Fire-fighting smoke exhaust skylight driving control method and system based on artificial intelligence

Abstract

The application discloses a fire-fighting smoke exhaust skylight driving control method and system based on artificial intelligence, which comprises the following steps: collecting smoke concentration, temperature, wind pressure, opening degree, driving feedback, sampling time and topological linkage state parameters of a target fire-fighting smoke exhaust skylight and adjacent skylights, and constructing a space-time coupling feature matrix; intercepting a previous period feature matrix and a current period feature matrix according to a preset control period and writing them into a training sample buffer; calculating smoke distribution drift intensity by using a RuLSIF algorithm and generating a drift rising mark; generating a dual-mode parameter set; training an improved PILCO probabilistic dynamics model to obtain a probabilistic dynamics model and an uncertainty boundary; combining strategy search parameters to generate a driving opening degree control quantity and a predicted driving feedback, and writing back an actual driving feedback and a prediction deviation. The application completes model updating, strategy searching and feedback correction according to smoke distribution drift and topological linkage state.

Description

Technical Field

[0001] This invention relates to the field of building fire protection control technology, and in particular to a fire smoke exhaust skylight drive control method and system based on artificial intelligence. Background Technology

[0002] Fire-fighting smoke exhaust skylights are crucial actuators in building fire-fighting smoke exhaust systems, primarily installed on the rooftops of factories, warehouses, and large-span public buildings. In the event of a fire, the skylights adjust their opening to create smoke exhaust channels, allowing smoke, heat, and pressure to escape along the building's roof. They work in conjunction with fans, smoke curtains, fire detectors, and building fire-fighting linkage control equipment to achieve smoke exhaust control. In current engineering applications, the drive control of smoke exhaust skylights is mostly triggered by smoke concentration, temperature, manual alarm signals, and fire-fighting linkage signals. The control logic primarily relies on threshold judgment, zone linkage, and preset opening commands, which can meet basic opening and closing control requirements.

[0003] In existing technologies, the drive control schemes for fire-fighting smoke exhaust skylights mostly employ programmable logic controllers (PLCs), fire alarm control systems, or building automation systems for centralized management. The control system collects detection data from the smoke-proof area and issues an opening command when the detection data reaches a preset threshold, or invokes a pre-set linkage scheme based on the fire zone location to bring the corresponding skylight to a fixed opening degree. Some schemes incorporate historical operating data and simple prediction algorithms to estimate smoke change trends and then correct the opening command; however, their data objects are mostly concentrated on a single skylight or a single smoke-proof area, and the spatial linkage state and drive feedback deviation fail to form a continuous closed-loop constraint.

[0004] It still has limitations in scenarios involving rapid smoke diffusion, wind pressure disturbances, and multi-skylight linkage control. Fixed threshold triggering methods struggle to distinguish between local eddy current disturbances and regional smoke spread, easily leading to delayed opening commands or frequent switching; single-point time-series control methods lack topological constraints between adjacent skylights, making it difficult to reflect synchronous changes caused by smoke propagation along the building space; predictive control schemes do not adequately utilize model uncertainties and actual drive feedback deviations, resulting in the control model being difficult to correct in a timely manner when the fire situation changes dynamically.

[0005] Therefore, how to provide a fire smoke exhaust skylight drive control method and system based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a fire smoke exhaust skylight drive control method and system based on artificial intelligence. This invention generates an opening control quantity by adaptively updating the model based on smoke drift, topological linkage and drive feedback, thereby reducing decision lag under fixed rule triggering.

[0007] A fire smoke exhaust skylight drive control method based on artificial intelligence according to an embodiment of the present invention includes:

[0008] Collect smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights, and construct a spatiotemporal coupling feature matrix;

[0009] The spatiotemporal coupling feature matrix is ​​extracted according to the preset control period to obtain the feature matrix of the previous period and the feature matrix of the current period. After establishing the adjacent correspondence according to the sampling time, it is written into the training sample cache.

[0010] Input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to calculate the flue gas distribution drift intensity and generate drift rise markers;

[0011] A dual-mode parameter set containing model update parameters and policy search parameters is generated based on the flue gas distribution drift intensity and drift rise marker.

[0012] The model update parameters and flue gas distribution drift intensity from the dual-mode parameter set are written into the training process of the improved PILCO probabilistic dynamics model. Weighted updates are performed on the training sample cache to obtain the probabilistic dynamics model and uncertainty boundary.

[0013] The strategy search parameters and uncertainty boundary in the dual-mode parameter set are written into the strategy search process of the probabilistic dynamics model. The uncertainty boundary is used as a dynamic penalty term and constraint boundary. The probabilistic dynamics model is combined to generate the driving opening control quantity and predict the driving feedback.

[0014] The process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

[0015] Optionally, the construction of the spatiotemporal coupled feature matrix specifically involves:

[0016] Collect the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage status parameters of the target fire smoke exhaust skylight and adjacent skylights, and establish candidate sampling records according to the skylight identification;

[0017] Based on the sampling time of the target fire smoke exhaust skylight, neighbor matching is performed on the sampling times of adjacent skylights in the candidate sampling records to obtain the same sampling record group;

[0018] The topological connection position between the target fire smoke exhaust skylight and the adjacent skylight is determined based on the topological linkage state parameters, and the sampling record group at the same moment is spatially sorted to obtain the spatially ordered sampling record group.

[0019] By comparing the direction of wind pressure and opening change in the spatially ordered sampling record group, a topological linkage consistency result is constructed.

[0020] Correction is performed on the spatially ordered sampling record group based on the topological linkage consistency results. When the topological linkage consistency results are consistent, the corresponding adjacent skylight sampling records are retained. When the topological linkage consistency results are inconsistent, the corresponding adjacent skylight sampling records are marked as delayed linkage records, thus obtaining the corrected sampling record group.

[0021] According to the sampling time sequence and topological connection position sequence, the flue gas concentration, temperature, wind pressure, opening degree, drive feedback and delay linkage records in the correction sampling record group are recombined to obtain the spatiotemporal coupling feature matrix.

[0022] Optionally, the step of extracting the spatiotemporal coupling feature matrix according to a preset control period specifically involves:

[0023] Extract the matrix rows from the spatiotemporal coupling feature matrix according to the sampling time sequence, and mark the candidate period boundaries according to the preset control period to obtain the candidate period boundary sequence;

[0024] Extract matrix rows from two adjacent control periods based on the candidate period boundary sequence, and calculate the period connection interval;

[0025] When the period connection interval is inconsistent with the sampling interval corresponding to the preset control cycle, the wind pressure change rate and temperature change rate of adjacent skylights in the corresponding time period are extracted and compared with the preset change rate threshold. The candidate period boundary is moved according to the comparison result until the period connection interval is consistent with the sampling interval corresponding to the preset control cycle, and the corrected period boundary sequence is obtained.

[0026] When the movement span of the candidate cycle boundary exceeds the preset fire evolution safety tolerance time threshold, the movement of the candidate cycle boundary is stopped, and the feature matrix corresponding to the adjacent skylight with the highest topological connection position is extracted according to the topological linkage state parameters to obtain the alternative feature matrix.

[0027] Based on the correction cycle boundary sequence, matrix rows within two adjacent control cycles are extracted from the spatiotemporal coupling feature matrix to obtain the feature matrix of the previous cycle and the feature matrix of the current cycle.

[0028] When stopping the candidate period boundary movement and obtaining the alternative feature matrix, the alternative feature matrix is ​​written into the corresponding missing position in the current period feature matrix according to the sampling time;

[0029] Establish the adjacent correspondence between the feature matrix of the previous period and the feature matrix of the current period according to the sampling time, and write it into the training sample cache.

[0030] Optionally, the step of inputting the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm specifically involves:

[0031] Based on the adjacent correspondence, the matrix rows corresponding to the sampling time are extracted from the feature matrix of the previous period and the feature matrix of the current period, and alignment is performed according to the topological connection position order to obtain the reference sample sequence and the sample sequence to be tested;

[0032] Delayed linkage records are extracted from the reference sample sequence and the sample sequence to be tested, and linkage effective labels are established for flue gas concentration, temperature, wind pressure, opening degree and drive feedback under the same topological connection position to obtain the labeled sample sequence.

[0033] The labeled sample sequence is hierarchically aggregated according to the topological connection position. The sample corresponding to the target fire smoke exhaust skylight is assigned to the target layer, and the sample corresponding to the adjacent skylight is assigned to the adjacent layer according to the topological connection position, thus obtaining a hierarchical sample set.

[0034] The hierarchical sample set is input into the RuLSIF algorithm, and the participation weight of the corresponding sample in the relative density ratio estimation is adjusted according to the effective linkage label to obtain the hierarchical relative density ratio sequence.

[0035] Based on the deviation of each relative density ratio in the layered relative density ratio sequence from the preset reference density ratio, accumulation is performed according to the arrangement order of the target layer and adjacent layers to obtain the flue gas distribution drift intensity;

[0036] The flue gas distribution drift intensity is compared with the flue gas distribution drift intensity corresponding to the previous control cycle. When the flue gas distribution drift intensity increases and the deviation direction of the target layer and the adjacent layer is consistent, a drift increase mark is generated.

[0037] Optionally, the generation of a dual-mode parameter set, including model update parameters and policy search parameters, specifically involves:

[0038] By comparing the flue gas distribution drift intensity corresponding to the current control cycle with the flue gas distribution drift intensity corresponding to the previous control cycle, the relationship of drift intensity change can be obtained.

[0039] The drift intensity of the flue gas distribution corresponding to the current control cycle is matched with the preset drift interval sequence to obtain the drift intensity level. The drift intensity level and the relationship between the drift intensity change are marked accordingly to obtain the drift state marking result.

[0040] When the drift rise marker is established, the training sample weight adjustment amount, the model update step size adjustment amount, and the uncertainty boundary adjustment amount are increased according to the drift state marker results to obtain the model update parameters.

[0041] When the drift rise marker is not established, the model update parameters are obtained by maintaining the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment according to the drift state marker results.

[0042] When the drift rise flag is established, the policy penalty adjustment amount and the opening change constraint amount are increased according to the drift state flag result, and the rolling prediction step size is shortened to obtain the policy search parameters.

[0043] When the drift rise marker is not established, the policy search parameters are obtained by maintaining the policy penalty adjustment amount, the opening change constraint amount, and the rolling prediction step size according to the drift state marker results.

[0044] Establish a correspondence between the model update parameters and the policy search parameters according to the current control cycle, and generate a dual-mode parameter set.

[0045] Optionally, obtaining the probabilistic dynamics model and uncertainty boundary specifically involves:

[0046] Extract the previous period feature matrix, the current period feature matrix and adjacent correspondence from the training sample cache, and extract the model update parameters from the dual-mode parameter set;

[0047] Based on the adjacent correspondence, the matrix rows corresponding to the sampling time in the previous period feature matrix and the current period feature matrix are paired to obtain periodic paired samples. According to the topological connection position, the target fire smoke exhaust skylight corresponding features and adjacent skylight corresponding features are extracted from the periodic paired samples to obtain spatiotemporal paired training samples.

[0048] Extract the flue gas concentration, temperature, wind pressure, opening degree and driving feedback corresponding to the previous period from the spatiotemporal paired training samples to form state input features, and extract the change of the corresponding fields in the current period relative to the previous period to form state transition output features.

[0049] Based on the topological connection position, the state input features are divided into target layer input features and adjacent layer input features, and a correspondence is established with the state transition output features to obtain a hierarchical state transition sample set.

[0050] Extract the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment from the model update parameters;

[0051] Based on the flue gas distribution drift intensity and the training sample weight adjustment, weight labeling is performed on the stratified state transition sample set to obtain the drift-weighted training sample set and weight labeling results.

[0052] According to the topological connection positions of the target layer and adjacent layers, kernel response calculation is performed on the state input features in the drift-weighted training sample set, and the kernel response calculation results are weighted according to the weight labeling results to obtain the spatiotemporal coupled kernel response matrix;

[0053] The PILCO probabilistic dynamics model is trained and improved based on the spatiotemporal coupling kernel response matrix and state transition output features to obtain candidate probabilistic dynamics model parameters and candidate prediction variances.

[0054] Based on the model update step size adjustment, the parameters of the candidate probabilistic dynamics model and the probabilistic dynamics model parameters of the previous control cycle are fused to obtain the probabilistic dynamics model of the current control cycle.

[0055] The uncertainty boundary for the current control cycle is generated based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment, and a correspondence is established between the probabilistic dynamics model and the uncertainty boundary for the current control cycle.

[0056] Optionally, the improved PILCO probabilistic dynamics model includes a spatiotemporal feature input layer, a topological hierarchical indexing layer, an initial kernel response calculation layer, a drift weight injection layer, a state transition regression layer, a variance boundary generation layer, and a parameter adaptive fusion output layer.

[0057] The spatiotemporal feature input layer receives target layer input features, adjacent layer input features, and state transition output features from the hierarchical state transition sample set, and establishes state transition training input according to sampling time and field order;

[0058] The topology hierarchical index layer divides the state transition training input into target layer index records and adjacent layer index records according to the topology connection position.

[0059] The initial kernel response calculation layer performs kernel response calculation on the target layer input features and the adjacent layer input features based on the target layer index record and the adjacent layer index record to generate an initial kernel response matrix;

[0060] The drift weight injection layer receives the initial kernel response matrix and the weight labeling result, and maps the weight labeling result as an environmental attenuation penalty term to the matrix elements in the initial kernel response matrix corresponding to the sampling time and topological connection position, to obtain the spatiotemporal coupled kernel response matrix.

[0061] The state transition regression layer receives the spatiotemporal coupling kernel response matrix and state transition output features, and generates candidate probabilistic dynamic model parameters and candidate prediction variances in the order of flue gas concentration, temperature, wind pressure, opening degree and drive feedback.

[0062] The variance boundary generation layer generates the uncertainty boundary for the current control cycle based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment amount, and then passes it to the parameter adaptive fusion output layer.

[0063] The parameter adaptive fusion output layer receives candidate probabilistic dynamics model parameters, probabilistic dynamics model parameters from the previous control cycle, model update step size adjustment, and uncertainty boundary. It adjusts the allocation ratio of the model update step size adjustment in the parameter fusion process based on the magnitude change of the uncertainty boundary, and outputs the probabilistic dynamics model and uncertainty boundary for the current control cycle.

[0064] Optionally, the step of combining the probabilistic dynamics model to generate the driving opening control quantity and predict the driving feedback specifically involves:

[0065] Extract the policy penalty adjustment, opening change constraint, and rolling prediction step size from the policy search parameters in the dual-mode parameter set, and call the probabilistic dynamics model and uncertainty boundary of the current control cycle;

[0066] A candidate opening trajectory set is established based on the current opening degree and opening degree change constraint of the target fire smoke exhaust skylight, and a prediction time domain index is established for the candidate opening trajectory set according to the rolling prediction step size;

[0067] The candidate opening trajectory set is input into the probabilistic dynamics model, and the predicted states of flue gas concentration, temperature, wind pressure, opening degree and driving feedback are recursively deduced according to the prediction time domain index to obtain the candidate prediction state sequence.

[0068] The uncertainty boundary is written into the candidate prediction state sequence in the order of the fields in the boundary comparison process. The positions of the fields that exceed the uncertainty boundary and the prediction time domain index are counted to obtain the boundary deviation record.

[0069] The dynamic penalty value is calculated based on the boundary deviation record and the policy penalty adjustment amount, and the dynamic penalty value is written into the trajectory cost calculation process of the candidate opening trajectory set to obtain the candidate trajectory cost sequence;

[0070] The candidate opening trajectory with the lowest cost is selected from the candidate trajectory cost sequence, and the opening value corresponding to the selected candidate opening trajectory in the first control cycle is determined as the driving opening control quantity.

[0071] Extract the driving feedback prediction state corresponding to the selected candidate opening trajectories, and determine the driving feedback prediction state as the prediction driving feedback.

[0072] Optionally, the process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle specifically includes:

[0073] According to the sampling time of the current control cycle, establish a control issuance record between the drive opening control quantity and the skylight identification of the target fire smoke exhaust skylight, and transmit the control issuance record to the target skylight drive mechanism;

[0074] The actual drive feedback returned by the target sunroof drive mechanism within the corresponding control cycle of the control issuance record is collected, and feedback alignment records are established between the actual drive feedback and the predicted drive feedback according to the sampling time.

[0075] The absolute value and direction of the difference between the actual driving feedback and the predicted driving feedback are calculated based on the feedback alignment record to obtain the prediction deviation record;

[0076] The prediction deviation record is compared with the opening change constraint. A feedback lag flag is generated when the prediction deviation record exceeds the opening change constraint, and a feedback follow flag is generated when the prediction deviation record does not exceed the opening change constraint.

[0077] The drive feedback field corresponding to the target fire smoke exhaust skylight in the next control cycle is corrected based on the feedback lag mark and feedback follow mark, and the prediction deviation record is written into the delay linkage record at the corresponding sampling time.

[0078] The corrected drive feedback field, prediction deviation record, and delay linkage record are recombined with the flue gas concentration, temperature, wind pressure, opening degree, and topology linkage state parameters collected in the next control cycle to obtain the spatiotemporal coupling feature matrix construction input for the next control cycle.

[0079] Optionally, an artificial intelligence-based fire smoke exhaust skylight drive control system includes:

[0080] The spatiotemporal coupling matrix construction module is used to generate a spatiotemporal coupling feature matrix based on the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights.

[0081] The periodic truncation cache module is used to truncate the spatiotemporal coupled feature matrix according to a preset control period, generate the feature matrix of the previous period, the feature matrix of the current period and the adjacent correspondence, and write them into the training sample cache.

[0082] The drift detection module is used to input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to generate the flue gas distribution drift intensity and drift rise marker.

[0083] The dual-mode parameter generation module is used to generate a dual-mode parameter set containing model update parameters and strategy search parameters based on the flue gas distribution drift intensity and drift rise marker.

[0084] The probabilistic dynamics training module is used to write the model update parameters and flue gas distribution drift intensity into the training process of the improved PILCO probabilistic dynamics model, perform weighted updates on the training sample cache, and generate the probabilistic dynamics model and uncertainty boundary.

[0085] The strategy search decision module is used to write the strategy search parameters and uncertainty boundary into the strategy search process of the probabilistic dynamics model, and generate the driving opening control quantity and predict the driving feedback.

[0086] The control feedback write-back module is used to transmit the drive opening control quantity to the target sunroof drive mechanism, generate a prediction deviation based on the actual drive feedback and the predicted drive feedback, and write the actual drive feedback and the prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

[0087] The beneficial effects of this invention are:

[0088] (1) This invention collects the smoke concentration, temperature, wind pressure, opening degree, driving feedback, sampling time and topological linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights, and constructs a spatiotemporal coupling feature matrix to realize the synchronous expression of the operating state of a single skylight and the spatial linkage state of adjacent skylights. Thus, it can provide a data foundation with spatiotemporal correlation constraints for subsequent drift detection, model training and opening degree control in building fire protection scenarios where local smoke disturbance and regional spread coexist.

[0089] (2) This invention inputs the previous period feature matrix and the current period feature matrix into the RuLSIF algorithm to calculate the flue gas distribution drift intensity and generate a drift rise marker. Then, based on the flue gas distribution drift intensity and the drift rise marker, a dual-mode parameter set containing model update parameters and strategy search parameters is generated. This realizes the linkage adjustment between fire state changes and model updates and strategy searches, thereby reducing the control lag and unnecessary frequent updates caused by fixed rule triggering methods under dynamic changes in flue gas.

[0090] (3) The present invention improves the PILCO probabilistic dynamics model by introducing the smoke distribution drift intensity, uncertainty boundary, spatiotemporal coupling kernel response matrix and feedback deviation write-back mechanism in the probabilistic dynamics model training and strategy search stages. This realizes weighted update of training samples, adaptive fusion of model parameters, opening trajectory constraint search and drive feedback correction, thereby enhancing the adaptability, stability and feedback closed-loop capability of fire smoke exhaust skylight drive control in fire dynamic environment. Attached Figure Description

[0091] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0092] Figure 1 This is a flowchart of a fire smoke exhaust skylight drive control method based on artificial intelligence proposed in this invention;

[0093] Figure 2 This is a system structure block diagram of a fire smoke exhaust skylight drive control system based on artificial intelligence proposed in this invention;

[0094] Figure 3 This is a structural diagram of the improved PILCO probabilistic dynamics model for a fire smoke exhaust skylight drive control method based on artificial intelligence proposed in this invention. Detailed Implementation

[0095] Example 1: Reference Figures 1-3 A fire smoke exhaust skylight drive control method based on artificial intelligence, comprising:

[0096] Collect smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights, and construct a spatiotemporal coupling feature matrix;

[0097] The spatiotemporal coupling feature matrix is ​​extracted according to the preset control period to obtain the feature matrix of the previous period and the feature matrix of the current period. After establishing the adjacent correspondence according to the sampling time, it is written into the training sample cache.

[0098] Input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to calculate the flue gas distribution drift intensity and generate drift rise markers;

[0099] A dual-mode parameter set containing model update parameters and policy search parameters is generated based on the flue gas distribution drift intensity and drift rise marker.

[0100] The model update parameters and flue gas distribution drift intensity from the dual-mode parameter set are written into the training process of the improved PILCO probabilistic dynamics model. Weighted updates are performed on the training sample cache to obtain the probabilistic dynamics model and uncertainty boundary.

[0101] The strategy search parameters and uncertainty boundary in the dual-mode parameter set are written into the strategy search process of the probabilistic dynamics model. The uncertainty boundary is used as a dynamic penalty term and constraint boundary. The probabilistic dynamics model is combined to generate the driving opening control quantity and predict the driving feedback.

[0102] The process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

[0103] In this embodiment, the construction of the spatiotemporal coupling feature matrix specifically involves:

[0104] Collect the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage status parameters of the target fire smoke exhaust skylight and adjacent skylights, and establish candidate sampling records according to the skylight identification;

[0105] Based on the sampling time of the target fire smoke exhaust skylight, neighbor matching is performed on the sampling times of adjacent skylights in the candidate sampling records to obtain the same sampling record group;

[0106] In Example 1, each sampling time of the target fire smoke exhaust skylight is set as the reference time. Sample records to be matched are extracted from the candidate sampling records according to the skylight identifiers of adjacent skylights. The absolute value of the time difference between the sampling time and the reference time of each sample record to be matched is calculated. Sample records to be matched with an absolute time difference not exceeding 0.5 seconds are included in the candidate matching set. When multiple candidate matching records exist for the same adjacent skylight, the candidate matching record with the smallest absolute time difference is selected. The candidate sampling records of the target fire smoke exhaust skylight at the reference time are merged with the candidate matching records selected from each adjacent skylight to obtain a simultaneous sampling record group.

[0107] The topological connection position between the target fire smoke exhaust skylight and the adjacent skylight is determined based on the topological linkage state parameters, and the sampling record group at the same moment is spatially sorted to obtain the spatially ordered sampling record group.

[0108] By comparing the direction of wind pressure and opening change in the spatially ordered sampling record group, a topological linkage consistency result is constructed.

[0109] Correction is performed on the spatially ordered sampling record group based on the topological linkage consistency results. When the topological linkage consistency results are consistent, the corresponding adjacent skylight sampling records are retained. When the topological linkage consistency results are inconsistent, the corresponding adjacent skylight sampling records are marked as delayed linkage records, thus obtaining the corrected sampling record group.

[0110] In Example 1, adjacent skylight sampling records in the spatially ordered sampling record group are checked one by one according to the topology connection position. When the topology linkage consistency results are consistent, the adjacent skylight sampling records along with the original topology connection positions are written into the correction sampling record group, and the delayed linkage record is recorded as 0. When the topology linkage consistency results are inconsistent, the sampling values ​​and original topology connection positions of the adjacent skylight sampling records are retained, and the delayed linkage record is recorded as 1.

[0111] According to the sampling time sequence and topological connection position sequence, the flue gas concentration, temperature, wind pressure, opening degree, drive feedback and delay linkage records in the correction sampling record group are recombined to obtain the spatiotemporal coupling feature matrix.

[0112] In Example 1, a matrix row index is established according to the sampling time order, and a matrix column index is established according to the topological connection position order, with the target fire smoke exhaust skylight placed in the first column; in the row corresponding to each sampling time, the sampling values ​​of each skylight are written in the fixed field order of smoke concentration, temperature, wind pressure, opening degree, drive feedback, and delay linkage record, and the position of missing matching record uses the field value of the same column of the previous sampling time and sets the delay linkage record to 1; the rows corresponding to each sampling time are arranged continuously in chronological order to form a spatiotemporal coupling feature matrix.

[0113] In this embodiment, the step of extracting the spatiotemporal coupling feature matrix according to a preset control period specifically involves:

[0114] The matrix rows in the spatiotemporal coupling feature matrix are extracted according to the sampling time sequence, and the candidate period boundaries are marked according to the preset control period to obtain the candidate period boundary sequence. The preset control period is set to 5 seconds, the sampling interval is set to 1 second, and the matrix rows corresponding to 5 consecutive sampling times are extracted in each control period.

[0115] Extract matrix rows from two adjacent control periods based on the candidate period boundary sequence, and calculate the period connection interval;

[0116] In Example 1, two adjacent candidate cycle boundaries in the candidate cycle boundary sequence are respectively determined as the end boundary of the previous control cycle and the start boundary of the current control cycle; the last sampling time is extracted from the matrix row corresponding to the end boundary of the previous control cycle, and the first sampling time is extracted from the matrix row corresponding to the start boundary of the current control cycle; the last sampling time is subtracted from the first sampling time to obtain the time difference; the absolute value of the time difference is converted, and the converted value is determined as the cycle connection interval; when the cycle connection interval is 1 second, it is determined that the cycle connection interval is consistent with the sampling interval corresponding to the preset control cycle.

[0117] When the periodic connection interval is inconsistent with the sampling interval corresponding to the preset control cycle, the wind pressure change rate and temperature change rate of adjacent skylights in the corresponding time period are extracted and compared with the preset change rate threshold. The candidate period boundary is moved according to the comparison result until the periodic connection interval is consistent with the sampling interval corresponding to the preset control cycle, and a corrected period boundary sequence is obtained. When either the wind pressure change rate or the temperature change rate exceeds the preset change rate threshold, the candidate period boundary is moved successively according to the sampling interval corresponding to the adjacent matrix rows in the spatiotemporal coupling feature matrix. When neither the wind pressure change rate nor the temperature change rate exceeds the preset change rate threshold, the candidate period boundary is moved according to the sampling interval corresponding to the preset control cycle. The preset change rate threshold is set to 0.2. The wind pressure change rate and temperature change rate are calculated as the ratio of the absolute value of the difference between adjacent sampled values ​​to the previous sampled value.

[0118] When the movement span of the candidate cycle boundary exceeds the preset fire evolution safety tolerance time threshold, the movement of the candidate cycle boundary is stopped, and the feature matrix corresponding to the adjacent skylight with the highest topological connection position is extracted according to the topological linkage state parameters to obtain the alternative feature matrix. The preset fire evolution safety tolerance time threshold is determined to be 10 seconds according to 2 preset control cycles. When the cumulative movement span of the candidate cycle boundary reaches 10 seconds, the movement of the candidate cycle boundary is stopped.

[0119] Based on the correction cycle boundary sequence, matrix rows within two adjacent control cycles are extracted from the spatiotemporal coupling feature matrix to obtain the feature matrix of the previous cycle and the feature matrix of the current cycle.

[0120] When stopping the candidate period boundary movement and obtaining the alternative feature matrix, the alternative feature matrix is ​​written into the corresponding missing position in the current period feature matrix according to the sampling time;

[0121] Establish the adjacent correspondence between the feature matrix of the previous period and the feature matrix of the current period according to the sampling time, and write it into the training sample cache.

[0122] In Example 1, the matrix rows in the previous period feature matrix and the current period feature matrix are labeled with sequential numbers from 1 to 5 according to time sequence; the matrix rows with the same sequential number in the previous period feature matrix and the matrix rows with the same sequential number in the current period feature matrix are paired records; two rows of sampling time, window position, field order and delay linkage records are written in the paired records to form an adjacent correspondence relationship.

[0123] In this embodiment, the step of inputting the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm specifically involves:

[0124] Based on the adjacent correspondence, the matrix rows corresponding to the sampling time are extracted from the feature matrix of the previous period and the feature matrix of the current period, and alignment is performed according to the topological connection position order to obtain the reference sample sequence and the sample sequence to be tested;

[0125] Delayed linkage records are extracted from the reference sample sequence and the sample sequence to be tested, and linkage effective labels are established for flue gas concentration, temperature, wind pressure, opening degree and drive feedback under the same topological connection position to obtain the labeled sample sequence.

[0126] In Example 1, delayed linkage records are extracted from the reference sample sequence and the test sample sequence according to the topological connection position. When the delayed linkage record is 0, the linkage validity flags of flue gas concentration, temperature, wind pressure, opening degree, and drive feedback under the same topological connection position are all recorded as 1. When the delayed linkage record is 1, the linkage validity flags of flue gas concentration, temperature, wind pressure, opening degree, and drive feedback under the same topological connection position are all recorded as 0. The linkage validity flags, sampling time, topological connection position, and field order are written into the labeled sample sequence.

[0127] The labeled sample sequence is hierarchically aggregated according to the topological connection position. The sample corresponding to the target fire smoke exhaust skylight is assigned to the target layer, and the sample corresponding to the adjacent skylight is assigned to the adjacent layer according to the topological connection position, thus obtaining a hierarchical sample set.

[0128] The hierarchical sample set is input into the RuLSIF algorithm, and the participation weight of the corresponding sample in the relative density ratio estimation is adjusted according to the effective linkage label to obtain the hierarchical relative density ratio sequence.

[0129] In Example 1, reference sample tables and test sample tables are established for the target layer and adjacent layers within the hierarchical sample set, respectively. Each sample is composed of a feature vector according to the field order of flue gas concentration, temperature, wind pressure, opening degree, and drive feedback. The RuLSIF algorithm uses a Gaussian kernel function, and the kernel function bandwidth is taken as the median of the Euclidean distance between adjacent feature vectors after merging the reference sample table and the test sample table in the same layer. The kernel center is extracted from the merged feature vector according to the sampling time interval. The ridge regularization coefficient is set to 0.001, and the relative mixing coefficient is set to 0.5. When the linkage effective flag is 1, the corresponding sample participation weight is recorded as 1; when the linkage effective flag is 0, the corresponding sample participation weight is recorded as 0.2. The participation weights are written into the kernel response weighted statistical process of the RuLSIF algorithm. The relative density ratio of the reference sample table to the test sample table is solved for the target layer and adjacent layers, respectively, to obtain the hierarchical relative density ratio sequence.

[0130] Based on the deviation of each relative density ratio in the stratified relative density ratio sequence from the preset reference density ratio, where the preset reference density ratio is set to 1, the absolute value of the stratified relative density ratios of the target layer and adjacent layers is obtained after subtracting 1, and the corresponding deviation is obtained. The accumulation is performed according to the arrangement order of the target layer and adjacent layers to obtain the flue gas distribution drift intensity.

[0131] In Example 1, the target layer is arranged first, and the adjacent layers are arranged from near to far according to their topological connection positions to form a hierarchical accumulation order. The initial accumulation value is set to 0. For each relative density ratio in the hierarchical relative density ratio sequence, a preset benchmark density ratio of 1 is subtracted and the absolute value is taken to obtain the individual deviation magnitude. According to the hierarchical accumulation order, the individual deviation magnitudes in the target layer are added to the initial accumulation value one by one, and then the individual deviation magnitudes in the adjacent layers are added to the accumulation value one by one according to their topological connection positions. After all individual deviation magnitudes are accumulated, the final accumulation value is recorded as the flue gas distribution drift intensity.

[0132] The flue gas distribution drift intensity is compared with the flue gas distribution drift intensity corresponding to the previous control cycle. When the flue gas distribution drift intensity increases and the deviation direction of the target layer and the adjacent layer is consistent, a drift increase mark is generated.

[0133] In Example 1, the number of categories with a relative density ratio greater than 1 and less than 1 within the target layer are counted, and the category with the larger number is taken as the deviation direction of the target layer. The number of categories with a relative density ratio greater than 1 and less than 1 within the adjacent layer are counted, and the category with the larger number is taken as the deviation direction of the adjacent layer. When the number of two categories within the same layer is the same, the deviation direction of this layer is recorded as neutral. When both the deviation direction of the target layer and the deviation direction of the adjacent layer are categories greater than 1, the deviation directions of the target layer and the adjacent layer are determined to be consistent. When both the deviation direction of the target layer and the deviation direction of the adjacent layer are categories less than 1, the deviation directions of the target layer and the adjacent layer are determined to be consistent. When the deviation direction of any layer is neutral, the deviation directions of the target layer and the adjacent layer are determined to be inconsistent.

[0134] In this embodiment, generating a dual-mode parameter set containing model update parameters and policy search parameters specifically involves:

[0135] By comparing the flue gas distribution drift intensity corresponding to the current control cycle with the flue gas distribution drift intensity corresponding to the previous control cycle, the relationship of drift intensity change can be obtained.

[0136] The flue gas distribution drift intensity corresponding to the current control cycle is matched with a preset drift interval sequence. The preset drift interval sequence is set as low drift interval 0 to 0.5, medium drift interval greater than 0.5 to 1.5, and high drift interval greater than 1.5. When the flue gas distribution drift intensity falls into the corresponding interval, the drift intensity level is obtained, and the drift intensity level and the drift intensity change relationship are marked accordingly to obtain the drift state marking result.

[0137] In Example 1, the drift intensity value of the flue gas distribution corresponding to the current control cycle is obtained. The value is then compared with the boundaries of the low drift range (0 to 0.5), the medium drift range (greater than 0.5 to 1.5), and the high drift range (greater than 1.5). When the value is not less than 0 and not greater than 0.5, the drift intensity level is recorded as a low drift level; when the value is greater than 0.5 and not greater than 1.5, the drift intensity level is recorded as a medium drift level; and when the value is greater than 1.5, the drift intensity level is recorded as a high drift level. The drift intensity level and the relationship between drift intensity changes are written into the drift state marking result.

[0138] When the drift rise marker is established, the training sample weight adjustment amount, the model update step size adjustment amount, and the uncertainty boundary adjustment amount are increased according to the drift state marker results to obtain the model update parameters.

[0139] When the drift rise marker is not established, the model update parameters are obtained by maintaining the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment according to the drift state marker results.

[0140] When the drift rise flag is established, the policy penalty adjustment amount and the opening change constraint amount are increased according to the drift state flag result, and the rolling prediction step size is shortened to obtain the policy search parameters.

[0141] When the drift rise marker is not established, the policy search parameters are obtained by maintaining the policy penalty adjustment amount, the opening change constraint amount, and the rolling prediction step size according to the drift state marker results.

[0142] In Example 1, the initial values ​​for the training sample weight adjustment are set to 1.00, the initial values ​​for the model update step size adjustment are set to 0.05, the initial values ​​for the uncertainty boundary adjustment are set to 0.10, the initial values ​​for the policy penalty adjustment are set to 1.00, the initial values ​​for the opening change constraint are set to 5 degrees, and the initial values ​​for the rolling prediction step size are set to 5 control cycles. When the drift state labeling result indicates a low drift level and the drift increase label is established, the training sample weight adjustment is increased to 1.10, the model update step size adjustment is increased to 0.06, the uncertainty boundary adjustment is increased to 0.12, the policy penalty adjustment is increased to 1.10, the opening change constraint is increased to 6 degrees, and the rolling prediction step size is shortened to 4 control cycles. When the drift state labeling result indicates a medium drift level and the drift increase label is established, the training sample weight adjustment is increased to 1.30, the model update step size adjustment is increased to 0.08, and the uncertainty boundary adjustment is increased to 0.1. 6. Increase the policy penalty adjustment to 1.30, increase the opening degree change constraint to 8 degrees, and shorten the rolling prediction step size to 3 control cycles. When the drift state labeling result indicates a high drift level and the drift rise label is established, increase the training sample weight adjustment to 1.60, increase the model update step size adjustment to 0.10, increase the uncertainty boundary adjustment to 0.20, increase the policy penalty adjustment to 1.60, increase the opening degree change constraint to 10 degrees, and shorten the rolling prediction step size to 2 control cycles. When the drift rise label is not established, use the training sample weight adjustment, model update step size adjustment, uncertainty boundary adjustment, policy penalty adjustment, opening degree change constraint, and rolling prediction step size corresponding to the previous control cycle. Combine the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment into model update parameters, and combine the policy penalty adjustment, opening degree change constraint, and rolling prediction step size into policy search parameters.

[0143] Establish a correspondence between the model update parameters and the policy search parameters according to the current control cycle, and generate a dual-mode parameter set.

[0144] In Example 1, a cycle number is assigned to the current control cycle, and the start and end sampling times of the current control cycle are recorded. The training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment in the model update parameters are associated with the cycle number. The policy penalty adjustment, opening change constraint, and rolling prediction step size in the policy search parameters are associated with the same cycle number. The two sets of associated labels are merged according to the cycle number to generate a dual-mode parameter set.

[0145] In this embodiment, obtaining the probabilistic dynamics model and uncertainty boundary specifically involves:

[0146] Extract the previous period feature matrix, the current period feature matrix and adjacent correspondence from the training sample cache, and extract the model update parameters from the dual-mode parameter set;

[0147] Based on the adjacent correspondence, the matrix rows corresponding to the sampling time in the previous period feature matrix and the current period feature matrix are paired to obtain periodic paired samples. According to the topological connection position, the target fire smoke exhaust skylight corresponding features and adjacent skylight corresponding features are extracted from the periodic paired samples to obtain spatiotemporal paired training samples.

[0148] In Example 1, each matrix row in the previous period's feature matrix is ​​labeled with row numbers 1 to 5 according to the sampling time sequence; each matrix row in the current period's feature matrix is ​​labeled with row numbers 1 to 5 according to the sampling time sequence; two matrix rows with the same row number and the same field order are paired up to form a pairing record; the pairing record contains the sampling time of the previous period, the sampling time of the current period, the topology connection position, the flue gas concentration, temperature, wind pressure, opening degree, drive feedback, and delay linkage record; 5 pairing records form a periodic pairing sample;

[0149] Extract the flue gas concentration, temperature, wind pressure, opening degree and driving feedback corresponding to the previous period from the spatiotemporal paired training samples to form state input features, and extract the change of the corresponding fields in the current period relative to the previous period to form state transition output features.

[0150] Based on the topological connection position, the state input features are divided into target layer input features and adjacent layer input features, and a correspondence is established with the state transition output features to obtain a hierarchical state transition sample set.

[0151] In Example 1, the status input features include skylight identifiers, topology connection positions, and field order markers. Status input features whose skylight identifiers match the target fire exhaust skylight identifiers are classified as target layer input features. Status input features whose skylight identifiers match adjacent skylight identifiers and whose topology connection positions are greater than 0 are classified as adjacent layer input features. Adjacent layer input features are arranged from nearest to farthest according to topology connection positions, and within the same topology connection position, they are arranged according to sampling time and field order. Both target layer input features and adjacent layer input features retain the fields of smoke concentration, temperature, wind pressure, opening degree, and drive feedback.

[0152] Extract the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment from the model update parameters;

[0153] Based on the flue gas distribution drift intensity and the training sample weight adjustment, weight labeling is performed on the stratified state transition sample set to obtain the drift-weighted training sample set and weight labeling results.

[0154] In Example 1, a basic weight is set for each sample in the hierarchical state transition sample set. The basic weight of the target layer sample is 1, the basic weight of the adjacent layer sample with a topological connection position of 1 is 0.8, the basic weight of the adjacent layer sample with a topological connection position of 2 is 0.6, and the basic weight of the adjacent layer sample with a topological connection position greater than 2 is 0.4. The drift coefficient is 1 when the flue gas distribution drift intensity is not greater than 0.5, 1.2 when the flue gas distribution drift intensity is greater than 0.5 but not greater than 1.5, and 1.5 when the flue gas distribution drift intensity is greater than 1.5. The basic weight, drift coefficient, and training sample weight adjustment are multiplied to obtain the weight label of each sample.

[0155] According to the topological connection positions of the target layer and adjacent layers, kernel response calculation is performed on the state input features in the drift-weighted training sample set, and the kernel response calculation results are weighted according to the weight labeling results to obtain the spatiotemporal coupled kernel response matrix;

[0156] In Example 1, the kernel center sample takes all state input features from the drift-weighted training sample set; the state input features are arranged in order of target layer first, and adjacent layers in order of topological connection position from near to far; the sum of squared differences between each state input feature and each kernel center sample is calculated field by field, divided by twice the square of the bandwidth, and the negative exponent of the natural base is taken to obtain the kernel response value; the bandwidth is set to 1; the kernel response value is multiplied by the corresponding weight label to obtain the weighted kernel response value; and written into the matrix in the order of arrangement to obtain the spatiotemporal coupled kernel response matrix.

[0157] The PILCO probabilistic dynamics model is trained and improved based on the spatiotemporal coupling kernel response matrix and state transition output features to obtain candidate probabilistic dynamics model parameters and candidate prediction variances.

[0158] Based on the model update step size adjustment, the parameters of the candidate probabilistic dynamics model and the probabilistic dynamics model parameters of the previous control cycle are fused to obtain the probabilistic dynamics model of the current control cycle.

[0159] In Example 1, a one-to-one correspondence is established between candidate probabilistic dynamics model parameters and probabilistic dynamics model parameters of the previous control cycle according to the model parameter names; for each model parameter with the same name, the parameter difference is obtained by subtracting the parameter value of the previous control cycle from the candidate parameter value; the parameter difference is multiplied by the model update step size adjustment to obtain the parameter increment; the parameter increment is added to the parameter value of the previous control cycle to obtain the parameter value of the current control cycle; all the parameter values ​​of the current control cycle constitute the probabilistic dynamics model parameter set of the current control cycle.

[0160] The uncertainty boundary for the current control cycle is generated based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment, and a correspondence is established between the probabilistic dynamics model and the uncertainty boundary for the current control cycle.

[0161] In Example 1, the candidate prediction variances are converted into standard deviations in the order of flue gas concentration, temperature, wind pressure, opening degree, and drive feedback fields. The square of the standard deviation is equal to the corresponding candidate prediction variance. The drift boundary coefficient is recorded as 1 when the flue gas distribution drift intensity is not greater than 0.5, 1.2 when it is greater than 0.5 but not greater than 1.5, and 1.5 when it is greater than 1.5. The standard deviation, drift boundary coefficient, and uncertainty boundary adjustment are multiplied to obtain the boundary half-width corresponding to each field. The boundary half-widths corresponding to each field are combined in the order of the fields to generate the uncertainty boundary for the current control cycle.

[0162] In this embodiment, the improved PILCO probabilistic dynamics model includes a spatiotemporal feature input layer, a topological hierarchical index layer, an initial kernel response calculation layer, a drift weight injection layer, a state transition regression layer, a variance boundary generation layer, and a parameter adaptive fusion output layer.

[0163] The spatiotemporal feature input layer receives target layer input features, adjacent layer input features, and state transition output features from the hierarchical state transition sample set, and establishes state transition training input according to sampling time and field order;

[0164] The topology hierarchical index layer divides the state transition training input into target layer index records and adjacent layer index records according to the topology connection position, and establishes a correspondence between target layer index records, adjacent layer index records and state transition output features;

[0165] The initial kernel response calculation layer performs kernel response calculation on the target layer input features and the adjacent layer input features based on the target layer index record and the adjacent layer index record to generate an initial kernel response matrix;

[0166] The drift weight injection layer receives the initial kernel response matrix and the weight labeling result, and maps the weight labeling result as an environmental attenuation penalty term to the matrix elements in the initial kernel response matrix corresponding to the sampling time and topological connection position, to obtain the spatiotemporal coupled kernel response matrix.

[0167] The state transition regression layer receives the spatiotemporal coupling kernel response matrix and state transition output features, and generates candidate probabilistic dynamic model parameters and candidate prediction variances in the order of flue gas concentration, temperature, wind pressure, opening degree and drive feedback.

[0168] The variance boundary generation layer generates the uncertainty boundary for the current control cycle based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment amount, and then passes it to the parameter adaptive fusion output layer.

[0169] The parameter adaptive fusion output layer receives candidate probabilistic dynamics model parameters, probabilistic dynamics model parameters from the previous control cycle, model update step size adjustment, and uncertainty boundary. It adjusts the allocation ratio of the model update step size adjustment in the parameter fusion process based on the magnitude change of the uncertainty boundary, and outputs the probabilistic dynamics model and uncertainty boundary for the current control cycle.

[0170] In this embodiment, the step of combining the probabilistic dynamics model to generate the driving opening control quantity and predict the driving feedback specifically involves:

[0171] Extract the policy penalty adjustment, opening change constraint, and rolling prediction step size from the policy search parameters in the dual-mode parameter set, and call the probabilistic dynamics model and uncertainty boundary of the current control cycle;

[0172] A candidate opening trajectory set is established based on the current opening degree and opening degree change constraint of the target fire smoke exhaust skylight, and a prediction time domain index is established for the candidate opening trajectory set according to the rolling prediction step size;

[0173] In Example 1, the mechanical fully closed opening of the target fire smoke exhaust skylight is set to 0 degrees, and the mechanical fully open opening is set to 90 degrees. The current opening minus the opening change constraint, the current opening, and the current opening plus the opening change constraint are taken as three candidate opening values. When the candidate opening value is less than 0 degrees, it is corrected to 0 degrees; when it is greater than 90 degrees, it is corrected to 90 degrees. A prediction time domain index of 1 to N is established according to the rolling prediction step size, where N is the value of the rolling prediction step size. The three candidate opening values ​​are arranged continuously according to the prediction time domain index to obtain a set of candidate opening trajectory.

[0174] The candidate opening trajectory set is input into the probabilistic dynamics model, and the predicted states of flue gas concentration, temperature, wind pressure, opening degree and driving feedback are recursively deduced according to the prediction time domain index to obtain the candidate prediction state sequence.

[0175] In Example 1, the flue gas concentration, temperature, wind pressure, opening degree, and drive feedback of the current control cycle are determined as the input state of prediction time domain index 1; the opening degree value corresponding to index 1 of the candidate opening degree trajectory is input into the probabilistic dynamics model together with the input state to obtain the state change of each field; the state change of each field is superimposed on the value of the same field in the input state to obtain the predicted state of index 1; the predicted state of the previous index and the opening degree value corresponding to the next index are input into the probabilistic dynamics model together, and the process is recursively repeated until the rolling prediction step ends.

[0176] The uncertainty boundary is written into the candidate prediction state sequence in the order of the fields in the boundary comparison process. The positions of the fields that exceed the uncertainty boundary and the prediction time domain index are counted to obtain the boundary deviation record.

[0177] The dynamic penalty value is calculated based on the boundary deviation record and the policy penalty adjustment amount, and the dynamic penalty value is written into the trajectory cost calculation process of the candidate opening trajectory set to obtain the candidate trajectory cost sequence;

[0178] In Example 1, for each field position in the boundary deviation record, the excess magnitude of the corresponding predicted state value exceeding the uncertainty boundary is calculated; all excess magnitudes are accumulated in the order of the prediction time domain index to obtain the cumulative boundary excess value; the cumulative boundary excess value is multiplied by the policy penalty adjustment amount to obtain the dynamic penalty value; for each candidate opening trajectory, the absolute values ​​of the predicted values ​​of flue gas concentration, temperature, and wind pressure in the candidate predicted state sequence are accumulated item by item to obtain the environmental state cost; the absolute values ​​of the opening difference between adjacent prediction time domain indices are accumulated item by item to obtain the opening action cost; the environmental state cost, the opening action cost, and the dynamic penalty value are added together to obtain the trajectory cost of the corresponding candidate opening trajectory.

[0179] The candidate opening trajectory with the lowest cost is selected from the candidate trajectory cost sequence, and the opening value corresponding to the selected candidate opening trajectory in the first control cycle is determined as the driving opening control quantity.

[0180] Extract the driving feedback prediction state corresponding to the selected candidate opening trajectories, and determine the driving feedback prediction state as the prediction driving feedback.

[0181] In Example 1, within the candidate prediction state sequence obtained by recursion from the probabilistic dynamics model, each prediction time-domain index contains five fields: smoke concentration, temperature, wind pressure, opening degree, and driving feedback. The driving feedback prediction state is the value of the driving feedback field in the prediction state corresponding to the selected candidate opening trajectory in the first control cycle. When there are multiple target fire smoke exhaust skylight sampling times in the first control cycle, the driving feedback field value in the last prediction state is extracted as the driving feedback prediction state according to the sampling time sequence.

[0182] In this embodiment, the process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle is as follows:

[0183] According to the sampling time of the current control cycle, establish a control issuance record between the drive opening control quantity and the skylight identification of the target fire smoke exhaust skylight, and transmit the control issuance record to the target skylight drive mechanism;

[0184] The actual drive feedback returned by the target sunroof drive mechanism within the corresponding control cycle of the control issuance record is collected, and feedback alignment records are established between the actual drive feedback and the predicted drive feedback according to the sampling time.

[0185] The absolute value and direction of the difference between the actual driving feedback and the predicted driving feedback are calculated based on the feedback alignment record to obtain the prediction deviation record;

[0186] The prediction deviation record is compared with the opening change constraint. A feedback lag flag is generated when the prediction deviation record exceeds the opening change constraint, and a feedback follow flag is generated when the prediction deviation record does not exceed the opening change constraint.

[0187] In Example 1, the sampling time, absolute value of the difference, and direction of the difference are extracted from the prediction deviation record. The absolute value of the difference is compared with the opening change constraint within the same control cycle. When the absolute value of the difference is greater than the opening change constraint, the feedback status field is recorded as feedback lag, the lag flag value is recorded as 1, the follow flag value is recorded as 0, and the difference direction is retained as the lag direction record. When the absolute value of the difference is not greater than the opening change constraint, the feedback status field is recorded as feedback follow, the follow flag value is recorded as 1, and the lag flag value is recorded as 0. The feedback status field, lag flag value, follow flag value, sampling time, and difference direction are all written into the feedback flag record.

[0188] The drive feedback field corresponding to the target fire smoke exhaust skylight in the next control cycle is corrected based on the feedback lag mark and feedback follow mark, and the prediction deviation record is written into the delay linkage record at the corresponding sampling time.

[0189] The corrected drive feedback field, prediction deviation record, and delay linkage record are recombined with the flue gas concentration, temperature, wind pressure, opening degree, and topology linkage state parameters collected in the next control cycle to obtain the spatiotemporal coupling feature matrix construction input for the next control cycle.

[0190] In this embodiment, a fire smoke exhaust skylight drive control system based on artificial intelligence includes:

[0191] The spatiotemporal coupling matrix construction module is used to generate a spatiotemporal coupling feature matrix based on the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights.

[0192] The periodic truncation cache module is used to truncate the spatiotemporal coupled feature matrix according to a preset control period, generate the feature matrix of the previous period, the feature matrix of the current period and the adjacent correspondence, and write them into the training sample cache.

[0193] The drift detection module is used to input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to generate the flue gas distribution drift intensity and drift rise marker.

[0194] The dual-mode parameter generation module is used to generate a dual-mode parameter set containing model update parameters and strategy search parameters based on the flue gas distribution drift intensity and drift rise marker.

[0195] The probabilistic dynamics training module is used to write the model update parameters and flue gas distribution drift intensity into the training process of the improved PILCO probabilistic dynamics model, perform weighted updates on the training sample cache, and generate the probabilistic dynamics model and uncertainty boundary.

[0196] The strategy search decision module is used to write the strategy search parameters and uncertainty boundary into the strategy search process of the probabilistic dynamics model, and generate the driving opening control quantity and predict the driving feedback.

[0197] The control feedback write-back module is used to transmit the drive opening control quantity to the target sunroof drive mechanism, generate a prediction deviation based on the actual drive feedback and the predicted drive feedback, and write the actual drive feedback and the prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

[0198] Example 2:

[0199] To verify the feasibility of this invention in practice, it was applied to a scenario of coordinated control of fire-fighting smoke exhaust skylights at the top of a building atrium. The application targets were a single target fire-fighting smoke exhaust skylight and adjacent skylights with which it had a topological linkage. The scenario involved controlling the opening degree of the skylight when smoke accumulated in the upper part of the atrium during the initial stage of a fire and was affected by wind pressure disturbances. In this scenario, existing control methods often issue fixed opening degree commands based on smoke concentration, temperature, or fire-fighting linkage signals, failing to continuously utilize the topological linkage status of adjacent skylights, smoke distribution drift, and drive feedback deviations. When local wind pressure changes cause rapid changes in smoke distribution, the opening degree control tends to lag, and the prediction model struggles to correct it in a timely manner. This embodiment addresses these problems by forming a continuous processing chain encompassing spatiotemporal coupling feature matrix construction, RuLSIF algorithm drift detection, dual-mode parameter generation, improved PILCO probabilistic dynamics model training, strategy search decision-making, and control feedback write-back.

[0200] Simultaneously collect smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time, skylight identification, and topology linkage status parameters on the target fire exhaust skylight and adjacent skylights. The sampling period is set to 1 second, and the preset control period is set to 5 seconds. Each control period extracts matrix rows corresponding to 5 consecutive sampling times. The system establishes candidate sampling records according to skylight identification. Using the sampling time of the target fire exhaust skylight as a benchmark, it performs proximity matching on the sampling times of adjacent skylights, with a preset proximity matching threshold of 0.5 seconds. Then, it determines the topology connection position based on the topology linkage status parameters, performs spatial sorting on the sampling record group at the same moment, and compares the direction of wind pressure and opening degree change to generate a topology linkage consistency result. When the topology linkage is consistent, the delayed linkage record is recorded as 0; when the topology linkage is inconsistent, the corresponding sampling value is retained and the delayed linkage record is recorded as 1. Subsequently, the system writes the smoke concentration, temperature, wind pressure, opening degree, drive feedback, and delayed linkage records into a matrix according to the sampling time order and the topology connection position order, forming a spatiotemporal coupling feature matrix.

[0201] During the periodic truncation and caching process, the system marks candidate period boundaries according to a preset control period and calculates the period connection interval between the end boundary of the previous control period and the start boundary of the current control period. When the period connection interval is inconsistent with the 1-second sampling interval, the system extracts the wind pressure change rate and temperature change rate of adjacent skylights within the corresponding time period, with a preset change rate threshold set to 0.2. When any change rate exceeds the threshold, the candidate period boundary is moved sequentially according to the sampling interval corresponding to the adjacent matrix rows. When the cumulative movement span of the candidate period boundary reaches 10 seconds, the boundary movement stops, and the feature matrix corresponding to the adjacent skylight with the highest topological connection position is truncated as the replacement feature matrix. The system obtains the feature matrix of the previous period and the feature matrix of the current period based on the corrected period boundary sequence, establishes an adjacent correspondence according to the matrix row order, and writes it into the training sample cache.

[0202] During drift detection, the system extracts the matrix rows corresponding to the sampling time from the feature matrix of the previous period and the feature matrix of the current period, aligns them according to the topological connection position order to form a reference sample sequence and a test sample sequence, and establishes a linkage validity marker based on the delay linkage record. When the delay linkage record is 0, the linkage validity marker of the corresponding field is recorded as 1; when the delay linkage record is 1, the linkage validity marker of the corresponding field is recorded as 0. The system assigns the sample corresponding to the target fire smoke exhaust skylight to the target layer, and assigns the sample corresponding to the adjacent skylight to the adjacent layer according to the topological connection position, forming a layered sample set. The RuLSIF algorithm uses a Gaussian kernel function, with the linkage validity marker recording a weight of 1 when it is 1 and 0.2 when it is 0. After obtaining the layered relative density ratio sequence, the system accumulates the deviation amplitude of the target layer and the adjacent layer based on a preset baseline density ratio of 1 to obtain the smoke distribution drift intensity, and generates a drift increase marker when the drift intensity increases and the target layer and the adjacent layer deviate in the same direction.

[0203] During the dual-mode parameter generation process, the preset drift interval sequence is set as follows: low drift interval (0 to 0.5), medium drift interval (greater than 0.5 to 1.5), and high drift interval (greater than 1.5). The initial values ​​of the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment are set to 1.00, 0.05, and 0.10, respectively; the initial values ​​of the policy penalty adjustment, opening change constraint, and rolling prediction step size are set to 1.00, 5 degrees, and 5 control cycles, respectively. When the drift increase flag is established and the drift level is medium, the model update parameters are adjusted to 1.30, 0.08, and 0.16, and the policy search parameters are adjusted to 1.30, 8 degrees, and 3 control cycles; when the drift level is high, the parameters are adjusted to 1.60, 0.10, and 0.20, and 1.60, 10 degrees, and 2 control cycles, respectively. The system extracts periodically paired samples from the training sample cache to form state input features and state transition output features. It then divides the target layer input features and adjacent layer input features according to topological connection positions, resulting in a hierarchical state transition sample set. The system combines the flue gas distribution drift intensity and the training sample weight adjustment to form a weight labeling result, calculates the spatiotemporal coupling kernel response matrix, and trains an improved PILCO probabilistic dynamics model. Model training uses 30 consecutive control cycle samples as the initial cache; subsequent control cycles perform parameter fusion once and generate the uncertainty boundary for the current control cycle.

[0204] During the strategy search and decision-making process, the target fire exhaust skylight's mechanical fully closed opening is set to 0 degrees, and its mechanical fully open opening is set to 90 degrees. The system establishes a set of candidate opening trajectories based on the current opening and opening change constraints, and performs boundary correction on candidate values ​​exceeding the range of 0 to 90 degrees. Subsequently, it recursively calculates the predicted states of smoke concentration, temperature, wind pressure, opening, and drive feedback according to the rolling prediction step size, writes the uncertainty boundary into the field comparison process, statistically records boundary deviations, calculates the dynamic penalty value and candidate trajectory cost sequence, and selects the candidate opening trajectory with the minimum cost, outputting the drive opening control quantity and predicted drive feedback. The system sends the drive opening control quantity to the target skylight drive mechanism, collects actual drive feedback, calculates prediction deviation records, and generates feedback lag markers or feedback following markers based on the opening change constraints. The prediction deviation records, corrected drive feedback fields, and delayed linkage records are jointly written back to the spatiotemporal coupling feature matrix of the next control cycle to construct the input.

[0205] Table 1 Implementation Data of Fire Exhaust Skylight Drive Control

[0206] Control cycle Smoke distribution drift intensity Drift Ascending Marker Uncertainty boundary half-width Drive opening control quantity / degree Predictive-driven feedback / degree Actual driving feedback / degree Prediction bias / degree Fixed rule deviation / degree C1 0.42 Not established 0.10 45 44.6 43.8 0.8 4.7 C2 0.76 Established 0.19 53 52.4 51.2 1.2 5.6 C3 1.18 Established 0.23 61 60.5 59.1 1.4 6.3 C4 1.72 Established 0.36 71 70.2 68.4 1.8 7.1 C5 1.31 Not established 0.25 68 67.3 66.0 1.3 5.9

[0207] As shown in Table 1, as the smoke distribution drift intensity increases from 0.42 to 1.72, the system can synchronously adjust the uncertainty boundary and drive opening control quantity according to the drift increase mark, so that the target fire smoke exhaust skylight opening increases from 45 degrees to 71 degrees; in C5, the smoke distribution drift intensity drops from the high drift level to the medium drift level, the drift increase mark is not established, the system maintains the corresponding parameters and adjusts the drive opening control quantity to 68 degrees.

[0208] The prediction deviation between the predicted driving feedback and the actual driving feedback is in the range of 0.8 degrees to 1.8 degrees, which is lower than the opening change constraint of the corresponding control cycle. Compared with the deviation value under the fixed rule, the prediction deviation remains in a low range, indicating that the improved PILCO probabilistic dynamics model can constrain the candidate opening trajectory search and form the data correction input for the next control cycle through feedback comparison.

[0209] The results of this embodiment show that the present invention can uniformly express the state relationship between the target fire smoke exhaust skylight and the adjacent skylight using a spatiotemporal coupled feature matrix, identify the smoke distribution drift using the RuLSIF algorithm and generate a dual-mode parameter set, and then complete the probabilistic dynamics training, opening trajectory search and feedback closed-loop update by improving the PILCO probabilistic dynamics model, so that the drive opening control quantity is adaptively generated with the change of smoke distribution, and the actual drive feedback and prediction deviation are continuously written back to the next control cycle.

[0210] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fire-fighting smoke exhaust skylight drive control method based on artificial intelligence, characterized in that, include: Collect smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights, and construct a spatiotemporal coupling feature matrix; The spatiotemporal coupling feature matrix is ​​extracted according to the preset control period to obtain the feature matrix of the previous period and the feature matrix of the current period. After establishing the adjacent correspondence according to the sampling time, it is written into the training sample cache. Input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to calculate the flue gas distribution drift intensity and generate drift rise markers; A dual-mode parameter set containing model update parameters and policy search parameters is generated based on the flue gas distribution drift intensity and drift rise marker. The model update parameters and flue gas distribution drift intensity from the dual-mode parameter set are written into the training process of the improved PILCO probabilistic dynamics model. Weighted updates are performed on the training sample cache to obtain the probabilistic dynamics model and uncertainty boundary. The strategy search parameters and uncertainty boundary in the dual-mode parameter set are written into the strategy search process of the probabilistic dynamics model. The uncertainty boundary is used as a dynamic penalty term and constraint boundary. The probabilistic dynamics model is combined to generate the driving opening control quantity and predict the driving feedback. The process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

2. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The construction of the spatiotemporal coupled feature matrix is ​​specifically as follows: Collect the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage status parameters of the target fire smoke exhaust skylight and adjacent skylights, and establish candidate sampling records according to the skylight identification; Based on the sampling time of the target fire smoke exhaust skylight, neighbor matching is performed on the sampling times of adjacent skylights in the candidate sampling records to obtain the same sampling record group; The topological connection position between the target fire smoke exhaust skylight and the adjacent skylight is determined based on the topological linkage state parameters, and the sampling record group at the same moment is spatially sorted to obtain the spatially ordered sampling record group. By comparing the direction of wind pressure and opening change in the spatially ordered sampling record group, a topological linkage consistency result is constructed. Correction is performed on the spatially ordered sampling record group based on the topological linkage consistency results. When the topological linkage consistency results are consistent, the corresponding adjacent skylight sampling records are retained. When the topological linkage consistency results are inconsistent, the corresponding adjacent skylight sampling records are marked as delayed linkage records, thus obtaining the corrected sampling record group. According to the sampling time sequence and topological connection position sequence, the flue gas concentration, temperature, wind pressure, opening degree, drive feedback and delay linkage records in the correction sampling record group are recombined to obtain the spatiotemporal coupling feature matrix.

3. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The step of extracting the spatiotemporal coupling feature matrix according to a preset control period is specifically as follows: Extract the matrix rows from the spatiotemporal coupling feature matrix according to the sampling time sequence, and mark the candidate period boundaries according to the preset control period to obtain the candidate period boundary sequence; Extract matrix rows from two adjacent control periods based on the candidate period boundary sequence, and calculate the period connection interval; When the period connection interval is inconsistent with the sampling interval corresponding to the preset control cycle, the wind pressure change rate and temperature change rate of adjacent skylights in the corresponding time period are extracted and compared with the preset change rate threshold. The candidate period boundary is moved according to the comparison result until the period connection interval is consistent with the sampling interval corresponding to the preset control cycle, and the corrected period boundary sequence is obtained. When the movement span of the candidate cycle boundary exceeds the preset fire evolution safety tolerance time threshold, the movement of the candidate cycle boundary is stopped, and the feature matrix corresponding to the adjacent skylight with the highest topological connection position is extracted according to the topological linkage state parameters to obtain the alternative feature matrix. Based on the correction cycle boundary sequence, matrix rows within two adjacent control cycles are extracted from the spatiotemporal coupling feature matrix to obtain the feature matrix of the previous cycle and the feature matrix of the current cycle. When stopping the candidate period boundary movement and obtaining the alternative feature matrix, the alternative feature matrix is ​​written into the corresponding missing position in the current period feature matrix according to the sampling time; Establish the adjacent correspondence between the feature matrix of the previous period and the feature matrix of the current period according to the sampling time, and write it into the training sample cache.

4. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The specific steps for inputting the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm are as follows: Based on the adjacent correspondence, the matrix rows corresponding to the sampling time are extracted from the feature matrix of the previous period and the feature matrix of the current period, and alignment is performed according to the topological connection position order to obtain the reference sample sequence and the sample sequence to be tested; Delayed linkage records are extracted from the reference sample sequence and the sample sequence to be tested, and linkage effective labels are established for flue gas concentration, temperature, wind pressure, opening degree and drive feedback under the same topological connection position to obtain the labeled sample sequence. The labeled sample sequence is hierarchically aggregated according to the topological connection position. The sample corresponding to the target fire smoke exhaust skylight is assigned to the target layer, and the sample corresponding to the adjacent skylight is assigned to the adjacent layer according to the topological connection position, thus obtaining a hierarchical sample set. The hierarchical sample set is input into the RuLSIF algorithm, and the participation weight of the corresponding sample in the relative density ratio estimation is adjusted according to the effective linkage label to obtain the hierarchical relative density ratio sequence. Based on the deviation of each relative density ratio in the layered relative density ratio sequence from the preset reference density ratio, accumulation is performed according to the arrangement order of the target layer and adjacent layers to obtain the flue gas distribution drift intensity; The flue gas distribution drift intensity is compared with the flue gas distribution drift intensity corresponding to the previous control cycle. When the flue gas distribution drift intensity increases and the deviation direction of the target layer and the adjacent layer is consistent, a drift increase mark is generated.

5. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The generation of the dual-mode parameter set, which includes model update parameters and policy search parameters, specifically involves: By comparing the flue gas distribution drift intensity corresponding to the current control cycle with the flue gas distribution drift intensity corresponding to the previous control cycle, the relationship of drift intensity change can be obtained. The drift intensity of the flue gas distribution corresponding to the current control cycle is matched with the preset drift interval sequence to obtain the drift intensity level. The drift intensity level and the relationship between the drift intensity change are marked accordingly to obtain the drift state marking result. When the drift rise marker is established, the training sample weight adjustment amount, the model update step size adjustment amount, and the uncertainty boundary adjustment amount are increased according to the drift state marker results to obtain the model update parameters. When the drift rise marker is not established, the model update parameters are obtained by maintaining the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment according to the drift state marker results. When the drift rise flag is established, the policy penalty adjustment amount and the opening change constraint amount are increased according to the drift state flag result, and the rolling prediction step size is shortened to obtain the policy search parameters. When the drift rise marker is not established, the policy search parameters are obtained by maintaining the policy penalty adjustment amount, the opening change constraint amount, and the rolling prediction step size according to the drift state marker results. Establish a correspondence between the model update parameters and the policy search parameters according to the current control cycle, and generate a dual-mode parameter set.

6. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The obtained probabilistic dynamic model and uncertainty boundary are specifically as follows: Extract the previous period feature matrix, the current period feature matrix and adjacent correspondence from the training sample cache, and extract the model update parameters from the dual-mode parameter set; Based on the adjacent correspondence, the matrix rows corresponding to the sampling time in the previous period feature matrix and the current period feature matrix are paired to obtain periodic paired samples. According to the topological connection position, the target fire smoke exhaust skylight corresponding features and adjacent skylight corresponding features are extracted from the periodic paired samples to obtain spatiotemporal paired training samples. Extract the flue gas concentration, temperature, wind pressure, opening degree and driving feedback corresponding to the previous period from the spatiotemporal paired training samples to form state input features, and extract the change of the corresponding fields in the current period relative to the previous period to form state transition output features. Based on the topological connection position, the state input features are divided into target layer input features and adjacent layer input features, and a correspondence is established with the state transition output features to obtain a hierarchical state transition sample set. Extract the training sample weight adjustment, model update step size adjustment, and uncertainty boundary adjustment from the model update parameters; Based on the flue gas distribution drift intensity and the training sample weight adjustment, weight labeling is performed on the stratified state transition sample set to obtain the drift-weighted training sample set and weight labeling results. According to the topological connection positions of the target layer and adjacent layers, kernel response calculation is performed on the state input features in the drift-weighted training sample set, and the kernel response calculation results are weighted according to the weight labeling results to obtain the spatiotemporal coupled kernel response matrix; The PILCO probabilistic dynamics model is trained and improved based on the spatiotemporal coupling kernel response matrix and state transition output features to obtain candidate probabilistic dynamics model parameters and candidate prediction variances. Based on the model update step size adjustment, the parameters of the candidate probabilistic dynamics model and the probabilistic dynamics model parameters of the previous control cycle are fused to obtain the probabilistic dynamics model of the current control cycle. The uncertainty boundary for the current control cycle is generated based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment, and a correspondence is established between the probabilistic dynamics model and the uncertainty boundary for the current control cycle.

7. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 6, characterized in that, The improved PILCO probabilistic dynamics model includes a spatiotemporal feature input layer, a topological hierarchical indexing layer, an initial kernel response calculation layer, a drift weight injection layer, a state transition regression layer, a variance boundary generation layer, and a parameter adaptive fusion output layer. The spatiotemporal feature input layer receives target layer input features, adjacent layer input features, and state transition output features from the hierarchical state transition sample set, and establishes state transition training input according to sampling time and field order; The topology hierarchical index layer divides the state transition training input into target layer index records and adjacent layer index records according to the topology connection position. The initial kernel response calculation layer performs kernel response calculation on the target layer input features and the adjacent layer input features based on the target layer index record and the adjacent layer index record to generate an initial kernel response matrix; The drift weight injection layer receives the initial kernel response matrix and the weight labeling result, and maps the weight labeling result as an environmental attenuation penalty term to the matrix elements in the initial kernel response matrix corresponding to the sampling time and topological connection position, to obtain the spatiotemporal coupled kernel response matrix. The state transition regression layer receives the spatiotemporal coupling kernel response matrix and state transition output features, and generates candidate probabilistic dynamic model parameters and candidate prediction variances in the order of flue gas concentration, temperature, wind pressure, opening degree and drive feedback. The variance boundary generation layer generates the uncertainty boundary for the current control cycle based on the candidate prediction variance, flue gas distribution drift intensity, and uncertainty boundary adjustment amount, and then passes it to the parameter adaptive fusion output layer. The parameter adaptive fusion output layer receives candidate probabilistic dynamics model parameters, probabilistic dynamics model parameters from the previous control cycle, model update step size adjustment, and uncertainty boundary. It adjusts the allocation ratio of the model update step size adjustment in the parameter fusion process based on the magnitude change of the uncertainty boundary, and outputs the probabilistic dynamics model and uncertainty boundary for the current control cycle.

8. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The process of combining the probabilistic dynamics model to generate the driving opening control quantity and predict the driving feedback specifically involves: Extract the policy penalty adjustment, opening change constraint, and rolling prediction step size from the policy search parameters in the dual-mode parameter set, and call the probabilistic dynamics model and uncertainty boundary of the current control cycle; A candidate opening trajectory set is established based on the current opening degree and opening degree change constraint of the target fire smoke exhaust skylight, and a prediction time domain index is established for the candidate opening trajectory set according to the rolling prediction step size; The candidate opening trajectory set is input into the probabilistic dynamics model, and the predicted states of flue gas concentration, temperature, wind pressure, opening degree and driving feedback are recursively deduced according to the prediction time domain index to obtain the candidate prediction state sequence. The uncertainty boundary is written into the candidate prediction state sequence in the order of the fields in the boundary comparison process. The positions of the fields that exceed the uncertainty boundary and the prediction time domain index are counted to obtain the boundary deviation record. The dynamic penalty value is calculated based on the boundary deviation record and the policy penalty adjustment amount, and the dynamic penalty value is written into the trajectory cost calculation process of the candidate opening trajectory set to obtain the candidate trajectory cost sequence; The candidate opening trajectory with the lowest cost is selected from the candidate trajectory cost sequence, and the opening value corresponding to the selected candidate opening trajectory in the first control cycle is determined as the driving opening control quantity. Extract the driving feedback prediction state corresponding to the selected candidate opening trajectories, and determine the driving feedback prediction state as the prediction driving feedback.

9. The fire smoke exhaust skylight drive control method based on artificial intelligence according to claim 1, characterized in that, The process of sending the drive opening control quantity to the target sunroof drive mechanism, comparing the actual drive feedback with the predicted drive feedback to obtain the prediction deviation, and writing the actual drive feedback and prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle is as follows: According to the sampling time of the current control cycle, establish a control issuance record between the drive opening control quantity and the skylight identification of the target fire smoke exhaust skylight, and transmit the control issuance record to the target skylight drive mechanism; The actual drive feedback returned by the target sunroof drive mechanism within the corresponding control cycle of the control issuance record is collected, and feedback alignment records are established between the actual drive feedback and the predicted drive feedback according to the sampling time. The absolute value and direction of the difference between the actual driving feedback and the predicted driving feedback are calculated based on the feedback alignment record to obtain the prediction deviation record; The prediction deviation record is compared with the opening change constraint. A feedback lag flag is generated when the prediction deviation record exceeds the opening change constraint, and a feedback follow flag is generated when the prediction deviation record does not exceed the opening change constraint. The drive feedback field corresponding to the target fire smoke exhaust skylight in the next control cycle is corrected based on the feedback lag mark and feedback follow mark, and the prediction deviation record is written into the delay linkage record at the corresponding sampling time. The corrected drive feedback field, prediction deviation record, and delay linkage record are recombined with the flue gas concentration, temperature, wind pressure, opening degree, and topology linkage state parameters collected in the next control cycle to obtain the spatiotemporal coupling feature matrix construction input for the next control cycle.

10. A fire smoke exhaust skylight drive control system based on artificial intelligence, comprising the fire smoke exhaust skylight drive control method based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, include: The spatiotemporal coupling matrix construction module is used to generate a spatiotemporal coupling feature matrix based on the skylight identification, smoke concentration, temperature, wind pressure, opening degree, drive feedback, sampling time and topology linkage state parameters of the target fire smoke exhaust skylight and adjacent skylights. The periodic truncation cache module is used to truncate the spatiotemporal coupled feature matrix according to a preset control period, generate the feature matrix of the previous period, the feature matrix of the current period and the adjacent correspondence, and write them into the training sample cache; The drift detection module is used to input the feature matrix of the previous period and the feature matrix of the current period into the RuLSIF algorithm to generate the flue gas distribution drift intensity and drift rise marker. The dual-mode parameter generation module is used to generate a dual-mode parameter set containing model update parameters and strategy search parameters based on the flue gas distribution drift intensity and drift rise marker. The probabilistic dynamics training module is used to write the model update parameters and flue gas distribution drift intensity into the training process of the improved PILCO probabilistic dynamics model, perform weighted updates on the training sample cache, and generate the probabilistic dynamics model and uncertainty boundary. The strategy search decision module is used to write the strategy search parameters and uncertainty boundary into the strategy search process of the probabilistic dynamics model, and generate the driving opening control quantity and predict the driving feedback. The control feedback write-back module is used to transmit the drive opening control quantity to the target sunroof drive mechanism, generate a prediction deviation based on the actual drive feedback and the predicted drive feedback, and write the actual drive feedback and the prediction deviation back to the spatiotemporal coupling feature matrix construction process of the next control cycle.

Images