywords; and perform entity relationship binding processing based on the certain keywords and the on-site context state information to generate a text feature sequence. A cross-modal fusion module is used to acquire a fire scene image stream, remove irrelevant visual background from the fire scene image stream based on the text feature sequence to obtain visual features to be fused, and perform feature mapping on the text feature sequence and the visual features to be fused to obtain high-order cross-modal fusion features; perform feature dimensionality reduction on the high-order cross-modal fusion features to generate a structured control alarm data stream; The safety linkage decision module is used to solve the optimization problem of physical safety boundary that satisfies the control barrier function based on the structured control alarm data stream, generate a legal safety action sequence, and broadcast the legal safety action sequence; The linkage logic evolution module is used to perform probability distribution fitting on the feedback time series data generated after the execution of the legal safety action sequence to obtain the actual observation distribution; to perform distribution divergence calculation on the actual observation distribution to quantify the decay offset of the currently used linkage control parameter matrix; to perform parameter reconstruction when the decay offset exceeds the decay offset threshold to generate a candidate strategy matrix; and to generate an updated linkage control parameter matrix based on the candidate strategy matrix.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.