Intelligent property management system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence and intelligent property management, in particular to an intelligent property management system based on artificial intelligence, which comprises a scene perception fusion module, a semantic cognition modeling module, a self-evolution learning engine module, an intelligent regulation and control execution module and an interaction module, according to the invention, through collaborative design of the scene perception fusion module and the semantic cognition modeling module, a semantic modeling mechanism of a knowledge graph and a graph attention network is introduced into intelligent property management, and the multi-modal information perception precision and dynamic response capability of the system in a complex scene are improved; the invention provides a self-evolution learning mechanism fusing a multi-step aggregation reward function and a depth policy mapping optimization generator, a task and resource relationship is modeled through a multilayer feedforward neural network, a parameter self-learning process is optimized by using a multi-step accumulation reward, and a candidate policy generation and enhancement mechanism based on a depth policy mapping function is introduced. And the intelligent level of property energy efficiency management and task scheduling is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and intelligent property management, and particularly relates to an intelligent property management system based on artificial intelligence. BACKGROUND

[0002] With the deepening of urban intelligentization and digital transformation, the traditional property management mode gradually exposes the following main problems:

[0003] (1) The problem of multi-source information isolation and data fragmentation

[0004] At present, the property management system is usually composed of independent subsystems such as video monitoring, access control system, energy consumption monitoring, equipment operation state monitoring, etc. There is a lack of unified data interface and semantic association between each subsystem, which leads to inefficient fusion and sharing of information, weak response ability to complex scene changes, and difficulty in forming global intelligent judgment.

[0005] (2) The problem of low intelligentization degree of resource scheduling and management decision

[0006] The existing system mainly relies on static rules or preset thresholds for device control and energy consumption management, and lacks an intelligent scheduling mechanism based on data self-learning and optimization. When the property environment or task load dynamically changes, the system is difficult to adjust the resource allocation strategy in time, resulting in energy waste or service response delay. SUMMARY

[0007] In view of the above situation, in order to overcome the defects of the prior art, the present application provides an intelligent property management system based on artificial intelligence. In view of the problem of multi-source information isolation and data fragmentation, the present application introduces the semantic modeling mechanism of knowledge graph and graph attention network through the collaborative design of scene perception fusion module and semantic cognition modeling module in intelligent property management, realizes the leap from "perception layer fusion" to "cognition layer understanding", and significantly improves the multi-modal information perception accuracy and dynamic response ability of the system in complex scenes. In view of the problem of low intelligentization degree of resource scheduling and management decision, the present application proposes a self-evolution learning mechanism that integrates multi-step aggregated reward function and deep policy mapping optimization generator. The task and resource relationship is modeled by a multi-layer feedforward neural network, the parameter self-learning process is optimized by multi-step cumulative reward, and the candidate strategy generation and enhancement mechanism based on deep policy mapping function is introduced. The scheduling strategy is continuously self-adaptively optimized in operation, realizing dynamic self-learning and self-evolution update of resource allocation, and significantly improving the intelligent level of property energy efficiency management and task scheduling.

[0008] The technical solution adopted in this invention is as follows: This invention provides an artificial intelligence-based smart property management system, including a scene perception and fusion module, a semantic cognition modeling module, a self-evolutionary learning engine module, an intelligent control and execution module, and an interaction module, specifically including the following:

[0009] The scene perception fusion module collects multi-dimensional property information in real time, including video surveillance streams, access control records, energy consumption sensor data, environmental quality parameters, equipment operating status and personnel behavior data. It performs feature decoupling and multi-source data fusion on the multi-dimensional property information to obtain scene feature vectors.

[0010] The semantic cognition modeling module performs semantic layer modeling on scene feature vectors, constructs a property semantic network based on knowledge graphs, and uses graph attention networks to embed and update complex semantic relationships to obtain the updated property semantic network.

[0011] The self-evolutionary learning engine module collects historical task data sets, constructs a dynamic task allocation mechanism based on a multi-layer feedforward neural network, introduces a multi-step aggregated reward function, and performs adaptive optimization of parameters based on a gradient self-update strategy to obtain the optimal resource scheduling strategy.

[0012] The intelligent control and execution module performs intelligent dynamic control of property equipment, energy consumption and environmental systems based on the optimal resource scheduling strategy;

[0013] The interactive module constructs an intelligent human-machine collaborative interface.

[0014] Furthermore, the self-evolutionary learning engine module specifically includes the following steps:

[0015] Step S1: Collect a set of historical task data, represented as Each of the tasks Includes the amount of resources required for execution and completion time Collect the current system resource status and represent it as R 0 =[ r 1 , r 2 , ... , r M ] ,in Indicates the first The availability of each resource unit;

[0016] Step S2: Generate the initial task allocation matrix. Based on the historical task data set and the current resource status, generate the initial task allocation matrix using an association mapping function, as shown below:

[0017] ;

[0018] wherein, denotes the initial task allocation matrix, denotes the association mapping function;

[0019] Step S3: Resource allocation probability modeling, a multi-layer feedforward neural network is constructed, denoted as follows:

[0020] π θ ( X t )=[ p 1 , p 2 , ... , p N ] ;

[0021] wherein, denotes the multi-layer feedforward neural network, denotes the task allocation matrix at the current time step, denotes the multi-layer feedforward neural network weight set, denotes the total number of tasks, denotes the task probability of being allocated to the optimal resource unit;

[0022] Step S4: Task allocation, for each task in the historical task data set, select the allocation action of the resource unit, the formula used is as follows:

[0023] ;

[0024] wherein, denotes the allocation action performed on task at time step ;

[0025] Update the task allocation matrix after performing the allocation action, denoted as follows:

[0026] ;

[0027] wherein, denotes the updated task allocation matrix, denotes the allocation matrix update function;

[0028] Step S5: Multi-step reward aggregation, define Step S5: Multi-step reward aggregation, define :

[0029] ;

[0030] wherein, denotes the aggregation step size, denotes the cumulative step index, denotes the immediate reward value, denotes the reward function, denotes the proportion allocated to the resource unit, denotes the task resource allocation state matrix, denotes the trade-off coefficient vector, denotes the environmental dynamic factor;

[0031] Step S6: Self-evolution strategy update, update the parameters of the multi-layer feedforward neural network according to the aggregated reward, denoted as follows:

[0032] ;

[0033] wherein, denotes the learning rate;

[0034] Step S7: Optimal resource scheduling strategy output, set the maximum number of iterations, repeat the iteration of steps S2 to S6 until the maximum number of iterations is reached, extract the parameters of the current multi-layer feedforward neural network, design an optimization generator based on a deep policy mapping function, and generate an optimal resource scheduling strategy through a parameter-driven policy reasoning mechanism.

[0035] Further, step S7, design an optimization generator based on a deep policy mapping function, and generate an optimal resource scheduling strategy through a parameter-driven policy reasoning mechanism, which specifically includes the following steps:

[0036] Step S71: Parameter encoding, encode the parameters of the current multi-layer feedforward neural network into a parameter vector, and initialize the deep policy mapping function using a high-dimensional nonlinear mapping function;

[0037] Step S72: Strategy candidate generation, generate a candidate strategy set through the deep policy mapping function, denoted as:

[0038] ;

[0039] ;

[0040] wherein, denotes the th candidate strategy, denotes the main strategy vector based on the parameter vector; denotes the candidate strategy disturbance vector;

[0041] Step S73: Candidate strategy evaluation, calculating the comprehensive payoff function for the candidate strategy set:

[0042] ;

[0043] in, This represents the overall return value of the candidate strategies. Indicates task throughput. Indicates the degree of satisfaction of the delay constraint. Indicates resource utilization rate, , and This is a vector of weighting coefficients;

[0044] Step S74: Strategy Enhancement. The comprehensive benefit values ​​of candidate strategies are sorted, and high-yield candidate strategies are selected. Enhanced strategies are generated based on the updated task allocation matrix and the fine-tuning function of the resource state. The formula used is as follows:

[0045] ;

[0046] in, This indicates the enhanced strategy;

[0047] Step S75: Optimal policy selection, combining the enhancement policies into a set of enhancement policies. The strategy with the highest overall return is selected from among them, using the following formula:

[0048] s opt = arg max s i ' ∈ S e [P s i ' · R s i ' ] ;

[0049] ;

[0050] in, This represents the return sensitivity coefficient. This represents the overall return value of the enhanced strategy. Indicating reinforcement strategy The probability of being selected as the optimal strategy. This represents the optimal resource scheduling strategy.

[0051] The beneficial effects achieved by the present invention using the above solution are as follows:

[0052] (1) To address the problems of isolated multi-source information and fragmented data, this invention introduces the semantic modeling mechanism of knowledge graph and graph attention network into smart property management through the collaborative design of scene perception fusion module and semantic cognition modeling module, realizing the leap from "perception layer fusion" to "cognition layer understanding", which significantly improves the system's multimodal information perception accuracy and dynamic response capability in complex scenarios;

[0053] (2) To address the problem of low intelligence in resource scheduling and management decision-making, this invention proposes a self-evolutionary learning mechanism that integrates a multi-step aggregated reward function and a deep policy mapping optimization generator. By modeling the relationship between tasks and resources through a multi-layer feedforward neural network, the self-learning process of parameters is optimized using a multi-step cumulative reward, and a candidate policy generation and enhancement mechanism based on a deep policy mapping function is introduced. During operation, the scheduling policy is continuously adaptively optimized, realizing dynamic self-learning and self-evolutionary updates of resource allocation, which significantly improves the intelligence level of property energy efficiency management and task scheduling. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of an artificial intelligence-based smart property management system proposed in this invention;

[0055] Figure 2 This is a flowchart illustrating the self-evolutionary learning engine module proposed in this invention.

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0057] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0058] Example 1, see Figure 1 This invention provides an artificial intelligence-based smart property management system, comprising a scene perception and fusion module, a semantic cognition modeling module, a self-evolutionary learning engine module, an intelligent control and execution module, and an interaction module, specifically including the following:

[0059] The scene perception fusion module collects multi-dimensional property information in real time, including video surveillance streams, access control records, energy consumption sensor data, environmental quality parameters, equipment operating status and personnel behavior data. It performs feature decoupling and multi-source data fusion on the multi-dimensional property information to obtain scene feature vectors.

[0060] The semantic cognition modeling module performs semantic layer modeling on scene feature vectors, constructs a property semantic network based on knowledge graphs, and uses graph attention networks to embed and update complex semantic relationships to obtain the updated property semantic network.

[0061] The self-evolutionary learning engine module collects historical task data sets, constructs a dynamic task allocation mechanism based on a multi-layer feedforward neural network, introduces a multi-step aggregated reward function, and performs adaptive optimization of parameters based on a gradient self-update strategy to obtain the optimal resource scheduling strategy.

[0062] The intelligent control and execution module performs intelligent dynamic control of property equipment, energy consumption and environmental systems based on the optimal resource scheduling strategy;

[0063] The interactive module constructs an intelligent human-machine collaborative interface.

[0064] Example 2, see Figure 2 This embodiment is based on the above embodiment, and the self-evolutionary learning engine module specifically includes the following steps:

[0065] Step S1: Collect a set of historical task data, represented as Each of the tasks Includes the amount of resources required for execution and completion time Collect the current system resource status and represent it as R 0 =[ r 1 , r 2 , ... , r M ] ,in Indicates the first The availability of each resource unit;

[0066] Step S2: Generate the initial task allocation matrix. Based on the historical task data set and the current resource status, generate the initial task allocation matrix using an association mapping function, as shown below:

[0067] ;

[0068] in, This represents the initial task assignment matrix. Indicates the associative mapping function;

[0069] Step S3: Model resource allocation probabilistics by constructing a multi-layer feedforward neural network, as shown below:

[0070] π θ ( X t )=[ p 1 , p 2 , ... , p N ] ;

[0071] in, This represents a multilayer feedforward neural network. This represents the task allocation matrix for the current time step. This represents the set of weights in a multilayer feedforward neural network. Indicates the total number of tasks. Indicates task The probability of being assigned to the optimal resource unit;

[0072] Step S4: Task allocation. For each task in the historical task dataset, select the resource unit allocation action. The formula used is as follows:

[0073] ;

[0074] in, Indicates at time step Always on the task The allocation action performed;

[0075] After performing the assignment action, update the task assignment matrix as follows:

[0076] ;

[0077] in, This represents the updated task assignment matrix. This represents the function for updating the allocation matrix;

[0078] Step S5: Multi-step reward aggregation, definition Step-by-step cumulative rewards :

[0079] ;

[0080] in, Indicates the aggregation step size. Indicates the cumulative step index. Indicates immediate earnings value. Represents the payoff function. express The proportion allocated to resource units, This represents the task resource allocation state matrix. Represents the vector of tradeoff coefficients. Indicates dynamic environmental factors;

[0081] Step S6: Self-evolutionary policy update. Update the parameters of the multilayer feedforward neural network based on the aggregated reward, as shown below:

[0082] ;

[0083] in, Indicates the learning rate;

[0084] Step S7: Output the optimal resource scheduling strategy. Set the maximum number of iterations and repeat steps S2 to S6 until the maximum number of iterations is reached. Extract the parameters of the current multi-layer feedforward neural network and design an optimization generator based on a deep policy mapping function. Generate the optimal resource scheduling strategy through a parameter-driven policy reasoning mechanism.

[0085] In this embodiment, in a comprehensive commercial complex with a building area of ​​approximately 300,000 square meters, there are 100 schedulable resource units for equipment resources and a total of 50 management tasks.

[0086] Collect property operation historical data for the past three months:

[0087] Each task Include:

[0088] Required resources c i ∈[1,10] ;

[0089] Average completion time t i ∈[5,60] minute;

[0090] The current resource status is represented as follows:

[0091] R 0 =[ r 1 , r 2 ,…, r 100 ] ;

[0092] The matching degree between the task requirement vector and the resource capability vector is calculated using cosine similarity, resulting in an initial allocation matrix of 50×100. Each element in the matrix x ij ∈[0,1] Indicates task Allocate resources The initial probability value;

[0093] The probability of allocating air conditioning temperature control tasks to air conditioning unit resources is 0.83;

[0094] The probability of assigning lighting scheduling tasks to the lighting control unit is 0.91;

[0095] Construct a multi-layer feedforward neural network with a three-layer structure:

[0096] Input layer dimension: 5000, unfolded into a 50×100 matrix;

[0097] Hidden layer 1: 256 neurons, ReLU activated;

[0098] Hidden layer 2: 64 neurons, Sigmoid activation;

[0099] Output layer: 50 neurons, corresponding to the optimal allocation probability for each task. [ p 1 , p 2 , ... , p 50 ] .

[0100] Model output results:

[0101] π θ ( X t )=[0.93,0.87,0.71, ... ,0.80] ;

[0102] The model assumes that the first task is most likely to be assigned to the air conditioning unit, with a probability of 0.93.

[0103] The fifth task, "Elevator Operation Monitoring," is currently assigned to the resource unit "Elevator Group Control System." After the assignment action is updated, the probability of other assignments in the related rows decreases by approximately 20%.

[0104] Let the aggregation step size be... The immediate return function is defined as:

[0105] ;

[0106] The benefits consist of three parts: task completion rate contribution, device energy consumption optimization rate, and user comfort feedback score;

[0107] ;

[0108] Update network parameters using gradient ascent:

[0109] ;

[0110] Learning rate ;

[0111] Assuming a maximum number of iterations of 50, the system extracts the final parameters after convergence and maps them through a depth policy function. Generate a set of candidate strategies:

[0112] ;

[0113] Through the disturbance term Generate Enhancement Strategy Calculate the comprehensive return function:

[0114] ;

[0115] The strategy with the highest overall return was ultimately selected. .

[0116] Example 3, based on the above examples, in step S7, designs an optimization generator based on a deep policy mapping function, which generates the optimal resource scheduling policy through a parameter-driven policy reasoning mechanism, specifically including the following steps:

[0117] Step S71: Parameter encoding, encode the parameters of the current multilayer feedforward neural network into a parameter vector, and initialize the deep policy mapping function using a high-dimensional nonlinear mapping function;

[0118] Step S72: Policy candidate generation. A set of candidate policies is generated through the deep policy mapping function, represented as:

[0119] ;

[0120] ;

[0121] in, Indicates the first One candidate strategy, This represents the main policy vector based on the parameter vector; Represents the candidate policy perturbation vector;

[0122] Step S73: Candidate strategy evaluation, calculating the comprehensive payoff function for the candidate strategy set:

[0123] ;

[0124] in, This represents the overall return value of the candidate strategies. Indicates task throughput. Indicates the degree of satisfaction of the delay constraint. Indicates resource utilization rate, , and This is a vector of weighting coefficients;

[0125] Step S74: Strategy Enhancement. The comprehensive benefit values ​​of candidate strategies are sorted, and high-yield candidate strategies are selected. Enhanced strategies are generated based on the updated task allocation matrix and the fine-tuning function of the resource state. The formula used is as follows:

[0126] ;

[0127] in, This indicates the enhanced strategy;

[0128] Step S75: Optimal policy selection, combining the enhancement policies into a set of enhancement policies. The strategy with the highest overall return is selected from among them, using the following formula:

[0129] s opt = arg max s i ' ∈ S e [P s i ' · R s i ' ] ;

[0130] ;

[0131] in, This represents the return sensitivity coefficient. This represents the overall return value of the enhanced strategy. Indicating reinforcement strategy The probability of being selected as the optimal strategy. This represents the optimal resource scheduling strategy.

[0132] In this embodiment, the core code used is as follows:

[0133] import numpy as np

[0134] class SystemState:

[0135] def __init__(self, num_tasks=8, num_resources=4):

[0136] self.num_tasks = num_tasks

[0137] self.num_resources = num_resources

[0138] self.X_t = np.random.randint(0, num_resources, size=(num_tasks,))

[0139] self.R_t = np.random.uniform(0.5, 1.0, size=(num_resources,))

[0140] def parameter_encoding(model_params):

[0141] Θ = np.concatenate([p.flatten() for p in model_params])

[0142] F_Ω = lambda Θ: np.tanh(Θ)

[0143] return Θ, F_Ω

[0144] def generate_candidate_strategies(F_Ω, Θ, num_candidates=12, noise_scale=0.05):

[0145] base = F_Ω(Θ)

[0146] candidates = []

[0147] for _ in range(num_candidates):

[0148] ϵ_i = np.random.normal(0, noise_scale, size=base.shape)

[0149] s_i = base + ϵ_i

[0150] candidates.append(s_i)

[0151] return candidates

[0152] def evaluate_strategies(candidates, weights=(0.4, 0.3, 0.3)):

[0153] w1, w2, w3 = weights

[0154] results = []

[0155] for s in candidates:

[0156] T = np.clip(np.mean(s), 0, 1)

[0157] D = np.clip(1 - np.std(s), 0, 1)

[0158] U = np.clip(np.sum(np.abs(s)) / len(s), 0, 1)

[0159] R = w1*T + w2*D + w3*U

[0160] results.append((s, R, T, D, U))

[0161] return sorted(results, key=lambda x: x[1], reverse=True)

[0162] def enhance_strategies(top_candidates, X_t, R_t, learning_rate=0.1):

[0163] enhanced = []

[0164] for s, R, _, _, _ in top_candidates:

[0165] delta = learning_rate * (np.mean(X_t) + np.mean(R_t)) *np.sign(np.random.randn(*s.shape))

[0166] s_prime = s + delta

[0167] enhanced.append(s_prime)

[0168] return enhanced

[0169] def select_optimal_strategy(enhanced_strategies, weights=(0.4, 0.3,0.3), beta=5.0):

[0170] evaluated = evaluate_strategies(enhanced_strategies, weights)

[0171] Rs = np.array([r[1] for r in evaluated])

[0172] P = np.exp(beta * Rs) / np.sum(np.exp(beta * Rs))

[0173] idx = np.argmax(P * Rs)

[0174] s_opt = evaluated[idx]

[0175] return s_opt, P

[0176] if __name__ == "__main__":

[0177] system = SystemState()

[0178] model_params = [np.random.randn(8, 4), np.random.randn(4, 1)]

[0179] Θ, F_Ω = parameter_encoding(model_params)

[0180] candidates = generate_candidate_strategies(F_Ω, Θ)

[0181] ranked = evaluate_strategies(candidates)

[0182] top_candidates = ranked[:4]

[0183] enhanced = enhance_strategies(top_candidates, system.X_t,system.R_t)

[0184] s_opt, prob = select_optimal_strategy(enhanced)

[0185] print("Optimal strategy overall return R = =", round(s_opt[1], 4))

[0186] print("Corresponding policy vector (partially truncated) =", np.round(s_opt[0][:8], 3)).

[0187] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0188] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0189] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent property management system based on artificial intelligence, characterized in that: It includes a scene perception fusion module, a semantic cognition modeling module, a self-evolutionary learning engine module, an intelligent control and execution module, and an interaction module, specifically including the following: The scene perception fusion module collects multi-dimensional property information in real time, performs feature decoupling and multi-source data fusion on the multi-dimensional property information, and obtains scene feature vectors. The semantic cognition modeling module performs semantic layer modeling on scene feature vectors, constructs a property semantic network based on knowledge graphs, and uses graph attention networks to embed and update complex semantic relationships to obtain the updated property semantic network. The self-evolutionary learning engine module collects historical task data sets, constructs a dynamic task allocation mechanism based on a multi-layer feedforward neural network, introduces a multi-step aggregated reward function, and performs adaptive optimization of parameters based on a gradient self-update strategy to obtain the optimal resource scheduling strategy. The intelligent control and execution module performs intelligent dynamic control of property equipment, energy consumption and environmental systems based on the optimal resource scheduling strategy; The interactive module constructs an intelligent human-machine collaborative interface.

2. The intelligent property management system based on artificial intelligence according to claim 1, characterized in that: The self-evolutionary learning engine module specifically includes the following steps: Step S1: Collect a set of historical task data, represented as Each of the tasks Includes the amount of resources required for execution and completion time Collect the current system resource status and represent it as ,in Indicates the first The availability of each resource unit; Step S2: Generate the initial task allocation matrix. Based on the historical task data set and the current resource status, generate the initial task allocation matrix using an association mapping function, as shown below: ； in, This represents the initial task assignment matrix. Indicates the associative mapping function; Step S3: Model resource allocation probabilistics by constructing a multi-layer feedforward neural network, as shown below: ； in, This represents a multilayer feedforward neural network. This represents the task allocation matrix for the current time step. This represents the set of weights in a multilayer feedforward neural network. Indicates the total number of tasks. Indicates task The probability of being assigned to the optimal resource unit; Step S4: Task allocation. For each task in the historical task dataset, select the resource unit allocation action. The formula used is as follows: ； in, Indicates at time step Always on the task The allocation action performed; After performing the assignment action, update the task assignment matrix as follows: ； in, This represents the updated task assignment matrix. This represents the function for updating the allocation matrix; Step S5: Multi-step reward aggregation, definition Step-by-step cumulative rewards : ； in, Indicates the aggregation step size. Indicates the cumulative step index. Indicates immediate earnings value. Represents the payoff function. express The proportion allocated to resource units, This represents the task resource allocation state matrix. Represents the vector of tradeoff coefficients. Indicates dynamic environmental factors; Step S6: Self-evolutionary policy update. Update the parameters of the multilayer feedforward neural network based on the aggregated reward, as shown below: ； in, Indicates the learning rate; Step S7: Output the optimal resource scheduling strategy. Set the maximum number of iterations and repeat steps S2 to S6 until the maximum number of iterations is reached. Extract the parameters of the current multi-layer feedforward neural network and design an optimization generator based on a deep policy mapping function. Generate the optimal resource scheduling strategy through a parameter-driven policy reasoning mechanism.

3. The intelligent property management system based on artificial intelligence according to claim 2, characterized in that: Step S7: Design an optimization generator based on a deep policy mapping function to generate the optimal resource scheduling policy through a parameter-driven policy inference mechanism. This includes the following steps: Step S71: Parameter encoding, encode the parameters of the current multilayer feedforward neural network into a parameter vector, and initialize the deep policy mapping function using a high-dimensional nonlinear mapping function; Step S72: Policy candidate generation. A set of candidate policies is generated through the deep policy mapping function, represented as: ； ； in, Indicates the first One candidate strategy, This represents the main policy vector based on the parameter vector; Represents the candidate policy perturbation vector; Step S73: Candidate strategy evaluation, calculating the comprehensive payoff function for the candidate strategy set: ； in, This represents the overall return value of the candidate strategies. Indicates task throughput. Indicates the degree of satisfaction of the delay constraint. Indicates resource utilization rate, , and This is a vector of weighting coefficients; Step S74: Strategy Enhancement. The comprehensive benefit values ​​of candidate strategies are sorted, and high-yield candidate strategies are selected. Enhanced strategies are generated based on the updated task allocation matrix and the fine-tuning function of the resource state. The formula used is as follows: ； in, This indicates the enhanced strategy; Step S75: Optimal policy selection, combining the enhancement policies into a set of enhancement policies. The strategy with the highest overall return is selected from among them, using the following formula: ； ； in, This represents the return sensitivity coefficient. This represents the overall return value of the enhanced strategy. Indicating reinforcement strategy The probability of being selected as the optimal strategy. This represents the optimal resource scheduling strategy.

Images