Building fire prediction and intelligent decision optimization method based on deep learning

Abstract

The invention discloses a building fire prediction and intelligent decision optimization method based on deep learning, and the method comprises the steps: collecting monitoring data in a building, and carrying out the preprocessing of the monitoring data, and constructing a multivariable time sequence sample; constructing an improved TimeMixer model to obtain a risk prediction result and a risk trend sequence; adopting sharpness perception minimization training to obtain target model parameters; inputting parameters with a target model to obtain a prediction result and a risk trend of the building; constructing an emergency decision optimization state vector, and determining candidate linkage actions; solving the candidate linkage action to obtain a linkage control instruction; and issuing an evacuation guide scheme, and completing emergency disposal closed-loop control. According to the method, the improved TimeMixer model is constructed, and sharpness perception minimization is adopted, so that fire risk prediction and risk trend generation of building multi-source time sequence monitoring data are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire safety, and in particular to a building fire prediction and intelligent decision optimization method based on deep learning. BACKGROUND

[0002] The existing building fire warning platform usually takes single-point data collected by fire detectors and building monitoring terminals as the triggering basis, adopts fixed threshold, fixed time criterion or static logic combination to realize alarm and linkage. The current method is relatively simple in engineering implementation, but it is not adaptive to the changes of building operation state. In the actual building environment, air conditioning ventilation, kitchen fume, construction dust, dense flow of people, and equipment aging factors will cause non-fire fluctuations in monitoring data. At the same time, monitoring data may also appear out of synchronization, short packet loss, sensor drift, and noise spike conditions. In the initial stage of fire, there are obvious response lags and asynchronous changes between different monitoring variables. The time sequence and amplitude of temperature rise, smoke diffusion, and gas concentration change are not consistent in different areas. Single threshold or static rules are difficult to express the dynamic correlation across variables and time. Due to the lack of joint modeling and trend prediction ability of multi-source time series information, the existing warning technology is prone to frequent false alarms, increased risk of missed reports, and insufficient warning advance under complex working conditions, which is difficult to meet the needs of early fire risk identification and reliable warning in high-rise buildings and large complex scenarios.

[0003] The existing fire emergency disposal relies on pre-prepared fixed plans or manual judgment and linkage operation by on-duty personnel according to experience. The common process is to start sprinkling, smoke exhaust, air supply, broadcasting according to the established zoning after alarm, or to guide personnel according to static evacuation instructions. There is a lack of coupling calculation link with real-time fire situation prediction, and linkage control and evacuation decision often do not form a closed loop of prediction, decision, execution and feedback. In the actual emergency process, the fire situation has the characteristics of rapid change and spatial diffusion, and the key conditions such as the density of personnel in the building, the state of access control, the passable capacity of the channel, the availability of the elevator, and the running state of the fire-fighting equipment will also change over time. If fixed plans or manual selection of actions are still used, it is difficult to consider the risk trend, zoning connectivity, channel capacity limit, equipment availability state and action execution time sequence in a short time, so as to generate executable linkage control instruction sequence and evacuation guidance scheme.

[0004] Therefore, how to provide a building fire prediction and intelligent decision optimization method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to propose a building fire prediction and intelligent decision optimization method based on deep learning, which comprehensively utilizes multi-source monitoring data fusion, improved TimeMixer model structure innovation and sharpness perception minimization training optimization technology, and completely describes the whole process from building monitoring data acquisition and preprocessing, sliding window construction of multivariate time sequence samples, risk prediction and risk trend generation, emergency decision state vector construction, candidate linkage action screening, linkage control and evacuation guidance multi-objective solution to execution feedback closed loop control; the improved TimeMixer model innovatively introduces variable time migration, frequency dimension embedding, state cache matrix and interlayer alternating flow structure, realizes efficient representation and trend prediction of multivariate time sequence information, and the sharpness perception minimization improves the model training robustness and generalization ability through disturbance space transformation and dislocation gradient update strategy, so as to reduce false alarm and improve early warning lead time, and at the same time make the linkage control and evacuation decision consistent with the prediction result and quickly output executable optimization strategy.

[0006] According to the building fire prediction and intelligent decision optimization method based on deep learning, the method comprises the following steps:

[0007] Acquire monitoring data in the building, preprocess the monitoring data, and construct multivariate time sequence samples according to a sliding window;

[0008] Construct an improved TimeMixer model, perform time sequence dislocation processing between variables on the multivariate time sequence samples through a variable time migration module, introduce a frequency dimension embedding channel for cycle component modulation processing, perform cache processing based on a state cache matrix, and adopt an interlayer alternating flow structure for alternating transmission processing, to obtain a risk prediction result and a risk trend sequence;

[0009] Improve the TimeMixer model by using sharpness perception minimization training, map the loss gradient to a disturbance subspace by using a disturbance space transformation structure, introduce a dislocation gradient update strategy to calculate an update gradient, and obtain target model parameters;

[0010] Input the multivariate time sequence samples into the improved TimeMixer model with the target model parameters, to obtain a fire prediction result and a risk trend sequence of each building partition;

[0011] Construct an emergency decision optimization state vector based on the fire prediction result and the risk trend sequence, and set an action set to determine a candidate linkage action;

[0012] Solve the candidate linkage action to obtain a linkage control instruction sequence and an evacuation guidance scheme, with the optimization objectives of minimizing risk exposure and evacuation time and minimizing equipment linkage cost;

[0013] Execute the linkage control instruction sequence and publish the evacuation guide scheme, complete the building fire emergency disposal closed-loop control.

[0014] Optionally, the monitoring data in the building includes temperature, smoke concentration, flammable gas concentration, fire-fighting equipment operating status, and personnel density.

[0015] Optionally, the constructing the multivariate time series sample according to the sliding window comprises:

[0016] Collecting monitoring data in each partition of the building, recording the collection time and the partition identifier for each monitoring data, and forming a continuous time sequence for the monitoring data in the same partition in the order of collection time;

[0017] Pretreating the monitoring data, the pretreatment comprising aligning different monitoring variables to a unified time index within the same sampling period, completing missing data, removing abnormal data, and normalizing each monitoring variable;

[0018] Constructing the multivariate time series sample according to the sliding window, the sliding window being determined by a preset window length and a preset step length, and the monitoring data of a continuous window length in the continuous time sequence being intercepted for each partition as a multivariate time series sample, and the next multivariate time series sample being generated by moving along the time axis by the preset step length.

[0019] Optionally, the obtaining the risk prediction result and the risk trend sequence comprises:

[0020] Constructing an improved TimeMixer model, which comprises an inter-variable time migration module, a frequency dimension embedding channel, a state cache matrix, an inter-layer alternating flow structure, and a channel mixing sub-layer;

[0021] The inter-variable time migration module performs time series misalignment processing on the multivariate time series sample, sets an integer sampling step time migration amount for each monitoring variable and performs shifting, a positive value indicating backward shifting and a negative value indicating forward shifting, and positions exceeding the sample boundary being padded with a padding value to obtain a time series alignment sample;

[0022] The frequency dimension embedding channel performs periodic component modulation processing on the time series alignment sample, generates a frequency embedding vector for each monitoring variable and combines it with the value of the monitoring variable at each sampling time, the frequency embedding vector containing a plurality of frequency parameter corresponding sine sequences and cosine sequences, to obtain an embedded sample;

[0023] The state cache matrix is used to cache the embedded sample, the state cache matrix storing intermediate features of key sampling times, the key sampling times being determined by event determination rules, the intermediate features of the key sampling times being written into the state cache matrix, and the corresponding intermediate features and the current intermediate features being spliced to obtain a cache enhanced feature sequence when the channel mixing sub-layer is calculated.

[0024] The interlayer alternating flow structure is adopted to alternately transmit and process the cache enhancement feature sequence, the TimeMixer blocks of each layer are arranged according to layer numbers, each TimeMixer block includes a time mixing sublayer and a channel mixing sublayer, an alternating transmission path between adjacent two layers is established, dimension matching is performed to obtain an alternating transmission feature sequence, and a risk prediction result and a risk trend sequence are obtained through an output layer.

[0025] Optionally, the obtaining the target model parameter comprises:

[0026] The loss function of the improved TimeMixer model is calculated based on the multivariate time sequence samples, and a gradient vector of the loss is obtained by solving the gradient of the model parameter, the loss function includes a classification loss of the risk prediction result and a regression loss of the risk trend sequence;

[0027] The loss gradient vector is mapped to a perturbation subspace by using a perturbation space transformation structure, a perturbation projection matrix is set, and a mapped gradient vector is obtained by performing matrix multiplication operation on the loss gradient vector, when generating the parameter perturbation vector, the parameter perturbation vector is obtained according to the product of the perturbation radius and the unit vector of the mapped gradient vector, and the unit vector of the mapped gradient vector is obtained by dividing the mapped gradient vector by the sum of the two-norm of the mapped gradient vector and a minimum constant;

[0028] The parameter perturbation vector is added to the model parameter to obtain a perturbation model parameter, under the perturbation model parameter, a staggered gradient update strategy is adopted to select a staggered sample of a time window adjacent to the multivariate time sequence sample, and a perturbation loss function is calculated, and an update gradient is obtained by solving the gradient of the perturbation model parameter;

[0029] The model parameter is updated based on the update gradient, and the parameter update is performed according to the model parameter after the update being equal to the model parameter before the update minus the product of the learning rate and the update gradient, and the target model parameter is obtained.

[0030] Optionally, the obtaining the fire prediction result and the risk trend sequence of each building partition comprises:

[0031] Real-time monitoring data of each partition of the building are acquired, pre-processing and sliding window construction sample processing are performed, and multivariate time sequence samples corresponding to each partition are obtained;

[0032] The multivariate time sequence samples of each partition are input into the improved TimeMixer model with the target model parameter, time sequence staggering processing between variables, periodic component modulation processing, key time point intermediate feature writing and reading processing, and interlayer alternating transmission processing are performed, and feature sequences of each partition are obtained.

[0033] The feature sequence of each partition is input into an output layer of the improved TimeMixer model to obtain a fire prediction result of each partition and a risk trend sequence in a future time domain, and the fire prediction result and the risk trend sequence are associated with the corresponding partition identifier and output.

[0034] Optionally, the determining the candidate linkage action comprises:

[0035] An emergency decision optimization state vector is constructed according to the fire prediction result and the risk trend sequence of each building partition, and the emergency decision optimization state vector comprises a partition identifier, a fire prediction result, a risk trend sequence, a personnel density, a passable state of a passage, a door access state and an available state of a fire-fighting device.

[0036] An action set is set, and an action type, an action object, an action execution duration and an action execution sequence identifier are defined for each action, and the action set comprises smoke exhaust, air supply, sprinkler partition start, door access control, fire door control and elevator control.

[0037] The candidate linkage action is screened from the action set based on the emergency decision optimization state vector, and the screening condition comprises a partition corresponding to the fire prediction result, a time domain corresponding to the risk trend sequence, an available state of a fire-fighting device and a passable state of a passage, and a candidate linkage action set is obtained.

[0038] Optionally, the obtaining the linkage control instruction sequence and the evacuation guide scheme comprises:

[0039] A decision variable is defined, and the decision variable comprises an execution flag of the candidate linkage action, an action execution start time, an action execution duration, an action object, an evacuation path of each building partition to a safety exit and an evacuation flow distribution of each passage at each discrete time.

[0040] An optimization objective function is constructed, and is set to sequentially minimize a risk exposure, an evacuation time and a device linkage cost, the risk exposure is obtained by multiplying a risk trend sequence value and a number of personnel at a corresponding time of a partition and summing all partitions and all time, the evacuation time is obtained by taking a maximum value of arrival times of all personnel to a safety exit, and the device linkage cost is obtained by summing action values of executed linkage actions.

[0041] A constraint condition set is constructed, and the optimization objective function is solved under the premise of meeting the constraint condition set, the constraint condition set comprises a building partition connectivity, a passage passable capacity, a device available state, an action time sequence and an evacuation feasibility, and a linkage control instruction sequence and an evacuation guide scheme are output.

[0042] Optionally, the completing the building fire emergency disposal closed-loop control comprises:

[0043] The linkage control instruction sequence is sent to a fire linkage controller or a building automatic control platform, and a candidate linkage action is executed according to an action execution starting time and an action execution duration, and an evacuation guide scheme is sent to a building broadcasting platform or a mobile terminal;

[0044] During the execution, execution feedback data is collected in real time to form feedback time sequence data, the execution feedback data includes linkage action execution state, fire equipment operation state, access control state, passable state of a passageway and personnel density change data, the emergency decision optimization state vector is updated based on the feedback time sequence data, and closed-loop control of building fire emergency disposal is completed.

[0045] The beneficial effects of the present application are:

[0046] The present application collects multi-source monitoring data in the building and constructs multivariate time sequence samples, and realizes joint prediction of fire risk and risk trend by combining the improved TimeMixer model, thereby improving the identification ability of early fire hazards under complex working conditions. Compared with the existing single sensor threshold or static rule warning mode, the present application can depict the asynchronous response and periodic disturbance characteristics between monitoring variables by using time migration between variables, frequency dimension embedding, state buffer matrix and interlayer alternating flow structure; at the same time, the sharpness perception minimization training method containing disturbance space transformation structure and staggered gradient update strategy is adopted, the adaptability of the model to noise, missing and drift data is improved, thereby reducing false alarm and improving warning stability.

[0047] The present application maps the fire prediction result and the risk trend sequence into an emergency decision optimization state vector, and selects a candidate linkage action based on an action set, and further outputs a linkage control instruction sequence and an evacuation guide scheme through multi-objective solving, thereby realizing integration of prediction and disposal. Compared with the existing disposal mode which relies on fixed preplan or manual experience and is disconnected with the prediction result of linkage control and evacuation decision, the present application integrates risk exposure, evacuation time and equipment linkage cost into a unified optimization framework, and forms a closed-loop control by updating the state vector through execution feedback, thereby being capable of quickly generating an executable strategy under real operation conditions of the building and continuously rolling correction, thereby improving emergency disposal efficiency and overall safety level. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation of the present application. In the drawings:

[0049] Figure 1 A flowchart of a building fire prediction and intelligent decision optimization method based on deep learning is provided in the present application;

[0050] Figure 2A structural block diagram of an improved TimeMixer model of a building fire prediction and intelligent decision optimization method based on deep learning is provided in the present application.

[0051] Figure 3 A sharpness perception minimization function diagram of a building fire prediction and intelligent decision optimization method based on deep learning is provided in the present application. DETAILED DESCRIPTION

[0052] The present application will now be further described in greater detail in conjunction with the accompanying drawings, in which these drawings are simplified schematic diagrams only showing the basic structure of the present application in a schematic manner and thus only show the components relevant to the present application.

[0053] Reference Figure 1 , Figure 2 and Figure 3 A building fire prediction and intelligent decision optimization method based on deep learning, comprising:

[0054] Collecting monitoring data in the building, pre-processing the monitoring data, and constructing multivariate time series samples according to a sliding window;

[0055] Constructing an improved TimeMixer model, performing variable time migration processing on the multivariate time series samples through a variable time migration module, introducing a frequency dimension embedding channel for cycle component modulation processing, performing cache processing based on a state cache matrix, and performing alternating transmission processing using an inter-layer alternating flow structure to obtain risk prediction results and risk trend sequences;

[0056] Training the improved TimeMixer model using sharpness perception minimization, mapping the loss gradient to a perturbation subspace using a perturbation space transformation structure, introducing a misalignment gradient update strategy to calculate the update gradient, and obtaining target model parameters;

[0057] Inputting the multivariate time series samples into the improved TimeMixer model with the target model parameters to obtain fire prediction results and risk trend sequences for each building partition;

[0058] Constructing an emergency decision optimization state vector based on the fire prediction results and risk trend sequences, and setting an action set to determine candidate linkage actions;

[0059] Solving the candidate linkage actions to obtain linkage control instruction sequences and evacuation guidance schemes with the optimization objectives of minimizing risk exposure and evacuation time and minimizing equipment linkage cost;

[0060] Executing the linkage control instruction sequences and issuing the evacuation guidance schemes to complete the building fire emergency disposal closed-loop control.

[0061] In the embodiment, the monitoring data in the building includes temperature, smoke concentration, flammable gas concentration, fire-fighting equipment operating state and personnel density.

[0062] In the embodiment, the multivariate time series sample is constructed according to a sliding window, including:

[0063] The monitoring data is collected in each partition of the building, the collection time and the partition identifier are recorded for each monitoring data record, and the monitoring data of the same partition is sequentially formed into a continuous time sequence according to the collection time;

[0064] The monitoring data is preprocessed, including aligning different monitoring variables to a unified time index within the same sampling period, completing missing data, removing abnormal data, and normalizing each monitoring variable;

[0065] The multivariate time series sample is constructed according to a sliding window, the sliding window is determined by a preset window length and a preset step length, the monitoring data corresponding to the continuous window length in the continuous time sequence is intercepted for each partition as a multivariate time series sample, and the next multivariate time series sample is generated by moving along the time axis according to the preset step length.

[0066] In the embodiment, the risk prediction result and the risk trend sequence are obtained, including:

[0067] The improved TimeMixer model is constructed, which is composed of an inter-variable time migration module, a frequency dimension embedding channel, a state cache matrix, an inter-layer alternating flow structure and a channel mixing sub-layer, wherein:

[0068] The improved TimeMixer model takes the multivariate time series sample as input, performs time series dislocation processing on each monitoring variable sequence by shifting by an integer sampling step through the inter-variable time migration module at the input end, and combines the frequency embedding vector composed of sine sequence and cosine sequence with the variable sequence through the frequency dimension embedding channel to complete the periodic component modulation processing;

[0069] A plurality of TimeMixer blocks containing channel mixing sub-layers are stacked in the backbone structure to realize multivariate feature interaction, a state cache matrix is set to write and read intermediate features at key sampling times and splice them with current intermediate features to form cache enhanced features, an inter-layer alternating flow structure is introduced between adjacent TimeMixer blocks to establish an alternating transmission path and perform dimension matching, and finally the risk prediction result and the risk trend sequence are obtained through the output layer;

[0070] The time migration module between variables performs time sequence misalignment processing on the multivariate time sequence sample, sets an integer sampling step time migration amount for each monitoring variable and performs shifting, a positive value indicates backward shifting, a negative value indicates forward shifting, and positions exceeding the sample boundary are padded to obtain time sequence alignment samples. The time migration amount is set to an integer sampling step in units of sampling periods, and a time migration amount is specified for each monitoring variable in the multivariate time sequence sample, and a corresponding table of variables and time migration amounts is established. The time sequence of each monitoring variable is indexed and translated according to the corresponding time migration amount. When the time migration amount is positive, the sequence is shifted backward by the corresponding number of steps along the time axis and the missing positions at the front end of the sequence are padded with padding values. When the time migration amount is negative, the sequence is shifted forward by the corresponding number of steps along the time axis and the missing positions at the end of the sequence are padded with padding values. Thus, the time sequence alignment samples of the variables after shifting are obtained.

[0071] The time migration amount is set to an integer sampling step in units of sampling periods, and a time migration amount is specified for each monitoring variable in the multivariate time sequence sample, and a corresponding table of variables and time migration amounts is established. The time sequence of each monitoring variable is indexed and translated according to the corresponding time migration amount. When the time migration amount is positive, the sequence is shifted backward by the corresponding number of steps along the time axis and the missing positions at the front end of the sequence are padded with padding values. When the time migration amount is negative, the sequence is shifted forward by the corresponding number of steps along the time axis and the missing positions at the end of the sequence are padded with padding values. Thus, the time sequence alignment samples of the variables after shifting are obtained.

[0072] The frequency dimension embedding channel performs periodic component modulation processing on the time sequence alignment sample, generates a frequency embedding vector for each monitoring variable, and combines it with the numerical value of the monitoring variable at each sampling time. The frequency embedding vector contains multiple frequency parameter corresponding sine and cosine sequences, and the embedded sample is obtained. The time sequence alignment sample is processed by periodic component modulation, specifically:

[0073] A set of frequency parameters is preset for each monitoring variable, and the sampling time sequence is used as the independent variable to calculate the sine and cosine sequences of the corresponding frequency parameters. The multiple sets of sine and cosine sequences corresponding to the same monitoring variable are concatenated according to the dimensions to form a frequency embedding vector. At each sampling time, the numerical value of the current monitoring variable is combined with the frequency embedding vector corresponding to the sampling time. The combination method is to copy the monitoring variable value to the same dimension as the frequency embedding vector and then concatenate it with the frequency embedding vector to form a variable representation with frequency information. The embedded sample is obtained by repeating the above process for all monitoring variables and all sampling times.

[0074] The frequency parameters are preset as a set of positive integers in units of sampling steps, and each period is converted to an angular frequency parameter by setting the angular frequency equal to two times the circumference divided by the period. The angular frequency parameter and the sampling time index are used to calculate the corresponding sine and cosine sequences and concatenate them into a frequency embedding vector. The embedded sample is obtained by combining it with the monitoring variable value.

[0075] The state buffer matrix is used to cache the embedded sample, the state buffer matrix is set to store the intermediate features of the key sampling time, the key sampling time is determined by the event determination rule, the intermediate features of the key sampling time are written into the state buffer matrix, and the corresponding intermediate features are read during the calculation of the channel mixing sublayer to concatenate the current intermediate features and obtain the cache enhanced feature sequence.

[0076] The state cache matrix, in particular: the state cache matrix is set according to the number of cache slots multiplied by the feature dimension, the number of cache slots is set to eight, and is used to store the intermediate feature vectors of the same partition at different key sampling moments; each cache slot contains a timestamp field and a feature vector field; the timestamp field records the key sampling moment index; the feature vector field stores the intermediate features corresponding to the sampling moment; when writing, the rule of empty slot first and then covering is followed; when reading, the cache features are obtained by matching the key sampling moment index corresponding to the current sampling moment according to the timestamp field;

[0077] The event determination rule, in particular: the event determination rule includes amplitude mutation determination and cumulative change determination; the amplitude mutation determination is to calculate the difference value of adjacent sampling moments for each monitoring variable and take the absolute value; if the absolute value exceeds the first preset threshold, the sampling moment is determined as the key sampling moment; the cumulative change determination is to calculate the absolute value of the difference between the first and last sampling moment values of each monitoring variable within a time period; if the result exceeds the second preset threshold, the end sampling moment within the current time period is determined as the key sampling moment; the first preset threshold is set to 0.2, and the second preset threshold is set to 0.5;

[0078] The channel mixing sub-layer calculation, in particular: the channel mixing sub-layer takes the multivariate feature vector of the same sampling moment as input; the input feature vector and the cache feature vector read from the state cache matrix are first spliced according to the feature dimension to form an expanded feature vector; then the expanded feature vector is sequentially subjected to first linear transformation, nonlinear activation operation and second linear transformation to obtain a mixed output feature vector; and the mixed output feature vector is used as the intermediate feature of the sampling moment to generate the cache enhanced feature sequence;

[0079] An interlayer alternating flow structure is adopted to perform alternating transmission processing on the cache enhanced feature sequence; the TimeMixer blocks of each layer are arranged according to layer number; each TimeMixer block contains a time mixing sub-layer and a channel mixing sub-layer; an alternating transmission path between adjacent two layers is established; dimension matching is performed to obtain an alternating transmission feature sequence; a risk prediction result and a risk trend sequence are obtained through an output layer; the establishment of the alternating transmission path between adjacent two layers and the dimension matching to obtain the alternating transmission feature sequence are as follows:

[0080] The alternate transmission path between two adjacent layers is realized by establishing a connection path between the time mixing sublayer output of the previous layer and the channel mixing sublayer input of the next layer, and establishing a connection path between the channel mixing sublayer output of the previous layer and the time mixing sublayer input of the next layer, and respectively splicing the features transmitted by the two connection paths with the original input of the corresponding sublayer of the next layer in the feature dimension, and dimension matching is realized by linearly transforming the spliced feature vectors to convert the feature dimension into the input dimension of the corresponding sublayer of the next layer, so as to form the feature representation after alternate transmission at each sampling time and form the alternate transmission feature sequence in time sequence.

[0081] In the embodiment, the target model parameter includes:

[0082] The loss function of the improved TimeMixer model is calculated based on the multivariate time sequence sample, and a loss gradient vector is obtained by calculating the gradient of the model parameter, the loss function includes a classification loss of a risk prediction result and a regression loss of a risk trend sequence, and wherein:

[0083] The loss gradient vector is obtained by calculating the gradient of the model parameter, and specifically: in the training stage, the multivariate time sequence sample is input into the improved TimeMixer model to obtain the risk prediction result and the risk trend sequence, the classification loss and the regression loss are calculated according to the corresponding label and added to obtain the loss function value, then the loss function value is calculated with respect to all trainable parameters of the model, the chain rule is applied layer by layer from the output layer to the front according to the calculation graph, the partial derivative of each layer parameter is obtained and is sequentially flattened and spliced to form the loss gradient vector, each element of the loss gradient vector corresponds to the gradient value of a trainable parameter;

[0084] The classification loss and the regression loss of the risk trend sequence are specifically: the classification loss adopts cross-entropy loss, the calculation method is to one-hot encode the risk category label of each partition, multiply the natural logarithm of each risk category prediction probability output by the improved TimeMixer model by the corresponding one-hot label and sum to take the negative value, and then average all samples in a batch to obtain the classification loss, the regression loss of the risk trend sequence adopts mean absolute error loss, the calculation method is to take the absolute value of the difference between the predicted risk trend value and the real risk trend value at each discrete time in the prediction time domain, sum all discrete times, and then average all samples in a batch to obtain the regression loss;

[0085] The loss gradient vector is mapped to a perturbation subspace by using a perturbation space transformation structure, a perturbation projection matrix is set, and a matrix multiplication operation is performed on the loss gradient vector to obtain a mapped gradient vector, and when generating a parameter perturbation vector, the parameter perturbation vector is obtained according to the product of the perturbation radius and the unit vector of the mapped gradient vector, the unit vector of the mapped gradient vector is obtained by dividing the mapped gradient vector by the sum of the two-norm of the mapped gradient vector and a minimum constant, and the minimum constant is set to 10 raised to the power of -8.

[0086] The perturbation projection matrix is set, and a matrix multiplication operation is performed on the loss gradient vector to obtain a mapped gradient vector, and the specific operation is as follows:

[0087] The perturbation projection matrix is set to be a matrix for mapping the loss gradient vector from the original parameter dimension to a preset perturbation dimension, the perturbation dimension is first determined to be smaller than the original parameter dimension, then a sparse matrix with element values of zero or nonzero is generated according to a random seed and row vector normalization is performed, so that the two-norm of each row is equal to one, the loss gradient vector is arranged in the form of a column vector, and then a matrix multiplication operation of the perturbation projection matrix and the loss gradient vector is performed, and the output column vector of the matrix multiplication is the mapped gradient vector, the dimension is equal to the preset perturbation dimension, and the preset perturbation dimension is set to 1024*1024.

[0088] The minimum constant is specifically used to add the two-norm when calculating the two-norm of the mapped gradient vector, so as to avoid division by zero caused by the two-norm being zero or close to zero, and ensure that the calculation of the mapped gradient vector unit vector and the parameter perturbation vector is executable and numerically stable, and the minimum constant is set to 10 raised to the power of -8.

[0089] The parameter perturbation vector is added to the model parameter to obtain a perturbed model parameter, and under the perturbed model parameter, a staggered gradient update strategy is used to select staggered samples adjacent to a time window of a multivariate time series sample, and a perturbed loss function is calculated, and an update gradient is obtained by taking the gradient of the perturbed model parameter, wherein:

[0090] The staggered gradient update strategy is specifically as follows: after the perturbed model parameter is generated, the original time window sample used to calculate the loss gradient vector is not used, but a time window adjacent to the original time window is selected as a staggered sample from a continuous time sequence in the same partition, the starting time of the staggered sample is different from the starting time of the original time window by a preset staggered step, and the staggered step is set to one sampling step. When the original time window is a continuous segment from the starting time to the ending time, the staggered sample is a continuous segment from the starting time shifted by one sampling step to the ending time shifted by one sampling step.

[0091] The computing perturbation loss function is specifically: the perturbation loss function is calculated forwardly by taking the misaligned samples as inputs under the perturbed model parameters to obtain a risk prediction result and a risk trend sequence, and the risk prediction result and the risk trend sequence are respectively calculated with a risk category label corresponding to the misaligned samples and a real risk trend sequence to obtain a classification loss and a regression loss, and the classification loss and the regression loss are added to obtain a numerical value of the perturbation loss function.

[0092] The model parameters are updated based on the update gradient, and the parameter updating is performed according to that the model parameters after the updating are equal to the model parameters before the updating minus a product of a learning rate and the update gradient to obtain target model parameters, and the obtaining target model parameters is specifically:

[0093] The target model parameters are obtained through iterative training, and each iteration first calculates a loss gradient vector from the multivariate time sequence samples and generates a parameter perturbation vector through a perturbation space transformation, then calculates an update gradient from the misaligned samples under the perturbed model parameters, and then performs parameter updating on all trainable parameters of the model element by element, that is, the current value of each parameter is reduced by a product of a learning rate and an update gradient value corresponding to the parameter to obtain a new value of the parameter, and one parameter updating is completed, and the iterative updating is repeatedly performed until a training number or a training set loss does not decrease continuously for several rounds, and the parameter set after the last iterative updating is the target model parameters.

[0094] In the embodiment, the obtaining of the fire prediction result and the risk trend sequence of each building partition includes:

[0095] Real-time monitoring data of each partition of the building are acquired and preprocessed and processed by a sliding window to obtain multivariate time sequence samples corresponding to each partition;

[0096] The multivariate time sequence samples of each partition are input into an improved TimeMixer model with target model parameters to perform time sequence misalignment processing between variables, cycle component modulation processing, intermediate feature writing and reading processing at key moments, and interlayer alternating transmission processing to obtain feature sequences of each partition.

[0097] The feature sequences of each partition are input into an output layer of the improved TimeMixer model to obtain a fire prediction result and a risk trend sequence in a future time domain of each partition, and the fire prediction result and the risk trend sequence are associated with the corresponding partition identifier and output.

[0098] In the embodiment, the determining of the candidate linkage action includes:

[0099] An emergency decision optimization state vector is constructed according to the fire prediction result and the risk trend sequence of each building partition, and the emergency decision optimization state vector includes a partition identifier, a fire prediction result, a risk trend sequence, a personnel density, a passable state of a passageway, an access control state, and an available state of a fire-fighting device, and the constructing of the emergency decision optimization state vector is specifically:

[0100] The emergency decision optimization state vector is constructed by partition one by one. For each partition, the partition identifier is coded into a fixed dimension identifier vector, the fire prediction result of the current partition is coded into a risk level value or a risk probability vector, the risk trend sequence of the partition is unfolded into a sequence vector according to the discrete time of the prediction time domain in order, and the personnel density value of the partition at the corresponding time, the passable state flag of the channel connected with the partition, the access control state flag of the partition and the available state flag of the fire-fighting equipment of the partition are synchronously obtained. The identifier vector, risk coding, trend sequence vector and each state flag are spliced in order to obtain the emergency decision optimization state vector of the current partition, and the state vectors of all partitions are sequentially combined into a state vector set for solving;

[0101] An action set is set, and an action type, an action object, an action execution duration and an action execution sequence identifier are defined for each action. The action set includes smoke exhaust, air supply, sprinkler partition start, access control, fire door control and elevator control. The setting of the action set is specifically:

[0102] The action set is set in the form of an action item table. An action name, an action type, an action object, an action execution duration and an action execution sequence identifier are written for each action item. The action name includes smoke exhaust, air supply, sprinkler partition start, access control, fire door control and elevator control in turn. The action type is set as fan control, fan control, sprinkler control, access control, partition control and elevator control. The action object is bound to the corresponding device or device group with the building partition as the index. The action objects of smoke exhaust and air supply are the smoke exhaust fan, the air supply fan and the air valve group of the corresponding partition. The action object of sprinkler partition start is the sprinkler valve group of the corresponding partition.

[0103] The action object of access control is the access controller of the corresponding partition. The action object of fire door control is the fire door controller of the corresponding partition. The action object of elevator control is the elevator group control interface associated with the corresponding partition. The action execution duration is uniformly set as thirty seconds. The action execution sequence identifier is set as access control first, fire door control and elevator control in the middle, smoke exhaust and air supply later, and sprinkler partition start last. The action item table is used as the input for candidate linkage action screening and instruction sequence solving.

[0104] The candidate linkage action is screened from the action set based on the emergency decision optimization state vector. The screening conditions include the partition corresponding to the fire prediction result, the time domain corresponding to the risk trend sequence, the available state of the fire-fighting equipment and the passable state of the channel. The candidate linkage action set is obtained. The screening conditions are specifically:

[0105] The screening conditions include four categories, the partition matching condition is to only keep the action whose action object contains the partition with the medium risk or high risk indicated by the fire prediction result and the adjacent connected partition, and the time domain matching condition is to only keep the action whose action execution starting time falls within the decision time domain defined by the risk trend sequence and the action execution duration is not more than the length of the decision time domain;

[0106] The device available condition is to only keep the action corresponding to the device with the available device available state flag in the emergency decision optimization state vector, and to eliminate the action corresponding to the device with the unavailable device available state flag, the passage passable condition is to only keep the action which does not set the evacuation main passage door as closed and does not take the passage marked as impassable as the entrance of the evacuation path, and to eliminate the action which causes the key passage to be impassable, thereby forming the candidate linkage action set.

[0107] In the embodiment, the linkage control instruction sequence and the evacuation guide scheme are obtained, including:

[0108] The decision variables are defined, including the execution flag of the candidate linkage action, the action execution starting time, the action execution duration, the action object, and the evacuation path of each building partition to the safety exit and the evacuation flow distribution of each passage at each discrete time, and the definition of the decision variable is specifically:

[0109] A binary execution flag is set for each candidate linkage action, taking the value of zero or one, an action execution starting time is set, taking the discrete time index within the decision time domain, an action execution duration is set, taking the discrete duration value in the preset allowed set, and the device or device group identifier bound in the action set is taken as the action object. For the evacuation part, an evacuation path variable is defined for each building partition, which is composed of partition nodes and passage edges and terminates at the safety exit, and the path variable takes a feasible path in the connected graph. At the same time, an evacuation flow distribution variable is defined for each passage at each discrete time, the flow distribution variable takes a non-negative real number and is configured according to the connection relationship from the partition to the passage, thereby forming a decision variable set for solving;

[0110] An optimization objective function is constructed, which is set to sequentially minimize the risk exposure, the evacuation time and the device linkage cost, the risk exposure is obtained by multiplying the risk trend sequence value and the number of personnel at the corresponding time of each partition and summing all partitions and all time, the evacuation time is obtained by the maximum value of the arrival time of all personnel to the safety exit, and the device linkage cost is obtained by summing the action value of the executed linkage action;

[0111] The constraint condition set is constructed and the optimization objective function is solved under the premise of meeting the constraint condition set, the constraint condition set includes building partition connectivity, channel traffic capacity, equipment available state, action timing and evacuation feasibility, and outputs linkage control instruction sequence and evacuation guidance scheme.

[0112] In the embodiment, the closed-loop control of the building fire emergency disposal includes:

[0113] The linkage control instruction sequence is sent to the fire linkage controller or the building automation platform, and the candidate linkage action is executed according to the action execution starting time and the action execution duration, and the evacuation guidance scheme is published to the building broadcast platform or the mobile terminal.

[0114] During the execution, the execution feedback data is collected in real time and the feedback time series data is formed, the execution feedback data includes linkage action execution state, fire equipment running state, access control state, channel passable state and personnel density change data, the emergency decision optimization state vector is updated based on the feedback time series data, and the closed-loop control of the building fire emergency disposal is completed.

[0115] Embodiment 1:

[0116] In order to verify the feasibility of the application in the implementation, the application is applied to an underground equipment area and an adjacent public walkway area of a large comprehensive building, temperature, smoke, flammable gas, fire equipment running state, access control state, ventilation related quantity and personnel density multi-source monitoring points are arranged in the building, and the partition monitoring sequence is continuously collected according to the fixed sampling period. In the current area, there are ventilation switching, equipment starting and stopping, personnel flow and dust disturbance non-fire factors, and the monitoring data is prone to fluctuation, peak and short-time loss; at the same time, the fire incubation stage often shows slow change of multiple variables and response lag between different variables, which leads to difficulty in identifying early risk in time by traditional single threshold or static rule, and easy to produce false alarm or miss report, and fixed preplan is difficult to automatically adjust linkage action and evacuation guidance according to risk change, causing prediction and disposal to be disconnected.

[0117] In the current scenario, time alignment, missing completion, anomaly removal and normalization processing are performed on the collected monitoring data, and a sliding window is used to construct a multivariate time series sample input to improve the TimeMixer model. The time migration module between variables performs a shift on the sequence according to the integer sampling step migration set by each monitoring variable, realizes the time sequence misalignment processing between variables; the frequency dimension embedding channel generates a frequency embedding vector composed of a plurality of frequency parameters corresponding to the sine sequence and the cosine sequence for each variable and combines it with the variable value, realizes the periodic component modulation processing; the state buffer matrix writes the intermediate features at the key sampling time triggered by the event determination rule, and reads and splices the current intermediate features when calculating the channel mixing sublayer, forming a cache enhanced feature sequence; the interlayer alternating flow structure establishes an alternating transmission path between adjacent TimeMixer blocks and completes the dimension matching, outputs the risk prediction result and the risk trend sequence. Subsequently, based on the fire prediction result and the risk trend sequence, an emergency decision optimization state vector is constructed, candidate linkage actions are selected from the action set, and a linkage control instruction sequence and evacuation guidance scheme are obtained by solving, and are issued to the fire linkage controller or the building automation platform and the evacuation guidance information is issued simultaneously.

[0118] In order to verify the beneficial effects, a number of exercise data and real alarm data were selected from the historical records and exercise records of the same building platform for playback and retesting, and the pre-alarm output, false alarm and missed alarm, decision generation delay, evacuation results, risk exposure and linkage conflict indicators were recorded in multiple linkage exercises. The results show that after using the improved TimeMixer model and training with sharpness perception minimization, the model can maintain stable output in the presence of missing, spike and periodic disturbance in the monitoring data, improve the timeliness of fire risk identification and reduce false positives and false negatives; At the same time, the prediction results and trend sequence are directly used for decision optimization, so that the linkage control and evacuation strategy can be updated with the situation and remain executable.

[0119] Table 1 Comparison of multi-source fire prediction and linkage decision effect

[0120] Contrast index Invention method Threshold rule Fixed plan LSTM+rule Transformer+rule Original TimeMixer Early warning time (s) 210 60 0 165 180 195 False alarm rate (%) 1.8 9.6 6.2 4.9 4.3 3.2 Miss rate (%) 2.1 8.4 7.1 3.6 3.1 2.7 Trend sequence MAE 0.031 0.092 0.086 0.049 0.044 0.038 Decision generation delay (s) 0.78 0.10 0.05 1.05 1.22 0.92 Evacuation time (s) 410 520 545 455 448 432 Risk exposure integral 1180 1760 1895 1390 1335 1245 Number of conflict instructions 0 3 4 1 1 1

[0121] As can be seen from Table 1, in terms of early warning capability, the early warning lead of the method reaches 210 seconds, which is obviously higher than 60 seconds of the threshold rule, and is also better than 165 seconds of LSTM+rule, 180 seconds of Transformer+rule and 195 seconds of the original TimeMixer; meanwhile, in terms of accuracy, the false alarm rate of the method is 1.8%, and the false negative rate is 2.1%, which is greatly reduced compared with 9.6% and 8.4% of the threshold rule, 6.2% and 7.1% of the fixed plan, and is also better than 4.9%, 3.6% of LSTM+rule, 4.3%, 3.1% of Transformer+rule and 3.2%, 2.7% of the original TimeMixer, which embodies the more stable risk identification capability under complex working conditions and multi-source fluctuations.

[0122] In terms of trend prediction quality, the trend sequence MAE of the method is 0.031, which is lower than 0.092 of the threshold rule, 0.086 of the fixed plan, and also lower than 0.049 of LSTM+rule, 0.044 of Transformer+rule and 0.038 of the original TimeMixer, which indicates that the fitting error of the risk trend is smaller; in terms of response speed, the decision generation delay of the method is 0.78 seconds, which is higher than the direct triggering mode of 0.10 seconds of the threshold rule and 0.05 seconds of the fixed plan, but still maintains sub-second output, and is obviously faster than 1.05 seconds of LSTM+rule and 1.22 seconds of Transformer+rule, and is better than 0.92 seconds of the original TimeMixer, which can meet the real-time requirement of linkage control.

[0123] In terms of disposal effect, the evacuation completion time of the method is 410 seconds, which is shorter than 520 seconds of the threshold rule, 545 seconds of the fixed plan, and also shorter than 455 seconds of LSTM+rule, 448 seconds of Transformer+rule and 432 seconds of the original TimeMixer; the corresponding risk exposure integral is 1180, which is lower than 1760 of the threshold rule, 1895 of the fixed plan, and also lower than 1390 of LSTM+rule, 1335 of Transformer+rule and 1245 of the original TimeMixer, which indicates that the method is more effective in reducing the risk of personnel exposure; the number of conflict instructions is 0, which is compared with 3 of the threshold rule, 4 of the fixed plan and learning model scheme 1, indicating that the consistency of the linkage control instruction sequence is better, and the demand for manual intervention caused by instruction mutual exclusion or time sequence conflict is reduced.

[0124] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical range disclosed by the present application according to the technical solution and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A deep learning-based building fire prediction and intelligent decision optimization method, characterized in that, The method comprises the following steps: Collecting monitoring data in the building, preprocessing the monitoring data, and constructing multivariate time series samples according to a sliding window; An improved TimeMixer model is constructed, the multivariate time series samples are processed by a time migration module between variables, a frequency dimension embedding channel is introduced for periodic component modulation processing, a state buffer matrix is used for buffer processing, and an alternating flow structure between layers is used for alternating transmission processing, to obtain a risk prediction result and a risk trend sequence; The improved TimeMixer model is trained by sharpness perception minimization, the loss gradient is mapped to a perturbation subspace by using a perturbation space transformation structure, and a misalignment gradient update strategy is introduced to calculate an update gradient, to obtain target model parameters; The multivariate time series samples are input into the improved TimeMixer model with the target model parameters, to obtain a fire prediction result and a risk trend sequence for each building partition; An emergency decision optimization state vector is constructed based on the fire prediction result and the risk trend sequence, and a set of actions is set to determine a candidate linkage action; The candidate linkage action is solved by minimizing the risk exposure, the evacuation time and the equipment linkage cost, to obtain a linkage control instruction sequence and an evacuation guidance scheme; The linkage control instruction sequence is executed and the evacuation guidance scheme is issued, to complete the closed-loop control of building fire emergency disposal.

2. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The monitoring data in the building includes temperature, smoke concentration, flammable gas concentration, fire-fighting equipment operating status and personnel density.

3. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The multivariate time series samples are constructed according to a sliding window, which comprises the following steps: Monitoring data is collected in each partition of the building, the collection time and the partition identifier are recorded for each monitoring data, and the monitoring data of the same partition is sequentially formed into a continuous time sequence according to the collection time; The monitoring data is preprocessed, which includes aligning different monitoring variables to a unified time index within the same sampling period, completing missing data, removing abnormal data, and normalizing each monitoring variable; The multivariate time series samples are constructed according to a sliding window, the sliding window is determined by a preset window length and a preset step length, the monitoring data corresponding to the continuous window length in the continuous time sequence is intercepted for each partition as a multivariate time series sample, and the next multivariate time series sample is generated by moving along the time axis by the preset step length.

4. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The risk prediction result and the risk trend sequence are obtained, which comprises the following steps: An improved TimeMixer model is constructed, which comprises a time migration module between variables, a frequency dimension embedding channel, a state buffer matrix, an alternating flow structure between layers and a channel mixing sublayer; The time migration module between variables processes the multivariate time series samples, sets an integer sampling step time migration amount for each monitoring variable and performs shifting, a positive value represents backward shifting, a negative value represents forward shifting, and a position exceeding the sample boundary is padded with a padding value to obtain a time-aligned sample; The frequency dimension embedding channel modulates the periodic component of the time-aligned sample, generates a frequency embedding vector for each monitoring variable, and combines it with the numerical value of the monitoring variable at each sampling time, the frequency embedding vector contains a plurality of frequency parameter corresponding sine sequences and cosine sequences, and an embedded sample is obtained; The state cache matrix is used for cache processing of the embedded sample, the state cache matrix is configured to store intermediate features at key sampling time points, the key sampling time points are determined by event determination rules, the intermediate features at the key sampling time points are written into the state cache matrix, and the corresponding intermediate features are read and spliced with current intermediate features to obtain a cache enhanced feature sequence during channel mixing sublayer calculation; The cache enhanced feature sequence is processed by using an interlayer alternating flow structure, the TimeMixer blocks of each layer are arranged according to layer numbers, each TimeMixer block includes a time mixing sublayer and a channel mixing sublayer, an alternating transmission path is established between adjacent two layers, dimension matching is performed to obtain an alternating transmission feature sequence, and a risk prediction result and a risk trend sequence are obtained through an output layer.

5. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The target model parameters are obtained, including: A loss function of the improved TimeMixer model is calculated based on the multivariate time sequence samples, and a loss gradient vector is obtained by calculating the gradient of the model parameters, the loss function includes a classification loss of the risk prediction result and a regression loss of the risk trend sequence; The loss gradient vector is mapped to a perturbation subspace by using a perturbation space transformation structure, a perturbation projection matrix is set, and a mapped gradient vector is obtained by performing matrix multiplication operation on the loss gradient vector, a parameter perturbation vector is obtained by multiplying a perturbation radius and a unit vector of the mapped gradient vector, and the unit vector of the mapped gradient vector is obtained by dividing the mapped gradient vector by the sum of the two-norm of the mapped gradient vector and a minimum constant; The parameter perturbation vector is added to the model parameters to obtain perturbed model parameters, and a misalignment sample adjacent to the multivariate time sequence sample is selected by using a misalignment gradient update strategy under the perturbed model parameters, and a perturbed loss function is calculated, and an update gradient is obtained by calculating the gradient of the perturbed model parameters. The model parameters are updated based on the update gradient, and the parameter update is performed according to the updated model parameters equal to the model parameters before the update minus the product of the learning rate and the update gradient, and the target model parameters are obtained.

6. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The fire prediction result and the risk trend sequence of each building partition are obtained, including: Real-time monitoring data of each partition of the building are obtained, pre-processing and sliding window sample processing are performed, and multivariate time sequence samples corresponding to each partition are obtained; The multivariate time sequence samples of each partition are input into the improved TimeMixer model with target model parameters, time sequence misalignment processing between variables, periodic component modulation processing, intermediate feature writing and reading processing at key time points, and interlayer alternating transmission processing are performed, and feature sequences of each partition are obtained; The feature sequences of each partition are input into the output layer of the improved TimeMixer model, the fire prediction result and the risk trend sequence in the future time domain of each partition are obtained, and the fire prediction result and the risk trend sequence are associated with the corresponding partition identifier and output.

7. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The candidate linkage action is determined, including: An emergency decision optimization state vector is constructed according to the fire prediction result and the risk trend sequence of each building partition, the emergency decision optimization state vector includes a partition identifier, a fire prediction result, a risk trend sequence, a personnel density, a passable state of a channel, a door access state and an available state of a fire fighting equipment; The action set is set and the action type, action object, action execution duration and action execution sequence are defined for each action, and the action set includes smoke exhaust, air supply, sprinkler partition start, access control, fire door control and elevator control; The candidate linkage action set is filtered from the action set based on the emergency decision optimization state vector, and the filtering conditions include the partition corresponding to the fire prediction result, the time domain corresponding to the risk trend sequence, the available state of the fire fighting equipment and the passable state of the passage, and the candidate linkage action set is obtained.

8. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The linkage control instruction sequence and the evacuation guidance scheme are obtained, including: The decision variable is defined, including the execution flag of the candidate linkage action, the action execution start time, the action execution duration, the action object, and the evacuation path of each building partition to the safety exit and the evacuation flow distribution of each passage at each discrete time; The optimization objective function is constructed, which is set to minimize the risk exposure, the evacuation time and the equipment linkage cost in sequence, the risk exposure is obtained by multiplying the risk trend sequence value and the number of personnel at the corresponding time of each partition and summing all partitions and all time, the evacuation time is obtained by the maximum value of the arrival time of all personnel to the safety exit, and the equipment linkage cost is obtained by summing the action value of the executed linkage action; The constraint condition set is constructed and the optimization objective function is solved under the premise of meeting the constraint condition set, the constraint condition set includes the building partition connectivity, the passage capacity, the equipment availability, the action timing and the evacuation feasibility, and the linkage control instruction sequence and the evacuation guidance scheme are output.

9. The building fire prediction and intelligent decision optimization method based on deep learning according to claim 1, characterized in that, The building fire emergency disposal closed-loop control is completed, including: The linkage control instruction sequence is sent to the fire linkage controller or the building automation platform, and the candidate linkage action is executed according to the action execution start time and the action execution duration, and the evacuation guidance scheme is published to the building broadcast platform or the mobile terminal; The execution feedback data is collected in real time during the execution process and the feedback time series data is formed, the execution feedback data includes the linkage action execution state, the fire fighting equipment running state, the access state, the passage passable state and the personnel density change data, the emergency decision optimization state vector is updated based on the feedback time series data, and the building fire emergency disposal closed-loop control is completed.

Images