Intelligent building state modeling method and system using deep learning

Abstract

The application provides an intelligent building state modeling method and system applying deep learning, acquires distributed state sensing flow of the building, performs cross-domain feature correlation modeling on the distributed state sensing flow, generates building state correlation feature set with dynamic correlation attributes through inter-domain feature interaction strength calculation, calls a pre-trained deep state evolution game network to perform state trend deduction on the building state correlation feature set, generates a multi-path state evolution graph and a state resilience index of the building, performs building state resilience evaluation based on the multi-path state evolution graph and the state resilience index, locates a state fragile area and a fragility evolution rate of the building, and generates a building state self-healing regulation strategy according to the state fragile area and the fragility evolution rate. The application can combine the dynamic change trend of the fragility to perform prospective regulation, improve the initiative and effectiveness of the building state management, and thus optimize the accuracy and comprehensiveness of the building state modeling as a whole.

Description

Technical Field

[0001] This invention relates to the field of building monitoring, and more specifically, to an intelligent building status modeling method and system that applies deep learning. Background Technology

[0002] With the development of intelligent building technology, building state modeling technology has been proposed. By sensing, analyzing, and modeling various state data during building operation, it enables dynamic assessment and control of building state, which is beneficial to improving building management efficiency and ensuring operational stability. Existing technologies often fail to systematically partition and dynamically mine multi-source state data during the data processing stage, making it difficult to effectively capture the inherent interaction relationships between different types of data and affecting the comprehensiveness of feature representation. In the state prediction stage, the dynamic game relationship between features of different domains is not considered, resulting in a lack of diversity and robustness in the characterization of state evolution trends. In the state assessment process, it is often limited to static judgments of whether the state parameters exceed the threshold at the current moment, without quantitative analysis of the vulnerability of the building state and its evolution rate over time, leading to insufficient accuracy in identifying potential state risks. Summary of the Invention

[0003] This invention provides a method and system for intelligent building state modeling using deep learning.

[0004] In a first aspect, embodiments of the present invention provide a method for intelligent building state modeling using deep learning, comprising: acquiring a distributed state perception stream of a building, the distributed state perception stream including an environmental domain perception sequence, a device domain perception sequence, and a spatial domain perception sequence generated by different sensing nodes within a continuous monitoring period; performing cross-domain feature association modeling on the distributed state perception stream, generating a building state association feature set with dynamic association attributes by calculating the interaction strength between inter-domain features; calling a pre-trained deep state evolution game network to perform state trend inference on the building state association feature set, generating a multi-path state evolution map and a state resilience index of the building; performing a building state resilience assessment based on the multi-path state evolution map and the state resilience index, locating the building's state vulnerable areas and vulnerability evolution rate; and generating a building state self-healing regulation strategy based on the state vulnerable areas and the vulnerability evolution rate.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the intelligent building state modeling method using deep learning as described above.

[0006] This invention provides a deep learning-based building state modeling method that acquires a distributed state perception stream of a building. By integrating environmental, equipment, and spatial perception sequences generated by different sensing nodes within a continuous monitoring period, it constructs a multi-dimensional perception data foundation with spatiotemporal integrity, providing comprehensive and coherent input support for subsequent cross-domain feature association modeling. Through cross-domain feature association modeling of the distributed state perception stream, a dynamic association feature set is generated based on the inter-domain feature interaction strength calculation. This set includes environment-equipment interaction features, equipment-space coupling features, and space-environment feedback features, enabling deep capture of the dynamic interaction relationships between features from different domains. This overcomes the limitations of traditional static feature fusion and improves the ability of feature representation to depict complex building states. Finally, a pre-trained deep state evolution game network is invoked to process the building state association feature set. The system performs state trend extrapolation, generating multi-path state evolution maps and state resilience indices. By simulating the evolutionary game process between multi-domain features, it avoids the limitations of single-path prediction and quantifies the building's anti-interference capability, providing a more comprehensive basis for state assessment. Based on the multi-path state evolution maps and state resilience indices, it conducts building state resilience assessment, locating vulnerable areas and vulnerability evolution rates, achieving a leap from static anomaly judgment to dynamic evolution risk assessment, and improving the accuracy of identifying potential state problems. Based on vulnerable areas and vulnerability evolution rates, it generates building state self-healing control strategies, enabling proactive control based on the dynamic changes in vulnerability trends, improving the initiative and effectiveness of building state management, thereby optimizing the accuracy, comprehensiveness, and forward-looking nature of building state modeling and control. Attached Figure Description

[0007] Figure 1 This is a flowchart of an intelligent building state modeling method using deep learning, provided by an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 , Figure 1 A flowchart illustrating a method for intelligent building state modeling using deep learning, provided as an embodiment of the present invention, is available. This method can be executed by a computer system and may include the following steps:

[0010] Step S100: Obtain the distributed state perception stream of the building. The distributed state perception stream includes the environmental domain perception sequence, device domain perception sequence and spatial domain perception sequence generated by different sensing nodes within a continuous monitoring period.

[0011] Distributed state-aware streams are collections of data gathered by sensing nodes distributed throughout a building. These sensing nodes can be various sensors that continuously collect data over a monitoring cycle, forming different types of sensing sequences. Environmental domain sensing sequences describe the building's internal environmental conditions, such as changes in environmental parameters like temperature, humidity, and light intensity over time, reflecting the building's environmental state at different times. Equipment domain sensing sequences are data about the operational status of various devices within the building, covering information such as device on / off status, operating power, and rotational speed, demonstrating the device's working status and performance at different times. Spatial domain sensing sequences are data reflecting the usage and characteristics of space within the building's spatial area, such as personnel density and space occupancy rates in different areas, reflecting the utilization status of the building space at different times.

[0012] To acquire these sensing sequences, corresponding sensors can be deployed at different locations within the building. For environmental sensing sequences, temperature sensors can be used to collect real-time temperature data from different areas within the building, and humidity sensors can be used to acquire humidity data for those areas. For equipment sensing sequences, power sensors can be used to acquire equipment operating power data, and speed sensors can be used to acquire equipment speed data. For spatial sensing sequences, people counters can be used to count the number of people in different areas, and space occupancy sensors can be used to acquire space occupancy data. For example, in a commercial building, temperature and humidity sensors can be installed in each room on each floor, power and speed sensors can be installed on critical equipment, and people counters and space occupancy sensors can be installed in public areas to continuously acquire distributed state sensing streams.

[0013] Step S200: Perform cross-domain feature association modeling on the distributed state-aware flow, and generate a building state association feature set with dynamic association attributes by calculating the inter-domain feature interaction strength.

[0014] Cross-domain feature association modeling comprehensively considers the associations and interactions between perception sequences from different domains (environmental domain, device domain, and spatial domain) in a distributed state perception flow, extracting key features that reflect the overall building state. Calculating the inter-domain feature interaction strength is the process of evaluating the degree of mutual influence between features from different domains. By calculating this interaction strength, the association relationships between different features can be determined. A building state association feature set with dynamic association attributes is a set of features used to describe the building state; the association relationships between these features dynamically adjust with time and changes in the building state.

[0015] As one implementation method, step S200 can be specifically implemented as the following steps S210~S250:

[0016] Step S210: Timestamp synchronize the environment domain sensing sequence, device domain sensing sequence and spatial domain sensing sequence in the distributed state sensing stream to generate a synchronized sensing sequence with a unified time base.

[0017] Timestamp synchronization aligns sensing sequences from different domains in time, ensuring they share the same time base. This guarantees that features from different domains can be correlated along the same time dimension in subsequent analyses. Synchronized sensing sequences with a unified time base, after timestamp synchronization, maintain temporal consistency across environmental, device, and spatial domains, facilitating cross-domain feature correlation analysis. Timestamp synchronization can be achieved through a combination of hardware and software synchronization. Hardware synchronization uses a high-precision clock source, such as a GPS clock, to provide a unified time signal to all sensing nodes. Software synchronization calibrates the clocks of each sensing node using the Network Time Protocol (NTP).

[0018] Step S220: Calculate the interaction strength between the environment domain sensing sequence and the device domain sensing sequence in the synchronous sensing sequence, extract the time lag correlation parameters between environmental feature fluctuations and device feature responses, and construct the environment-device interaction feature matrix based on the time lag correlation parameters.

[0019] Environmental characteristic fluctuations refer to the changes in environmental parameters over time within the environmental domain sensing sequence, such as increases or decreases in temperature or changes in light intensity. Equipment characteristic responses are the responses of equipment parameters to environmental changes within the equipment domain sensing sequence, such as changes in equipment power or operational status. Time lag correlation parameters describe the time delay relationship between environmental characteristic fluctuations and equipment characteristic responses, reflecting the equipment's response time to environmental changes. The environment-equipment interaction characteristic matrix describes the interaction strength between environmental and equipment characteristics; the element values ​​of the matrix represent the interaction strength of the corresponding environmental and equipment characteristics.

[0020] As one implementation method, step S220 can be specifically implemented as the following steps S221~S225:

[0021] Step S221: Extract continuous feature segments from the synchronous sensing sequence of the environmental domain sensing sequence, calculate the environmental feature fluctuation amount of adjacent timestamps, and generate an environmental feature fluctuation sequence.

[0022] A continuous feature segment is a continuous piece of environmental domain sensing data selected from a synchronous sensing sequence. Environmental feature fluctuation is the difference between environmental feature values ​​at adjacent time points, reflecting the degree of change in environmental features over time. An environmental feature fluctuation sequence is a data sequence composed of a series of environmental feature fluctuations, describing the fluctuation of environmental features over time. In practice, continuous feature segments of the environmental domain sensing sequence can be extracted from the synchronous sensing sequence according to a preset time window. For example, setting the time window to 10 minutes, a segment of environmental domain sensing data is extracted every 10 minutes. Then, the difference between environmental feature values ​​at adjacent time points is calculated to obtain the environmental feature fluctuation. Assuming the environmental domain sensing sequence is temperature data, with a temperature of 25℃ at time t1 and 25.5℃ at time t2 (t2 = t1 + 1 minute), the environmental feature fluctuation between adjacent time points is 25.5 - 25 = 0.5℃. Recording the environmental feature fluctuations of all adjacent time points sequentially generates the environmental feature fluctuation sequence.

[0023] Step S222: Extract continuous feature segments from the device domain sensing sequence from the synchronous sensing sequence, calculate the device feature response quantities of adjacent timestamps, and generate the device feature response sequence.

[0024] Similar to step S221, this operation is performed on the device domain sensing sequence. A continuous feature segment is a continuous piece of device domain sensing data selected from the synchronous sensing sequence. The device feature response quantity is the difference between device feature values ​​at adjacent timestamps, reflecting the degree of change of the device feature at adjacent times. The device feature response sequence is a data sequence composed of a series of device feature response quantities, describing the response of the device feature over time.

[0025] Similarly, continuous feature segments of the device domain sensing sequence are extracted from the synchronous sensing sequence according to a preset time window. For example, for the power data of a certain device, a time window of 10 minutes is set, and a segment of device domain sensing data is extracted every 10 minutes. Then, the difference between the device feature values ​​of adjacent timestamps is calculated to obtain the device feature response quantity. Assuming that the power of the device is 1000 watts at a certain time t1 and 1050 watts at a certain time t2 (t2=t1+1 minute), then the device feature response quantity of the adjacent timestamps is 1050-1000=50 watts. By recording the device feature response quantities of all adjacent timestamps in sequence, the device feature response sequence is generated.

[0026] Step S223: Input the environmental characteristic fluctuation sequence and the equipment characteristic response sequence into the interaction analysis unit, calculate the interaction strength value under different lag durations using the sliding window method, and generate the interaction strength-lag duration relationship curve.

[0027] The interaction analysis unit is a module used to analyze the interaction relationship between environmental characteristic fluctuation sequences and equipment characteristic response sequences. The sliding window method is a commonly used method in time series analysis, which calculates the correlation between data within a fixed-length window by sliding it across the time series. The interaction strength value is an indicator that measures the degree of mutual influence between environmental characteristic fluctuations and equipment characteristic responses. The interaction strength-lag duration curve describes the change of the interaction strength value with lag duration, and it can visually demonstrate the time-lag correlation between environmental characteristic fluctuations and equipment characteristic responses.

[0028] In practical implementation, the interaction analysis unit can use correlation-based algorithms to calculate the interaction strength value. For example, the Pearson correlation coefficient can be used to measure the correlation between the environmental characteristic fluctuation sequence and the equipment characteristic response sequence. A fixed-length sliding window is set, for example, a window length of 30 minutes, and this window is simultaneously slid across the environmental characteristic fluctuation sequence and the equipment characteristic response sequence. For each lag duration (e.g., 0 minutes, 1 minute, 2 minutes, etc.), the Pearson correlation coefficient between the environmental characteristic fluctuation sequence and the equipment characteristic response sequence within the window is calculated, and this correlation coefficient is used as the interaction strength value. The interaction strength values ​​corresponding to different lag durations are recorded, and the interaction strength-lag duration relationship curve is plotted.

[0029] Step S224: Extract the peak value from the interaction strength-lag time relationship curve, filter the peak points where the interaction strength value exceeds the preset threshold, record the lag time parameter and interaction strength value corresponding to the peak point, and construct the environment-equipment interaction lag record table.

[0030] Peak extraction identifies all peak points in the interaction strength-lag duration relationship curve. The preset threshold is a pre-defined interaction strength value used to filter out peak points with strong interaction relationships. The environment-equipment interaction lag record table is used to record interaction lag information between environmental and equipment characteristics; it includes the lag duration parameter and interaction strength value corresponding to each peak point.

[0031] Peak value extraction can be performed using a derivative-based method. The derivative of the interaction strength-lag duration relationship curve is calculated; the point where the derivative changes from positive to negative is the peak value. Then, peak values ​​exceeding a preset threshold are selected. For example, if the preset threshold is 0.8, a peak value of 0.9 satisfies the selection criteria. The lag duration parameter (e.g., 5 minutes) and interaction strength value (0.9) corresponding to this peak value are recorded and stored in an environment-device interaction lag record table.

[0032] Step S225: Using environmental feature type as the row dimension and device feature type as the column dimension, the interaction strength value in the environment-device interaction delay record table is used as the matrix element value, and the delay duration parameter is used as the element attribute to construct the environment-device interaction feature matrix. The element value of the environment-device interaction feature matrix represents the interaction strength between the corresponding environmental feature and the device feature.

[0033] Environmental feature types are different environmental parameters in the environmental domain sensing sequence, such as temperature, humidity, and light intensity. Device feature types are different device parameters in the device domain sensing sequence, such as power, speed, and on / off status. The environment-device interaction feature matrix is ​​a two-dimensional matrix, where the row dimension represents the environmental feature type and the column dimension represents the device feature type. The element values ​​of the matrix are the interaction strength values ​​from the environment-device interaction time delay record table, and the element attributes are the corresponding lag duration parameters.

[0034] Step S230: Perform coupling path mining on the device domain sensing sequence and spatial domain sensing sequence in the synchronous sensing sequence, identify the co-occurrence pattern of device feature changes and spatial feature distribution, and generate device-space coupling feature vector by combining the co-occurrence pattern.

[0035] Coupling path mining analyzes the correlation paths between device domain sensing sequences and spatial domain sensing sequences to identify potential connections between changes in device characteristics and distributions of spatial characteristics. Device characteristic changes refer to the changes in device parameters over time within the device domain sensing sequence, such as device startup, shutdown, and power variations. Spatial characteristic distributions refer to the distribution of spatial parameters within the building space within the spatial domain sensing sequence, such as the distribution of personnel density and space occupancy. Co-occurrence patterns are patterns where changes in device characteristics and spatial characteristic distributions occur simultaneously, reflecting the interaction between device and space. The device-space coupling feature vector is a vector describing the coupling strength between device characteristics and spatial characteristics; the element values ​​of the vector represent the coupling strength between the corresponding device type and the spatial distribution pattern.

[0036] As one implementation method, step S230 can be specifically implemented as the following steps S231~S235:

[0037] Step S231: Perform feature mutation point detection on the device domain sensing sequence in the synchronous sensing sequence, identify the mutation time points of device feature values ​​from steady state to non-steady state or from non-steady state to steady state, and generate a set of device feature mutation events.

[0038] Feature mutation point detection identifies the points in the device domain sensing sequence where device feature values ​​suddenly change. Steady state is when device feature values ​​remain relatively stable over a period of time, while non-steady state is when device feature values ​​undergo significant changes. The device feature mutation event set is a collection containing all device feature mutation events, with each event recording the time of the mutation and the device feature values ​​before and after the mutation.

[0039] A threshold-based method can be used to detect feature mutation points. A threshold is set; when the change in a device's feature value exceeds this threshold, a feature mutation is considered to have occurred. All such mutation points are recorded to generate a set of device feature mutation events.

[0040] Step S232: Perform distribution pattern clustering on the spatial domain sensing sequence in the synchronous sensing sequence, divide the spatial feature distribution into different distribution pattern categories, record the start and end timestamps corresponding to each category, and generate a spatial distribution pattern time interval set.

[0041] Distribution pattern clustering classifies spatial feature distribution data in a spatial domain-aware sequence to identify similar distribution patterns. Distribution pattern categories are the groups of different spatial feature distribution patterns obtained through clustering. A spatial distribution pattern time interval set is a collection containing all spatial distribution pattern time intervals, with each interval recording the start and end timestamps of that distribution pattern. In practical implementation, density-based clustering algorithms, such as the DBSCAN algorithm, can be used to cluster spatial feature distribution data. Spatial feature distribution data is viewed as points in a high-dimensional space; by calculating the density and distance between points, similar points are grouped into the same category. For each clustered category, its corresponding start and end timestamps are recorded. For example, for the distribution data of personnel density within a building, the DBSCAN algorithm can be used to divide it into different distribution pattern categories, such as densely populated areas and sparsely populated areas. The start and end times of each category are recorded in time to generate a spatial distribution pattern time interval set.

[0042] Step S233: Align the set of device feature mutation events with the set of spatial distribution pattern time intervals on the time axis, calculate the probability that each device feature mutation time point falls into the spatial distribution pattern time interval, and generate a time coupling probability value.

[0043] Time axis alignment aligns the set of abrupt changes in equipment characteristics and the set of time intervals for spatial distribution patterns on the time axis, allowing for comparison of the relationship between the time points of abrupt changes in equipment characteristics and the time intervals of spatial distribution patterns. The temporal coupling probability value is the probability that the time point of abrupt changes in equipment characteristics falls within the time interval of spatial distribution patterns, reflecting the degree of temporal coupling between changes in equipment characteristics and spatial feature distribution.

[0044] In the specific calculation, for each mutation time point in the set of device characteristic mutation events, it is checked whether it falls within a certain interval of the spatial distribution pattern time interval set. The number of mutation time points falling within the interval is counted, and the result is divided by the total number of mutation time points to obtain the temporal coupling probability value.

[0045] Step S234: Filter highly coupled event pairs based on temporal coupling probability values, extract the co-occurrence relationship between device feature mutation type and spatial distribution pattern category in the highly coupled event pairs, and generate a device-space co-occurrence pattern table.

[0046] Highly coupled event pairs are combinations of device feature mutation events with temporal coupling probability values ​​exceeding a preset threshold and spatial distribution pattern time intervals. The preset threshold is a pre-defined temporal coupling probability value used to filter event pairs with strong temporal coupling relationships. Co-occurrence relationships refer to the simultaneous occurrence of device feature mutation types and spatial distribution pattern categories within a highly coupled event pair. The device-space co-occurrence pattern table is used to record the co-occurrence relationships between device feature mutation types and spatial distribution pattern categories.

[0047] Set a preset threshold, such as 0.7, to filter out highly coupled event pairs whose temporal coupling probability exceeds this threshold. For each highly coupled event pair, record the equipment feature mutation type (e.g., equipment startup, equipment shutdown) and spatial distribution pattern category (e.g., densely populated area, sparsely populated area). Count the occurrence frequency of different combinations of equipment feature mutation types and spatial distribution pattern categories to generate an equipment-space co-occurrence pattern table.

[0048] Step S235: Sort the co-occurrence frequency values ​​in the device-space co-occurrence pattern table according to the preset device type priority, and generate a device-space coupling feature vector with the same dimension as the number of device types. The element values ​​of the coupling feature vector represent the coupling strength between the corresponding device type and the spatial distribution pattern.

[0049] The co-occurrence frequency value is the frequency with which device feature mutation types and spatial distribution pattern categories co-occur in highly coupled event pairs. The preset device type priority is a pre-defined device type sorting rule used to rank the co-occurrence frequency values. The device-spatial coupling feature vector is a vector with the same dimension as the number of device types; each element of the vector corresponds to a device type, and the element value represents the coupling strength between that device type and the spatial distribution pattern.

[0050] In practice, the co-occurrence frequency values ​​in the device-space co-occurrence pattern table are sorted according to a preset device type priority. For example, the preset device type priority is Device A > Device B > Device C. For Device A, its co-occurrence frequency values ​​with various spatial distribution pattern categories are calculated, and these frequency values ​​are normalized to obtain the coupling strength between Device A and the spatial distribution pattern. Similarly, the coupling strength of Device B and Device C is calculated. These coupling strength values ​​are arranged sequentially to generate a device-space coupling feature vector. Assuming there are 3 device types, the device-space coupling feature vector is a three-dimensional vector, with each element corresponding to the coupling strength of a device type.

[0051] Step S240: Construct feedback loops for the spatial domain sensing sequence and the environmental domain sensing sequence in the synchronous sensing sequence, analyze the influence weight of the spatial feature gradient on the steady state of the environmental features, and generate a spatial-environmental feedback feature map based on the influence weight.

[0052] Feedback loop construction analyzes the interaction between spatial domain perception sequences and environmental domain perception sequences, establishing a feedback mechanism between spatial and environmental features. Spatial feature gradient is the rate of change of spatial features within a building space, such as the gradient of population density or spatial temperature. Environmental feature steady state is the relatively stable state of environmental features over a period of time. Influence weight is an indicator that measures the degree of influence of spatial feature gradient on environmental feature steady state. The spatial-environmental feedback feature map is a map used to describe the feedback relationship between spatial and environmental features, visually demonstrating the impact of spatial feature gradient on environmental feature steady state.

[0053] As one implementation method, step S240 can be specifically implemented as the following steps S241~S245:

[0054] Step S241: Extract the spatial feature gradient parameters of the spatial domain sensing sequence and the environmental feature steady-state parameters of the corresponding timestamp environmental domain sensing sequence from the synchronous sensing sequence to generate a spatial-environment correlation parameter pair sequence.

[0055] Spatial feature gradient parameters are specific numerical values ​​describing spatial feature gradients; for example, the spatial temperature gradient can be represented by dividing the temperature difference by the distance. Environmental feature steady-state parameters are parameters describing the steady state of environmental features; for example, the average value of ambient temperature, the stable value of humidity, etc. A spatial-environmental correlation parameter pair sequence is a sequence composed of spatial feature gradient parameters and environmental feature steady-state parameters. Each parameter pair records the spatial feature gradient and environmental feature steady-state parameter at the same timestamp.

[0056] In practice, for spatial feature data in the spatial domain sensing sequence, its spatial gradient is calculated. For example, for temperature data within a building, temperature values ​​are measured at two adjacent spatial locations, the temperature difference is calculated, and then divided by the distance between the two locations to obtain the temperature gradient. Simultaneously, steady-state environmental feature parameters corresponding to the timestamps are extracted from the environmental domain sensing sequence. The spatial feature gradient parameters and environmental feature steady-state parameters at each timestamp are combined into a parameter pair, and recorded sequentially to generate a sequence of spatial-environment correlation parameter pairs.

[0057] Step S242: The spatial-environmental correlation parameter pair sequence is segmented, and the sequence is divided into multiple parameter intervals according to the continuous change interval of the spatial feature gradient parameters. Each parameter interval corresponds to a set of environmental feature steady-state parameter change data.

[0058] Segmentation divides the spatial-environment correlation parameter sequence according to the changes in spatial feature gradient parameters, ensuring that the spatial feature gradient parameters within each interval exhibit continuous variation. Each parameter interval, obtained through segmentation, corresponds to a set of environmental feature steady-state parameter variation data. A threshold-based method can be used for segmentation. A threshold is set; when the change in spatial feature gradient parameters exceeds this threshold, a new parameter interval is considered to have been entered. The variation data of the environmental feature steady-state parameters corresponding to each interval are then recorded.

[0059] Step S243: Calculate the rate of change of the steady-state parameters of the environmental features within each parameter interval, and calculate the weighted sum of the rate of change and the length of the parameter interval to obtain the influence weight of the spatial feature gradient on the steady-state of the environmental features.

[0060] The rate of change of a steady-state environmental characteristic parameter is the rate of change of that parameter over time within a parameter interval. The length of the parameter interval is the time span of that interval. The influence weight is calculated by weighting the rate of change of the steady-state environmental characteristic parameter with the length of the parameter interval, reflecting the degree of influence of the spatial characteristic gradient on the steady-state environmental characteristic. In actual calculations, the rate of change of the steady-state environmental characteristic parameter is calculated for each parameter interval. For example, for the steady-state parameter of ambient temperature, the change in temperature within the interval is calculated and divided by the time length of the interval to obtain the rate of change in temperature. Then, this rate of change is weighted with the length of the parameter interval. Assuming the length of the parameter interval is t, the rate of change of the steady-state environmental characteristic parameter is r, and the weighting coefficient is set to α, then the influence weight is α × r × t.

[0061] Step S244: Construct a directed spatial-environment feedback graph using the spatial regions of the building as nodes and the influence weights as edge weights. The node attributes of the directed graph include the mean gradient of the regional spatial features, and the edge attributes include the influence weights and the feedback direction.

[0062] A spatial-environment feedback directed graph is a graph used to describe the feedback relationship between spatial features and environmental features. Nodes in the graph represent spatial regions of a building, and edges represent the feedback relationships between these regions. Node attributes are information related to the node, including the mean gradient of the spatial features in that region. Edge attributes are information related to the edge, including the influence weight and feedback direction. When constructing the directed graph, the spatial regions of the building are first identified as nodes. For each node, the mean gradient of the spatial features in that region is calculated. For example, for a floor's spatial region, the average spatial temperature gradient within that region is calculated. Then, the influence weights calculated in step S243 are used as the edge weights. The feedback direction can be determined based on the relationship between the spatial feature gradient and the steady-state parameters of the environmental features. For example, if an increase in the spatial feature gradient leads to an increase in the steady-state parameters of the environmental features, the feedback direction is positive; otherwise, it is negative. This information is recorded in the node and edge attributes of the directed graph.

[0063] Step S245: Extract graph features from the spatial-environment feedback directed graph, and fuse node attributes and edge attributes into a graph adjacency matrix and a node feature matrix to generate a spatial-environment feedback feature graph.

[0064] Graph feature extraction extracts information describing the structure and features of a directed graph with spatial-environment feedback. The graph adjacency matrix is ​​a matrix representing the connections between nodes in the graph, where each element represents the edge weight. The node feature matrix represents the features of each node in the graph, with each row corresponding to a node and each column corresponding to a node attribute. The spatial-environment feedback feature graph is obtained by combining the graph adjacency matrix and the node feature matrix, and it can comprehensively describe the feedback relationship between spatial and environmental features.

[0065] In practical implementation, for a directed graph with spatial-environment feedback, the first step is to construct a graph adjacency matrix. If two nodes are connected by an edge, the corresponding element in the matrix represents the edge weight; otherwise, the element's value is 0. Next, a node feature matrix is ​​constructed, arranging the attributes of each node (such as the mean gradient of the regional spatial features) column-wise. Combining the graph adjacency matrix and the node feature matrix generates the spatial-environment feedback feature graph.

[0066] Step S250: Input the environment-device interaction feature matrix, device-space coupling feature vector, and space-environment feedback feature map into the dynamic association fusion module. Dynamically allocate the association weights of different domain features through the attention mechanism. Perform feature aggregation processing based on the association weights to generate a building status association feature set with dynamic association attributes.

[0067] The dynamic association fusion module is used to fuse features from different domains. It can dynamically adjust the weights of features based on the association relationships between features from different domains. The attention mechanism can dynamically assign weights based on the importance of the input data. The association weights are weights assigned to features from different domains through the attention mechanism, reflecting the importance of each feature in the fusion process. Feature aggregation processing combines features from different domains according to their association weights to obtain a comprehensive feature representation. The building status association feature set with dynamic association attributes is a feature set obtained through feature aggregation processing. The association relationships between these features are dynamically adjusted with time and changes in building status.

[0068] As one implementation method, in step S250, the association weights of different domain features are dynamically allocated through an attention mechanism, which can be specifically implemented as follows: steps S251~S255:

[0069] Step S251: Extract the average historical interaction strength of the environment-device interaction feature matrix, the device-space coupling feature vector, and the space-environment feedback feature map from the historical feature fusion records of the dynamic association fusion module, and generate the historical interaction strength benchmark matrix.

[0070] Historical feature fusion records are all feature fusion information recorded by the dynamic correlation fusion module over a past period. The historical average interaction strength is the average interaction strength of the environment-device interaction feature matrix, device-space coupling feature vector, and space-environment feedback feature map over a preset period. The historical interaction strength baseline matrix is ​​a matrix representing the historical average interaction strength; each element represents the historical average interaction strength of its corresponding feature. In practice, a preset period is set, and data on the environment-device interaction feature matrix, device-space coupling feature vector, and space-environment feedback feature map within the preset period are extracted from the historical records of the dynamic correlation fusion module. For the environment-device interaction feature matrix, the average value of each element within the preset period is calculated to obtain the historical average interaction strength. These averages are then combined into a matrix to generate the historical interaction strength baseline matrix. Similarly, the average values ​​of the device-space coupling feature vector and the space-environment feedback feature map within the preset period are also calculated.

[0071] Step S252: Calculate the absolute value of the difference between each element in the current environment-device interaction feature matrix and the corresponding element in the historical interaction strength benchmark matrix to obtain the environment-device interaction deviation matrix. Use the same method to generate the device-space coupling deviation vector and the space-environment feedback deviation map respectively.

[0072] The environment-equipment interaction deviation matrix is ​​a matrix representing the deviation between the current environment-equipment interaction feature matrix and the historical interaction strength benchmark matrix. The element values ​​represent the absolute values ​​of the differences between corresponding elements. The equipment-space coupling deviation vector is a vector representing the deviation between the current equipment-space coupling feature vector and the historical mean. The element values ​​represent the absolute values ​​of the differences between corresponding elements. The space-environment feedback deviation map is a map representing the deviation between the current space-environment feedback feature map and the historical mean. The edge weights of the map represent the deviation values ​​of the corresponding edges. In actual calculations, for each element in the current environment-equipment interaction feature matrix, the value of the corresponding element in the historical interaction strength benchmark matrix is ​​subtracted, and the absolute value of the difference is taken to obtain the element values ​​of the environment-equipment interaction deviation matrix. The equipment-space coupling deviation vector and the space-environment feedback deviation map are calculated using the same method.

[0073] Step S253: Quantify the degree of deviation of the environment-equipment interaction deviation matrix, the equipment-space coupling deviation vector, and the space-environment feedback deviation map. Take the mean of the elements of the deviation matrix as the environment-equipment domain deviation, the magnitude of the deviation vector as the equipment-space domain deviation, and the sum of the edge weight changes of the deviation map as the space-environment domain deviation.

[0074] Deviation quantification involves quantifying the deviation information in the environment-device interaction deviation matrix, the device-space coupling deviation vector, and the space-environment feedback deviation map to obtain a numerical value representing the degree of deviation. The environment-device domain deviation is obtained by calculating the mean of the elements of the environment-device interaction deviation matrix, reflecting the degree of deviation between environment-device domain characteristics and historical mean. The device-space domain deviation is obtained by calculating the magnitude of the device-space coupling deviation vector, reflecting the degree of deviation between device-space domain characteristics and historical mean. The space-environment domain deviation is obtained by calculating the sum of the edge weight changes in the space-environment feedback deviation map, reflecting the degree of deviation between space-environment domain characteristics and historical mean.

[0075] In practical calculations, for the environment-device interaction deviation matrix, the average value of all elements is calculated, and this average value is taken as the environment-device domain deviation. For the device-space coupling deviation vector, the magnitude of the vector (i.e., the square root of the sum of the squares of the vector elements) is calculated, and this magnitude is taken as the device-space domain deviation. For the space-environment feedback deviation map, the sum of the weight changes of all edges is calculated, and this sum is taken as the space-environment domain deviation.

[0076] Step S254: Based on the environment-device domain deviation, device-space domain deviation, and space-environment domain deviation, set the initial weight allocation ratio of the attention mechanism. The domain feature with higher deviation is assigned a higher initial weight ratio, and a preliminary weight set is generated.

[0077] The initial weight allocation ratio is the initial weight ratio assigned to different domain features in the attention mechanism. The initial weight set is a set containing the initial weights of all domain features, where each weight represents the initial importance of the corresponding domain feature in the fusion process.

[0078] Based on the magnitudes of the environment-device domain bias, device-space domain bias, and space-environment domain bias, the initial weight allocation ratios for the attention mechanism are set. For example, assuming the environment-device domain bias is 0.3, the device-space domain bias is 0.5, and the space-environment domain bias is 0.2, then the highest initial weight percentage (e.g., 0.5) is assigned to the device-space domain features; an initial weight percentage of 0.3 is assigned to the environment-device domain features; and an initial weight percentage of 0.2 is assigned to the space-environment domain features. These weights are recorded to generate a preliminary weight set.

[0079] Step S255: Input the initial weight set into the weight balancing processing unit, and control the difference in the proportion of feature weights of each domain within a preset threshold range through dynamic adjustment, generating dynamic association weights with a sum of 1. The dynamic association weights are used for subsequent feature aggregation processing.

[0080] The weight balancing processing unit is a virtual software module that dynamically adjusts the weight proportions of features in each domain based on a preset threshold. The degree of difference is the extent of variation between the weight proportions of features in different domains. The preset threshold is a pre-defined threshold for the degree of difference used to control the variation in the weight proportions of features in different domains. The dynamically associated weights are weights obtained after adjustment by the weight balancing processing unit; the sum of these weights is 1, and they are used for subsequent feature aggregation processing.

[0081] In practice, the initial weight set is input into the weight balancing processing unit. This unit continuously adjusts the weight proportions of each domain feature to keep their differences within a preset threshold. For example, if the preset threshold is 0.1, and the weight proportions of device-space domain features are 0.5, environment-device domain features are 0.3, and space-environment domain features are 0.2, the difference is 0.3, exceeding the threshold. The weight balancing processing unit will then appropriately reduce the weight proportion of device-space domain features and increase the weight proportions of other domain features until the difference is controlled within 0.1. The final weight sum is 1; these weights are the dynamic association weights, used for subsequent feature aggregation processing.

[0082] Step S300: Call the pre-trained deep state evolution game network to perform state trend inference on the building state association feature set, and generate the multi-path state evolution map and state resilience index of the building.

[0083] A pre-trained deep state evolution game network is a deep learning network that can be used to analyze the evolutionary trends of building states. State trend inference predicts the possible future state changes of a building based on its associated feature set. A multi-path state evolution map is a graph that displays multiple possible evolutionary paths of a building state, visually representing how the building state changes under different conditions. The state resilience index measures a building's ability to withstand disturbances, reflecting its ability to maintain a stable state when faced with external interference.

[0084] As one implementation method, step S300 can be specifically implemented as the following steps S310~S350:

[0085] Step S310: Input the building state association feature set into the association feature encoding layer of the deep state evolution game network, and perform spatiotemporal association modeling on the building state association feature set through the spatiotemporal graph convolutional network to generate the spatiotemporal association feature tensor.

[0086] The association feature encoding layer is a layer in a deep state evolutionary game network used to encode the input set of building state association features, transforming the features into a form suitable for network processing. The spatiotemporal graph convolutional network is a network that combines graph convolutional networks and spatiotemporal analysis, enabling it to model the spatiotemporal information in the set of building state association features and capture the spatiotemporal relationships between features. The spatiotemporal association feature tensor is a tensor containing the spatiotemporal association information of the building state association feature set, which can be used as input to subsequent network layers.

[0087] As one implementation method, step S310 can be specifically implemented as the following steps S311~S315:

[0088] Step S311: Convert the environment-device interaction feature matrix, device-space coupling feature vector and space-environment feedback feature map in the building status association feature set into a feature tensor of a unified dimension, ensuring that the time dimension and spatial dimension of each feature tensor are aligned.

[0089] Unified-dimensional feature tensors convert the environment-device interaction feature matrix, device-space coupling feature vector, and space-environment feedback feature map into tensors with the same dimension, facilitating subsequent processing. Aligning the temporal and spatial dimensions ensures that each feature tensor has consistent dimensions in both time and space, enabling spatiotemporal correlation analysis. In practice, padding and truncation methods can be used to convert features of different dimensions into feature tensors of the same dimension. For example, for the environment-device interaction feature matrix, assuming its dimension is m×n, the device-space coupling feature vector has a dimension of p, and the space-environment feedback feature map has a dimension of q×r, a unified dimension is determined, such as M×N×T (M and N are spatial dimensions, and T is the temporal dimension). For features with dimensions smaller than the unified dimension, padding operations are performed, such as padding the matrix edges with 0s. For features with dimensions larger than the unified dimension, truncation operations are performed, retaining only the first M×N×T data. At the same time, ensure that the temporal and spatial dimensions of each feature tensor are aligned in time and space, for example, using the same timestamp in time and the same spatial coordinates in space.

[0090] Step S312: Construct a spatiotemporal relational graph structure for the building, using the spatial location of the sensing node as the graph node coordinates and the temporal and spatial correlations between nodes as edge weights, to generate a spatiotemporal relational adjacency matrix.

[0091] A spatiotemporal correlation graph is a graph used to represent the spatiotemporal relationships between sensing nodes in a building. Nodes in the graph represent sensing nodes, and edges represent the relationships between nodes. The spatial location of a sensing node is its specific coordinates within the building space. Temporal and spatial correlations between nodes are indicators of the degree of connection between them in time and space. A spatiotemporal correlation adjacency matrix is ​​a matrix used to represent the spatiotemporal correlation graph structure; the element values ​​of the matrix represent the edge weights between nodes.

[0092] In practical construction, the spatial locations of all sensing nodes in the building are first determined, and these locations are used as the coordinates of the graph nodes. Then, the temporal and spatial correlations between nodes are calculated. For example, the Pearson correlation coefficient is used to calculate the temporal correlation between nodes, and Euclidean distance is used to calculate the spatial correlation. The temporal and spatial correlations are then weighted and combined to obtain the edge weights. All edge weights between nodes are recorded in a matrix to generate a spatiotemporal adjacency matrix.

[0093] Step S313: Input the feature tensor and the spatiotemporal correlation adjacency matrix into the temporal convolutional layer of the spatiotemporal graph convolutional network, extract the temporal dimension features of the feature tensor through causal convolution kernels, preserve the causal relationship of the time series, and generate a temporal correlation feature map.

[0094] A temporal convolutional layer is a layer in a spatiotemporal graph convolutional network used to perform convolution operations on feature tensors along the time dimension. A causal convolutional kernel is a special type of convolutional kernel that considers only past information in the time series, preserving the causal relationships within the time series. A temporally correlated feature map is a feature map obtained after processing by a temporal convolutional layer, containing the temporal correlation information of the feature tensors.

[0095] In practical implementations, the temporal convolutional layers of spatiotemporal graph convolutional networks can employ one-dimensional convolution for temporal feature extraction. The feature tensor and the spatiotemporal correlation adjacency matrix are input into the temporal convolutional layer, and a causal convolution kernel is used to convolve the feature tensor. The size and stride of the causal convolution kernel can be adjusted according to specific circumstances. During convolution, the causal convolution kernel only considers information from the current time step and previous time steps, thus preserving the causal relationships of the time series. The result of the convolution is used as a temporal correlation feature map.

[0096] Step S314: Input the temporal correlation feature map into the spatial graph convolutional layer of the spatiotemporal graph convolutional network, and aggregate the spatial neighbor feature information of each node through graph convolution operation to generate a spatial correlation feature map.

[0097] Spatial graph convolutional layers are layers in spatiotemporal graph convolutional networks used to perform convolution operations on temporally correlated feature maps in the spatial dimension. Graph convolution is a type of convolution operation performed on a graph structure, which can aggregate the spatial neighbor feature information of each node. Spatial correlated feature maps are feature maps obtained after processing by spatial graph convolutional layers, containing the spatial correlation information of feature tensors.

[0098] In practice, spatial graph convolutional layers can employ graph neural network-based methods for graph convolution operations. The temporal correlation feature map and the spatiotemporal correlation adjacency matrix are input into the spatial graph convolutional layer. For each node, its spatial neighbors are determined based on the spatiotemporal correlation adjacency matrix. Then, the feature information of these neighboring nodes is aggregated through graph convolution operations. Through multiple iterations, the spatial neighbor feature information of each node is continuously aggregated, ultimately yielding the spatial correlation feature map.

[0099] Step S315: Concatenate the temporal correlation feature map and the spatial correlation feature map along the channel dimension, and fuse the temporal and spatial correlation information to generate a spatiotemporal correlation feature tensor with dual spatiotemporal correlation attributes.

[0100] Dimensional concatenation is an operation that combines temporally and spatially correlated feature maps along a specific dimension, allowing their information to be fused together. The channel dimension refers to the channel direction in the feature maps; by fusing along the channel dimension, temporal and spatial correlation information can be integrated into a single tensor. A spatiotemporally correlated feature tensor with dual spatiotemporal correlation attributes is obtained through dimensional concatenation and fusion, simultaneously containing the correlation information of the feature tensor in both time and space dimensions.

[0101] In practical implementation, the temporal correlation feature map and the spatial correlation feature map are concatenated along the channel dimension. Assume the temporal correlation feature map has a dimension of M×N×C1 (M and N are spatial dimensions, C1 is the channel dimension), and the spatial correlation feature map has a dimension of M×N×C2. After concatenating them along the channel dimension, the resulting spatiotemporal correlation feature tensor has a dimension of M×N×(C1+C2). In this way, the temporal and spatial correlation information is integrated, forming a spatiotemporal correlation feature tensor with dual spatiotemporal correlation attributes.

[0102] Step S320: Input the spatiotemporal correlation feature tensor into the multi-agent game layer of the deep state evolution game network, initialize virtual game agents in the environment domain, device domain, and spatial domain, simulate the evolution game process of different domain features through policy interaction between agents, and generate multi-path evolution feature sequences.

[0103] The multi-agent game layer is a layer in a deep state evolutionary game network used to simulate the evolutionary game process between features of different domains. Virtual game agents are virtual agents representing the environment domain, device domain, and spatial domain; they can make decisions and interact according to their own strategies. Virtual game agents are software models with autonomous decision-making and interaction capabilities, simulating the evolutionary game process of features of the environment domain, device domain, and spatial domain within the multi-agent game layer of the deep state evolutionary game network. A virtual game agent can consist of a state perception module, a policy decision module, and an action execution module. The state perception module receives the feature sub-tensor of the corresponding domain from the spatiotemporal correlation feature tensor, using it as the agent's initial observation state. The policy decision module selects appropriate actions from a preset policy space based on the observation state provided by the state perception module. For example, for an environment domain agent, the policy space contains a set of environmental feature adjustment actions, such as increasing temperature, decreasing temperature, increasing humidity, and decreasing humidity. The policy decision-making module can employ rule-based decision-making methods, such as choosing to raise the temperature when it is observed to be below a preset comfort range; or it can use machine learning-based methods, such as deep reinforcement learning algorithms. Taking a Deep Q-Network (DQN) as an example, this network takes the observed state as input and outputs the Q-value corresponding to each action, where the Q-value represents the expected reward obtained after taking the action. The policy decision-making module selects the action with the largest Q-value as the current decision. The action execution module is responsible for converting the action selected by the policy decision-making module into actual control commands and executing them in the virtual control environment. For example, if the policy decision-making module of the environmental agent selects the action of raising the temperature, the action execution module will send control commands to the simulated air conditioning equipment, simulating the air conditioning equipment to start heating. The action execution module will also record the environmental changes after the action is executed and feed the new observed state back to the state perception module, forming a closed-loop decision-making process. Policy interaction is the process by which agents interact according to their respective policies. The multi-path evolution feature sequence is a feature sequence simulated through policy interactions between agents, with each sequence corresponding to a possible state evolution path.

[0104] As one implementation method, step S320 can be specifically implemented as the following steps S321~S325:

[0105] Step S321: Initialize the environment domain agent, device domain agent, and spatial domain agent, and take the feature sub-tensor of the corresponding domain in the spatiotemporal correlation feature tensor as the initial observation state of the agent.

[0106] Environment-domain agents, device-domain agents, and space-domain agents are virtual agents representing the environment, device, and space domains, respectively, possessing the ability to make autonomous decisions and interact. The feature sub-tensors corresponding to the corresponding domains in the spatiotemporal correlation feature tensor are feature data related to the environment, device, and space domains extracted from the spatiotemporal correlation feature tensor. The initial observation state is the environmental state observed by the agent at the beginning, serving as the basis for the agent's decision-making.

[0107] In actual initialization, based on the dimension and domain division of the spatiotemporal correlation feature tensor, it is decomposed into feature sub-tensors of the environment domain, device domain, and spatial domain. For example, suppose the dimension of the spatiotemporal correlation feature tensor is M×N×C, where the first C1 channels correspond to the environment domain features, the middle C2 channels correspond to the device domain features, and the last C3 channels correspond to the spatial domain features (C1+C2+C3=C). The data from the first C1 channels are used as the initial observation state of the environment domain agent, the data from the middle C2 channels are used as the initial observation state of the device domain agent, and the data from the last C3 channels are used as the initial observation state of the spatial domain agent.

[0108] Step S322: Set the policy space of the intelligent agent. The policy space of the environment domain intelligent agent contains a set of environmental feature adjustment actions, the policy space of the device domain intelligent agent contains a set of device operating state switching actions, and the policy space of the space domain intelligent agent contains a set of space resource allocation actions.

[0109] The policy space is the set of all possible policies that an agent can choose. The environmental characteristic adjustment action set is the set of actions that an agent in the environmental domain can take to adjust environmental characteristics, such as adjusting temperature, humidity, and light intensity. The device operating state switching action set is the set of actions that an agent in the device domain can take to switch device operating states, such as turning devices on, turning them off, and adjusting device power. The spatial resource allocation action set is the set of actions that an agent in the spatial domain can take to allocate spatial resources, such as adjusting personnel distribution and allocating space usage.

[0110] Step S323: Within a preset evolution time step, each agent selects an action from the policy space based on the current observation state, and influences the observation state of other agents through action interaction to generate a new joint observation state.

[0111] The preset evolution time step is the time interval for simulating the evolution process, for example, set to 1 hour. The current observation state is the environmental state observed by the agent at a given moment. Action interaction is the process by which agents influence each other through actions. The joint observation state is the environmental state commonly observed by all agents at a given moment. In the actual simulation, within each evolution time step, each agent selects an action from the policy space based on its current observation state. For example, if the environment domain agent observes that the current temperature is low, it selects an action to increase the temperature. If the device domain agent observes that the device power is too high, it selects an action to decrease the device power. If the space domain agent observes that a certain area is too crowded, it selects an action to move the people to other areas. The actions of each agent influence each other; for example, the environment domain agent's action to increase the temperature may cause the device domain agent to need to adjust the device power to adapt to the temperature change. Through this action interaction, a new joint observation state is generated.

[0112] Step S324: Record the joint observation state and corresponding action combination for each time step to form a set of evolutionary trajectories containing different action paths.

[0113] The evolutionary trajectory set is a collection containing all possible evolutionary trajectories. Each trajectory records the joint observation state and corresponding action combination at different time steps. The joint observation state is the environmental state jointly observed by all agents at a given moment. The corresponding action combination is the combination of actions chosen by each agent at a given moment.

[0114] In actual recording, after each evolutionary time step, the current joint observation state and the action combinations selected by each agent are recorded. For example, in the first time step, the joint observation state is a temperature of 20°C, equipment power of 5000 watts, and personnel distributed in areas A and B. The corresponding action combinations are: the environment domain agent increases the temperature, the equipment domain agent decreases the power, and the space domain agent moves the personnel. This information is recorded to form a point in an evolutionary trajectory. As the time step progresses, new joint observation states and action combinations are continuously recorded, forming a set of evolutionary trajectories containing different action paths.

[0115] Step S325: Extract features from the set of evolution trajectories, convert the state parameter sequence in the trajectory into a feature vector sequence, and generate a multi-path evolution feature sequence, with each feature sequence corresponding to a possible state evolution path.

[0116] Feature extraction extracts information that represents the state evolution characteristics from a set of evolutionary trajectories. The state parameter sequence is a sequence of parameters recorded in the evolutionary trajectory, representing the joint observation of states, such as the time-varying sequence of parameters like temperature, equipment power, and personnel density. The feature vector sequence is a sequence of feature vectors converted from the state parameter sequence, with each feature vector representing a state feature at a time step. The multi-path evolution feature sequence is a set of feature sequences obtained through feature extraction, with each feature sequence corresponding to a possible state evolution path.

[0117] In practical implementation, feature extraction can be performed using a principal component analysis (PCA)-based method. For each trajectory in the evolutionary trajectory set, its state parameter sequence is used as input. First, the state parameter sequence is standardized to ensure it has the same scale and range. Then, the covariance matrix of the state parameter sequence is calculated, and the principal components of the data are obtained by solving for the eigenvalues ​​and eigenvectors of the covariance matrix. The top K principal components with larger eigenvalues ​​are selected as principal features, and the state parameter sequence is projected onto these principal components to obtain the feature vector sequence. The feature vector sequences of all trajectories are combined to generate a multipath evolution feature sequence.

[0118] Step S330: Path filtering is performed on the multi-path evolution feature sequence, retaining the effective evolution paths whose evolution probability exceeds the preset threshold, decoding the feature sequence of the effective evolution path into a state parameter time series, and generating a multi-path state evolution map.

[0119] Path selection involves filtering paths with high evolution probabilities from multi-path evolution feature sequences. Evolution probability is the likelihood of each evolution path occurring. A preset threshold is a pre-defined evolution probability threshold used to select effective evolution paths. Effective evolution paths are those whose evolution probability exceeds the preset threshold. The state parameter time series is a sequence showing the changes in state parameters over time after decoding the feature sequences. A multi-path state evolution map is a map used to display multiple possible evolution paths of a building's state, visually illustrating how the building's state changes under different conditions.

[0120] In practical screening, probabilistic model-based methods can be used to calculate the evolutionary probability of each evolutionary path. For example, a Gaussian Mixture Model (GMM) can be used to model the multi-path evolutionary feature sequence, calculating the probability density value of each path and using it as the evolutionary probability. A preset threshold, such as 0.1, is set to retain valid evolutionary paths whose evolutionary probabilities exceed this threshold. For the feature sequences of valid evolutionary paths, they are decoded into state parameter time series using back projection. For example, during feature extraction, PCA is used to project the state parameter sequence onto principal components to obtain a feature vector sequence. During decoding, the feature vector sequence is projected back into the original state parameter space to obtain the state parameter time series. The state parameter time series of all valid evolutionary paths are plotted on a graph to generate a multi-path state evolution graph.

[0121] Step S340: Extract the fluctuation range of state parameters and path stability index of each path from the multi-path state evolution map, and calculate the average fluctuation range and average stability index of all effective evolution paths.

[0122] The range of state parameter fluctuations is the magnitude of change of state parameters over time within an evolutionary path. Path stability metrics measure the stability of an evolutionary path; for example, the standard deviation of state parameters can be used. The average range of fluctuations is the average of the ranges of state parameters across all effective evolutionary paths. The average stability metric is the average of the path stability metrics across all effective evolutionary paths.

[0123] In practical extraction, for each valid evolution path in the multi-path state evolution map, the maximum and minimum values ​​of its state parameters are calculated, and the difference between them represents the fluctuation range of the state parameters. The standard deviation of the state parameters is calculated and used as a path stability index. For example, for the temperature state parameter, its standard deviation under that path is calculated. The fluctuation ranges of the state parameters and the path stability index of all valid evolution paths are recorded, and their average values ​​are calculated to obtain the average fluctuation range and the average stability index.

[0124] Step S350: Input the average fluctuation range and the average stability index into the resilience assessment function to generate a state resilience index that characterizes the building's ability to resist interference. The state resilience index is negatively correlated with the average fluctuation range and positively correlated with the average stability index.

[0125] The resilience assessment function is used to evaluate the resilience of a building. It takes the average fluctuation range and average stability index as inputs and outputs a resilience index. The resilience index is an indicator used to measure the building's ability to withstand disturbances; it reflects the building's ability to maintain a stable state when faced with external disturbances.

[0126] In practical implementation, the resilience assessment function can be a linear combination, for example: State Resilience Index = A × Average Stability Index - B × Average Fluctuation Range; where A and B are weighting coefficients, the values ​​of which are determined according to the actual situation. Substituting the average fluctuation range and the average stability index into the resilience assessment function, the state resilience index is calculated. Since the state resilience index is negatively correlated with the average fluctuation range and positively correlated with the average stability index, the smaller the average fluctuation range and the higher the average stability index, the higher the state resilience index, indicating that the building has a stronger resistance to interference.

[0127] Step S400: Based on the multipath state evolution map and state resilience index, conduct a building state resilience assessment to locate the vulnerable areas of the building and the rate of vulnerability evolution.

[0128] Building resilience assessment is the process of evaluating a building's resilience based on a multipath state evolution map and a resilience index. Vulnerable areas are regions within a building that are easily affected by external disturbances and have poor state stability. The vulnerability evolution rate is the rate at which the vulnerability of a vulnerable area changes over time, reflecting the trend of vulnerability development.

[0129] As one implementation method, step S400 can be specifically implemented as the following steps S410~S450:

[0130] Step S410: Extract the state parameter sequence of each spatial region under different evolution paths from the multi-path state evolution map, calculate the deviation of the state parameter sequence of each spatial region from the preset toughness threshold, and generate the regional deviation sequence.

[0131] The state parameter sequence is a time-varying sequence of state parameters for each spatial region under different evolution paths in a multipath state evolution map. The preset resilience threshold is a pre-defined resilience standard for state parameters, used to measure whether the spatial region is in a stable state. The region deviation sequence is a sequence containing the deviations of the state parameter sequence for each spatial region from the preset resilience threshold.

[0132] In practical extraction, for each spatial region in the multi-path state evolution map, its state parameter sequence is extracted under different evolution paths. For example, for each floor area of ​​a building, under different state evolution paths, the changes in state parameters such as temperature, humidity, and personnel density over time are extracted. The preset resilience threshold can be determined according to the building's design standards and actual usage requirements. For example, for temperature parameters, the preset resilience threshold can be set to a reasonable temperature range.

[0133] When calculating the deviation of the state parameter sequence of each spatial region from the preset toughness threshold, for cases where the state parameters are continuous values, the absolute value error method can be used for calculation. Assume that the temperature state parameter value of a certain spatial region at a certain moment is T. i The lower limit of the preset toughness threshold is T. min The upper limit is T max Then the temperature deviation d at that moment i The following can be calculated: when T i < T min hour, When T i >T max hour, When T min ≤T i ≤T max At that time, d i =0.

[0134] For the entire sequence of state parameters, the deviation at each moment is recorded sequentially to obtain the regional deviation sequence for that spatial region. For other state parameters, such as humidity and population density, a similar method is used to calculate the deviation. Ultimately, a comprehensive regional deviation sequence is generated for each spatial region, which reflects the degree to which the state parameters of that region deviate from the preset resilience standard under different evolutionary paths.

[0135] Step S420: The regional deviation sequence and the state resilience index are weighted and fused to generate a regional resilience assessment value. The weights of the weighted fusion are dynamically adjusted according to the regional importance level.

[0136] The regional deviation sequence reflects the deviation of spatial region state parameters from preset resilience thresholds, while the state resilience index measures the building's overall resistance to interference. Weighted fusion combines these two factors to comprehensively assess the resilience of each spatial region. Regional importance levels are determined based on the spatial region's importance within the building's overall functional architecture; for example, core office areas and important equipment rooms have higher importance levels, while ordinary storage rooms have lower importance levels.

[0137] In practice, the weights for weighted fusion are first determined. A weight mapping table can be established based on the importance level of the regions. Regions with high importance are given higher weights for their state resilience index and lower weights for their deviation sequence; conversely, regions with low importance are given lower weights for their state resilience index and higher weights for their deviation sequence. Then, weighted fusion calculations are performed. This weighted fusion method comprehensively considers both the overall anti-interference capability of a region and the deviation of its own state parameters, resulting in a more accurate regional resilience assessment value that reflects the resilience status of each spatial region.

[0138] Step S430: Sort the regional resilience assessment values ​​and select spatial regions with assessment values ​​lower than the preset resilience threshold as candidate vulnerable regions.

[0139] A preset resilience threshold is used to determine whether a space region has a high degree of vulnerability. All space regions are sorted in ascending order of their regional resilience assessment values; the lower the assessment value, the worse the region's resilience and the more susceptible it is to external disturbances. When a region's resilience assessment value is below the preset resilience threshold, it indicates poor state stability and a significant risk of vulnerability, and such regions are selected as candidate vulnerable regions.

[0140] The predetermined resilience threshold can be determined based on the building's overall safety requirements and historical data. For example, by statistically analyzing the status data of similar buildings under different disturbance conditions, a reasonable lower limit for resilience assessment can be determined as the predetermined resilience threshold. The selected candidate vulnerable areas are the key targets for further analysis of vulnerability evolution rates and the development of control strategies.

[0141] Step S440: Perform trend fitting on the regional deviation sequence of candidate vulnerable regions, calculate the growth slope of deviation over time, and use the growth slope as the vulnerability evolution rate.

[0142] The regional deviation sequence reflects how the state parameters of candidate vulnerable regions deviate from a preset resilience threshold over time. Trend fitting uses mathematical methods to find a suitable function to approximate the trend of the regional deviation sequence, thereby calculating the slope of the deviation increase over time. This slope is the vulnerability evolution rate, reflecting the speed at which the vulnerability of the candidate vulnerable region develops over time.

[0143] In one implementation, step S440 can be specifically implemented as the following steps S441-S444:

[0144] Step S441: Perform outlier removal processing on the regional deviation sequence of candidate vulnerable regions, and use the moving average method to smooth the sequence fluctuations to generate a smooth deviation sequence.

[0145] Outliers are values ​​in a regional deviation sequence that significantly deviate from the normal data distribution. These outliers may be caused by sensor malfunctions, temporary interference, or other reasons. If left untreated, they can significantly impact subsequent trend fitting and slope calculation. Outlier removal can be achieved using statistical thresholding methods. For example, the mean and standard deviation of the regional deviation sequence can be calculated, and data points deviating from the mean by more than k times the standard deviation (e.g., k=3) can be considered outliers and removed. The moving average method reduces data fluctuation by calculating the average value of data within a certain window. Through outlier removal and the moving average method, the generated smoothed deviation sequence more accurately reflects the true trend of deviation in candidate vulnerable regions.

[0146] Step S442: Normalize the time axis of the smooth deviation sequence and convert the timestamps into relative time steps to ensure that the time dimension of different candidate regions is consistent.

[0147] The regional deviation sequences of different candidate vulnerable regions may be collected within different time ranges, with varying start and interval timestamps, which can complicate subsequent trend fitting and comparison. Time axis normalization converts the timestamps of different candidate regions into a uniform relative time step, ensuring that the time dimension of all candidate regions is consistent. By mapping the smoothed deviation sequences to relative time steps, the regional deviation sequences of different candidate regions can be compared and analyzed on the same time dimension.

[0148] Step S443: Use the least squares method to perform linear fitting on the normalized smooth deviation sequence, generate the fitted line equation, and extract the slope parameter in the equation; if the slope is positive, it indicates that the deviation increases with time, and the slope value is the vulnerability evolution rate; if the slope is negative, it indicates that the deviation decreases with time, and the vulnerability evolution rate is recorded as zero.

[0149] The least squares method is a common linear fitting method. It determines the parameters of the fitted line by minimizing the sum of squared errors between the observed data and the fitted line. It will not be described in detail here.

[0150] Step S444: Associate and store the vulnerability evolution rate of each candidate vulnerable region with the corresponding regional resilience assessment value to generate a vulnerability assessment table containing the region number, resilience assessment value, and evolution rate.

[0151] Linking the vulnerability evolution rate of each candidate vulnerable region with its corresponding regional resilience assessment value facilitates comprehensive analysis and management of these regions. The region ID serves as a unique identifier for each candidate vulnerable region, allowing for quick location and retrieval of relevant information. The vulnerability assessment table can be structured as a single table containing three columns: region ID, resilience assessment value, and evolution rate.

[0152] Step S450: Prioritize the candidate vulnerable regions according to their vulnerability evolution rate, select the candidate vulnerable region with the fastest evolution rate as the state vulnerable region, and use its corresponding vulnerability evolution rate as the final evaluation parameter.

[0153] The vulnerability evolution rate reflects how quickly the vulnerability of a candidate vulnerable area develops over time. The faster the evolution rate, the more rapidly the vulnerability of the area may deteriorate in a short period of time, and the greater the impact on the overall condition of the building. Therefore, prioritizing candidate vulnerable areas based on their vulnerability evolution rate can help identify areas that require priority attention.

[0154] Candidate vulnerable areas are sorted from highest to lowest vulnerability evolution rate, and the candidate vulnerable area with the fastest evolution rate is selected as the state vulnerable area. This area is the most easily disturbed and fastest developing vulnerability area in the current building, and its corresponding vulnerability evolution rate is used as the final evaluation parameter for subsequent self-healing control strategy formulation.

[0155] Step S500: Generate a building condition self-healing regulation strategy based on the vulnerable areas and vulnerability evolution rate.

[0156] Building self-healing control strategies are a series of measures implemented to address the vulnerability of vulnerable areas by adjusting factors such as the building's environment, equipment, and spatial resources to restore the building to a stable state. The vulnerable areas and their evolution rate are crucial factors in formulating self-healing control strategies; different vulnerable areas and evolution rates require different control strategies.

[0157] As one implementation method, step S500 can be specifically implemented as the following steps S510-S550:

[0158] Step S510: Analyze the functional attributes and numerical characteristics of the vulnerability evolution rate of the vulnerable area, extract the importance parameters of the area in the overall functional architecture of the building and the urgency index corresponding to the vulnerability evolution rate, and determine the allocation ratio of control resources by combining the importance parameters and urgency index.

[0159] The functional attributes of a vulnerable area refer to its function within the building, such as whether it is an office area, a server room, or a public area. Different functional attributes have varying degrees of impact on the overall function of the building. The numerical characteristic of the vulnerability evolution rate reflects the speed at which the vulnerability develops in that area; the higher the value, the more urgent the problem.

[0160] Importance parameters can be determined based on the status and role of vulnerable areas within the overall functional architecture of the building. For example, the core computer room area has a higher importance parameter, while ordinary storage rooms have a lower importance parameter. Urgency indicators can be categorized based on the rate of vulnerability evolution. For instance, vulnerability evolution rates can be divided into several levels, each corresponding to an urgency indicator. An evolution rate of 0-0.05 indicates low urgency, 0.05-0.1 indicates medium urgency, and greater than 0.1 indicates high urgency.

[0161] The allocation ratio of regulatory resources is determined by combining importance parameters and urgency indicators. Regions with both high importance and high urgency are allocated more regulatory resources; regions with both low importance and low urgency are allocated fewer resources. For example, a vulnerable region with an importance parameter of 0.8 and a high urgency indicator could be allocated 70% of its regulatory resources; a region with an importance parameter of 0.2 and a low urgency indicator could be allocated only 10%.

[0162] Step S520: Based on the allocation ratio, select a set of candidate control actions that match the functional attributes of the vulnerable area from the preset control strategy knowledge base. The set of candidate control actions includes environmental parameter adjustment actions, equipment operation mode switching actions, and spatial resource dynamic scheduling actions. The selection process should refer to the feature distribution pattern of the environment-equipment interaction feature matrix and space-environment feedback feature map of the area in the building status association feature set.

[0163] The pre-defined control strategy knowledge base is a database that stores various control strategies and actions, including control solutions for regions with different functional attributes and different state problems. Environmental parameter adjustment actions include actions to adjust environmental parameters such as temperature, humidity, and light intensity; equipment operation mode switching actions include actions to turn equipment on and off, and adjust equipment power and speed; and spatial resource dynamic scheduling actions include actions to adjust personnel distribution and allocate space usage.

[0164] When selecting candidate control actions, the first step is to filter potentially applicable control actions from the control strategy knowledge base based on the functional attributes of vulnerable areas. For example, for office areas, potentially applicable control actions include adjusting air conditioning temperature and optimizing lighting systems; for computer room areas, potentially applicable control actions include increasing ventilation equipment power and adjusting server operating modes. Simultaneously, the feature distribution patterns of the environment-equipment interaction feature matrix and space-environment feedback feature map for this area are referenced from the building status association feature set. The environment-equipment interaction feature matrix reflects the interaction relationship between environmental features and equipment features. By analyzing this matrix, the impact of changes in environmental parameters on equipment operation can be understood, thereby selecting appropriate control actions. The space-environment feedback feature map reflects the feedback relationship between spatial features and environmental features. Based on this map, the correlation between the scheduling of spatial resources and the adjustment of environmental parameters can be determined, further optimizing the candidate control action set.

[0165] Step S530: Construct a virtual control environment, simulate the execution of each action in the candidate control action set in the virtual control environment, record the change curve of the vulnerability evolution rate of the vulnerable region during the execution process, and determine the control effect evaluation value of each action by comparing the slope and convergence time of the change curves corresponding to different actions.

[0166] The virtual control environment is a simulation environment built on a building state model. It can simulate the actual operation of a building and the effects of various control actions. In the virtual control environment, each action in the candidate control action set can be simulated and executed sequentially to observe the changes in the vulnerability evolution rate of vulnerable areas.

[0167] The vulnerability evolution rate curve records the change in the vulnerability evolution rate of a vulnerable region over time during the simulated execution of control actions. The slope reflects the rate of change of the vulnerability evolution rate; the larger the slope, the faster the vulnerability evolution rate decreases, and the better the control effect. The convergence time is the time it takes for the vulnerability evolution rate to stabilize; the shorter the convergence time, the faster the control action can achieve a stabilizing effect.

[0168] By comparing the slope and convergence time of the change curves corresponding to different actions, the evaluation value of the regulatory effect of each action can be determined. For example, a weighted average method can be used to assign weights to the slope and convergence time to calculate the evaluation value of the regulatory effect.

[0169] Step S540: Select the candidate control action with the highest evaluation value as the core control action based on the control effect evaluation value. Adaptively adjust the intensity, scope and execution sequence of the core control action in combination with the dynamic change trend of vulnerability evolution rate to generate a preliminary building condition self-healing control strategy.

[0170] The evaluation value of the regulation effect reflects the regulation effect of each candidate regulation action on the vulnerable region. Selecting the candidate regulation action with the highest evaluation value as the core regulation action can maximize the effectiveness of regulation.

[0171] The dynamic trend of vulnerability evolution rate refers to the change in vulnerability evolution rate over time during the simulated execution of control actions. Based on this dynamic trend, the intensity, scope, and timing of core control actions can be adaptively adjusted. For example, if the vulnerability evolution rate decreases slowly, the intensity of the core control action can be appropriately increased; if the vulnerable area is large, the scope of the core control action can be expanded; if the vulnerability evolution rate changes significantly within a certain period, the timing of the core control action can be adjusted to strengthen control during critical periods.

[0172] Through such adaptive adjustments, the resulting preliminary building condition self-healing control strategy can better adapt to the actual situation of vulnerable areas, improving the accuracy and effectiveness of control.

[0173] Step S550: Input the preliminary building condition self-healing control strategy into the strategy verification model. Predict the vulnerability evolution trend of the vulnerable area within a preset time window after implementing the strategy. If the predicted trend shows that the vulnerability evolution rate can decrease to below a preset safety threshold, the preliminary building condition self-healing control strategy is determined as the final building condition self-healing control strategy, and the final building condition self-healing control strategy is pushed to the building control center for condition intervention. If the preset safety threshold is not reached, the core control actions are re-selected and the parameters are adjusted, and the simulation verification process is repeated until the conditions are met. The strategy verification model is a machine learning or deep learning-based model that can predict the vulnerability evolution trend of the vulnerable area after implementing the control strategy based on the building's historical data and current state. The preset time window is a pre-set time range within which the vulnerability evolution trend is observed. The preset safety threshold is a pre-set safety standard for the vulnerability evolution rate. When the vulnerability evolution rate decreases to below this threshold, it indicates that the vulnerability of the vulnerable area has been effectively controlled, and the building condition tends to stabilize.

[0174] The initial building condition self-healing control strategy is input into the strategy verification model for prediction. If the prediction results show that the vulnerability evolution rate can decrease to below a preset safety threshold, the initial strategy is determined as the final building condition self-healing control strategy and pushed to the building control center. The building control center is the core of the building automation system. It can control and adjust the building's equipment and environment in real time according to the received control strategy, realizing condition intervention operations. If the prediction results show that the vulnerability evolution rate does not reach the preset safety threshold, it indicates that the current initial strategy is not effective, and the core control actions need to be re-selected and the parameters adjusted. A suitable action is selected again from the candidate control action set, and its intensity, scope, and execution sequence are adjusted to generate a new initial strategy. This new strategy is then input into the strategy verification model for simulation verification. This process is repeated until the prediction results show that the vulnerability evolution rate can decrease to below the preset safety threshold, finally determining the final building condition self-healing control strategy that meets the conditions.

[0175] Through the above steps, comprehensive modeling, analysis, and control of the state of intelligent buildings are achieved. This enables timely detection of vulnerable areas in the building, prediction of their vulnerability evolution trends, and the formulation of effective self-healing control strategies to improve the stability and anti-interference capabilities of the building.

[0176] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0177] Please see Figure 2 , Figure 2This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space. In one embodiment, the processor 101 executes the intelligent building state modeling method using deep learning provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for intelligent building state modeling using deep learning, characterized in that, include: Obtain the distributed state perception stream of the building, which includes environmental domain perception sequence, device domain perception sequence and spatial domain perception sequence generated by different sensing nodes within a continuous monitoring period; Cross-domain feature association modeling is performed on the distributed state perception stream, and a building state association feature set with dynamic association attributes is generated by calculating the inter-domain feature interaction strength. Specifically, this includes: synchronizing the environmental domain perception sequence, equipment domain perception sequence, and spatial domain perception sequence in the distributed state perception stream with timestamps to generate a synchronized perception sequence with a unified time base; calculating the interaction strength between the environmental domain perception sequence and the equipment domain perception sequence in the synchronized perception sequence, extracting the time lag association parameters between environmental feature fluctuations and equipment feature responses, and constructing an environment-equipment interaction feature matrix based on the time lag association parameters; and coupling the equipment domain perception sequence and the spatial domain perception sequence in the synchronized perception sequence. Path mining identifies co-occurrence patterns of equipment feature changes and spatial feature distribution, and generates equipment-space coupling feature vectors based on these co-occurrence patterns. Feedback loops are constructed for the spatial domain sensing sequence and the environmental domain sensing sequence in the synchronous sensing sequence. The influence weight of spatial feature gradients on the steady state of environmental features is analyzed, and a spatial-environmental feedback feature map is generated based on the influence weights. The environmental-equipment interaction feature matrix, the equipment-space coupling feature vectors, and the spatial-environmental feedback feature map are input into a dynamic association fusion module. An attention mechanism dynamically allocates association weights for features from different domains, and feature aggregation is performed based on these association weights to generate a building status association feature set with dynamic association attributes. A pre-trained deep state evolution game network is invoked to perform state trend inference on the building state association feature set, generating a multi-path state evolution map and a state resilience index for the building. Specifically, this includes: inputting the building state association feature set into the association feature encoding layer of the deep state evolution game network; performing spatiotemporal association modeling on the building state association feature set using a spatiotemporal graph convolutional network to generate a spatiotemporal association feature tensor; inputting the spatiotemporal association feature tensor into the multi-agent game layer of the deep state evolution game network; initializing virtual game agents in the environment domain, device domain, and spatial domain; simulating the evolutionary game process of different domain features through policy interactions between agents to generate multiple... The path evolution feature sequence is used for path screening, retaining effective evolution paths with evolution probabilities exceeding a preset threshold. The feature sequence of the effective evolution paths is decoded into a state parameter time series to generate a multi-path state evolution map. The state parameter fluctuation range and path stability index of each path are extracted from the multi-path state evolution map, and the average fluctuation range and average stability index of all effective evolution paths are calculated. The average fluctuation range and average stability index are input into a resilience evaluation function to generate a state resilience index characterizing the building's ability to resist interference. The state resilience index is negatively correlated with the average fluctuation range and positively correlated with the average stability index. Building resilience assessment is performed based on the multi-path state evolution map and the state resilience index to locate vulnerable areas and vulnerability evolution rates. Specifically, this includes: extracting state parameter sequences for each spatial region under different evolution paths from the multi-path state evolution map; calculating the deviation of each spatial region's state parameter sequence from a preset resilience threshold to generate a regional deviation sequence; weighting and fusing the regional deviation sequence with the state resilience index to generate a regional resilience assessment value, where the weights of the weighting and fusing are dynamically adjusted according to the region's importance level; sorting the regional resilience assessment values ​​and selecting spatial regions with assessment values ​​below the preset resilience threshold as candidate vulnerable areas; performing trend fitting on the regional deviation sequences of the candidate vulnerable areas to calculate the slope of deviation growth over time, using the growth slope as the vulnerability evolution rate; prioritizing the candidate vulnerable areas according to the vulnerability evolution rate, selecting the candidate vulnerable area with the fastest evolution rate as the state vulnerable area, and using its corresponding vulnerability evolution rate as the final assessment parameter. A building condition self-healing regulation strategy is generated based on the state-vulnerable regions and the vulnerability evolution rate.

2. The method according to claim 1, characterized in that, The step involves calculating the interaction strength between the environment domain sensing sequence and the device domain sensing sequence in the synchronized sensing sequence, extracting the time lag correlation parameters between environmental feature fluctuations and device feature responses, and constructing an environment-device interaction feature matrix based on the time lag correlation parameters, including: Continuous feature segments of the environmental domain sensing sequence are extracted from the synchronous sensing sequence, and the environmental feature fluctuations of adjacent timestamps are calculated to generate an environmental feature fluctuation sequence. Extract continuous feature segments from the device domain sensing sequence from the synchronous sensing sequence, calculate the device feature response quantities of adjacent timestamps, and generate a device feature response sequence; The environmental characteristic fluctuation sequence and the equipment characteristic response sequence are input into the interaction analysis unit to calculate the interaction strength value under different lag times and generate the interaction strength-lag time relationship curve. Peak values ​​are extracted from the interaction strength-hysteresis duration relationship curve. Peak points with interaction strength values ​​exceeding a preset threshold are selected, and the hysteresis duration parameters and interaction strength values ​​corresponding to the peak points are recorded to construct an environment-equipment interaction hysteresis record table. Using environmental feature type as the row dimension and device feature type as the column dimension, the interaction strength value in the environment-device interaction time delay record table is used as the matrix element value, and the lag duration parameter is used as the element attribute to construct an environment-device interaction feature matrix. The element value of the environment-device interaction feature matrix represents the interaction strength between the corresponding environmental feature and the device feature.

3. The method according to claim 1, characterized in that, The process of performing coupling path mining on the device domain sensing sequence and spatial domain sensing sequence in the synchronous sensing sequence, identifying co-occurrence patterns of device feature changes and spatial feature distributions, and generating a device-space coupling feature vector based on the co-occurrence patterns includes: Feature mutation point detection is performed on the device domain sensing sequence in the synchronous sensing sequence to identify the mutation time points of device feature values ​​from steady state to non-steady state or from non-steady state to steady state, and to generate a set of device feature mutation events; The spatial domain sensing sequences in the synchronous sensing sequence are clustered according to their distribution patterns. The spatial feature distribution is divided into different distribution pattern categories. The start and end timestamps of each category are recorded to generate a set of spatial distribution pattern time intervals. Align the set of device feature mutation events with the set of spatial distribution pattern time intervals on the time axis, calculate the probability that each device feature mutation time point falls into the spatial distribution pattern time interval, and generate a time coupling probability value. Based on the temporal coupling probability value, highly coupled event pairs are screened, and the co-occurrence relationship between the device feature mutation type and the spatial distribution pattern category in the highly coupled event pairs is extracted to generate a device-space co-occurrence pattern table. The co-occurrence frequency values ​​in the device-space co-occurrence pattern table are sorted according to the preset device type priority to generate a device-space coupling feature vector with the same dimension as the number of device types. The element values ​​of the coupling feature vector represent the coupling strength between the corresponding device type and the spatial distribution pattern.

4. The method according to claim 1, characterized in that, The step of constructing a feedback loop for the spatial domain sensing sequence and the environmental domain sensing sequence in the synchronous sensing sequence, analyzing the influence weight of the spatial feature gradient on the steady state of environmental features, and generating a spatial-environment feedback feature map based on the influence weight includes: Extract the spatial feature gradient parameters of the spatial domain sensing sequence and the environmental feature steady-state parameters of the corresponding timestamp environmental domain sensing sequence from the synchronous sensing sequence to generate a spatial-environment correlation parameter pair sequence. The spatial-environment correlation parameter pair sequence is segmented, and divided into multiple parameter intervals according to the continuous change interval of the spatial feature gradient parameters. Each parameter interval corresponds to a set of environmental feature steady-state parameter change data. Calculate the rate of change of the steady-state parameters of the environmental features within each parameter interval, and weight the rate of change with the length of the parameter interval to obtain the influence weight of the spatial feature gradient on the steady-state of the environmental features. A directed spatial-environment feedback graph is constructed using the spatial regions of buildings as nodes and the influence weights as edge weights. The node attributes of the directed graph include the mean gradient of the regional spatial features, and the edge attributes include the influence weights and the feedback direction. The spatial-environment feedback directed graph is subjected to graph feature extraction. The node attributes and edge attributes are fused into a graph adjacency matrix and a node feature matrix to generate a spatial-environment feedback feature graph.

5. The method according to claim 1, characterized in that, The step of inputting the building state association feature set into the association feature encoding layer of the deep state evolution game network, and performing spatiotemporal association modeling on the building state association feature set through a spatiotemporal graph convolutional network to generate a spatiotemporal association feature tensor includes: The environment-device interaction feature matrix, device-space coupling feature vector and space-environment feedback feature map in the building status association feature set are converted into feature tensors of a unified dimension to ensure that the time dimension and spatial dimension of each feature tensor are aligned. Construct a spatiotemporal relational graph structure for buildings, using the spatial location of the sensing nodes as the graph node coordinates and the temporal and spatial correlations between nodes as edge weights, to generate a spatiotemporal relational adjacency matrix. The feature tensor and the spatiotemporal association adjacency matrix are input into the temporal convolutional layer of the spatiotemporal graph convolutional network. The temporal dimension features are extracted from the feature tensor through causal convolution kernels, preserving the causal relationship of the time series and generating a temporal association feature map. The temporally correlated feature map is input into the spatial graph convolutional layer of the spatiotemporal graph convolutional network. The spatial neighbor feature information of each node is aggregated through graph convolution operations to generate a spatially correlated feature map. The temporal correlation feature map and the spatial correlation feature map are concatenated dimensionally, and the temporal and spatial correlation information is fused along the channel dimension to generate a spatiotemporal correlation feature tensor with dual spatiotemporal correlation attributes.

6. The method according to claim 1, characterized in that, The process involves inputting the spatiotemporal correlation feature tensor into the multi-agent game layer of the deep state evolution game network, initializing virtual game agents in the environment domain, device domain, and spatial domain, simulating the evolutionary game process of different domain features through policy interactions between agents, and generating multi-path evolutionary feature sequences, including: Initialize the environment domain agent, device domain agent, and spatial domain agent, respectively using the feature sub-tensors of the corresponding domains in the spatiotemporal correlation feature tensor as the initial observation states of the agents; Set up the policy space of the intelligent agent. The policy space of the environment domain intelligent agent contains a set of actions to adjust environmental features. The policy space of the device domain intelligent agent contains a set of actions to switch device operating states. The policy space of the space domain intelligent agent contains a set of actions to allocate space resources. Within a preset evolution time step, each agent selects an action from the policy space based on its current observation state, and influences the observation state of other agents through action interaction to generate a new joint observation state. Record the joint observation state and corresponding action combination at each time step to form a set of evolutionary trajectories containing different action paths; Feature extraction is performed on the set of evolution trajectories, and the sequence of state parameters in the trajectory is converted into a sequence of feature vectors to generate a multi-path evolution feature sequence. Each feature sequence corresponds to a possible state evolution path.

7. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the intelligent building state modeling method using deep learning as described in any one of claims 1-6.

Images