Building resource collaborative optimization method and device, and electronic equipment

Abstract

The invention provides a building resource collaborative optimization method, a building resource collaborative optimization device and electronic equipment, which effectively solve the negative influence of the dynamic change of a project site on the accuracy, the efficiency and the like of the existing building resource collaborative optimization system. The method comprises the following steps: the input layer obtains various heterogeneous data of the constructional engineering of resources to be optimized, and constructs state vectors corresponding to the various heterogeneous data respectively; inputting the state vector to the hidden layer, and controlling an action space to extract low-order resource features from the state vector based on a nonlinear association relationship; fusing the low-order resource features to obtain high-order resource features, and screening out target high-order resource features from the high-order resource features to obtain multiple actions based on the target high-order resource features; and receiving a plurality of actions, selecting an optimal action from the plurality of actions, and realizing collaborative optimization of the to-be-optimized resources of the to-be-constructed project based on the optimal optimization strategy corresponding to the optimal action.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building resource optimization, and in particular, to a building resource collaborative optimization method and device and electronic equipment. BACKGROUND

[0002] To realize the design collaboration and resource management of a building project, the patent document with the application number: 202410504591.X in the prior art provides an integrated system for collaborative design and engineering management, which achieves certain effects, but has certain defects: Firstly, the data integration and sharing module is only used for the close integration and information sharing between the collaborative design module and the engineering management module, and realizes the coordination consistency of the functional modules in the system, but does not mention any technical design about real-time data interaction (such as real-time data transmission protocol, dynamic synchronization mechanism), nor does it explicitly mention the frequency or timeliness standard of data interaction; Secondly, the data sources of each module are only limited to the static data of the design end and the management end: the data of the collaborative design module comes from the design file (such as the component size and pipeline layout in the BIM model), and the data of the engineering management module comes from the manual operation and preset information (such as task allocation and process node in the project schedule), and the collection of dynamic data of the engineering site is completely not involved - neither the access interface of the IoT device (such as personnel bracelet, equipment sensor, RFID material label) nor the technical scheme of real-time collection of data such as “personnel load, equipment utilization rate, and material inventory dynamics”.

[0003] The limitation of such data interaction mechanism leads to the inability of the system to capture the dynamic changes of the engineering site (such as the real-time operation state of the tower crane and the consumption speed of the steel reinforcement inventory): for example, when the utilization rate of the tower crane in construction decreases due to sudden failure, the system cannot obtain the dynamic data in time, and the collaborative design module still promotes the design according to the preset resource state, which may generate a design scheme that depends on the operation of the faulty tower crane; the engineering management module also cannot adjust the resource allocation based on the dynamic data, which is easy to cause the contradiction of “design scheme and site resource mismatch”, affecting the project promotion efficiency. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a building resource collaborative optimization method, device and electronic equipment, which effectively solves the negative impact of the dynamic changes of the engineering site on the accuracy and efficiency of the existing building resource collaborative optimization system.

[0005] In a first aspect, the embodiments of the present application provide a building resource collaborative optimization method, which is suitable for an optimization system including an input layer, a hidden layer and an output layer, and the method comprises: The input layer acquires multiple heterogeneous data of a building project of a to-be-optimized resource, and constructs a state vector based on the multiple heterogeneous data; different heterogeneous data corresponds to different building physical quantities; at least one building physical quantity has a nonlinear correlation with another building physical quantity; The state vector is input to the hidden layer, and the hidden layer extracts low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity; High-order resource features are obtained by fusing the low-order resource features, and target high-order resource features are selected from the high-order resource features, so as to obtain multiple actions based on the target high-order resource features; the multiple actions have been verified for compliance; The output layer receives the multiple actions transmitted by the hidden layer, and selects an optimal action from the multiple actions based on expected returns, so as to realize the collaborative optimization of the to-be-optimized resource of the to-be-built project based on an optimal optimization strategy corresponding to the optimal action.

[0006] In combination with the first aspect, the second possible implementation manner of the first aspect is provided, and the optimization system further iterates based on the following steps: Priorities of multiple training samples in a training sample set are defined, so as to control the training samples to train the optimization system based on the priorities; the training samples are dynamically increased; In the training process, a dynamic epsilon-greedy strategy and a soft update strategy are controlled to iteratively train the optimization system.

[0007] In combination with the first aspect, the second possible implementation manner of the first aspect is provided, and the fusion of the low-order resource features to obtain high-order resource features, and the selection of target high-order resource features from the high-order resource features include: Based on the type of the high-order resource features, a corresponding weight matrix and a bias vector are set to extract the target high-order resource features; In combination with a sample-driven update mechanism and a construction period stage adaptation strategy, the target high-order resource features are optimized.

[0008] In combination with the first aspect, the third possible implementation manner of the first aspect is provided, and the setting of the corresponding weight matrix and the bias vector to extract the target high-order resource features includes: For neurons in a second hidden layer on which the target high-order resource features depend, corresponding weight values are allocated; Based on the activation difficulty of the neurons, a bias value of the bias vector is adjusted, so as to extract the target high-order resource features based on the bias value and the weight values.

[0009] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein the hidden layer extracts low-order resource features from the state vector based on the nonlinear correlation, including: The state vector is linearly weighted by the target neuron set in the first hidden layer to obtain the linear weighted result; The linear weighted result is processed by a pre-defined function network to obtain an activation value, and the low-order resource feature is obtained based on the activation value.

[0010] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein selecting the optimal action from the plurality of actions based on expected benefits includes: Based on multiple neurons in the output layer, calculate the expected reward value corresponding to each of the various actions; Based on the aforementioned benefit values, the degree of impact of the various actions on the construction project is determined, and the optimal action is selected.

[0011] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein, after selecting the optimal action, the following is included: Feedback is provided on the action result after the optimal action is executed, and multiple core indicator data in the action result are quantified; The reward value for the best action is calculated based on the core indicator data, and training samples are generated based on the reward value to update the optimization system.

[0012] In conjunction with the first aspect, this application provides a seventh possible implementation of the first aspect, wherein constructing the state vector based on the multiple heterogeneous data includes: The state space is invoked to call the normalization network pre-set for each type of building physical quantity in order to normalize the data of the corresponding building physical quantity; The normalized data for each type of building physical quantity are collected to construct the state vector corresponding to the various heterogeneous data.

[0013] Secondly, embodiments of this application provide a building resource collaborative optimization device, applicable to optimization systems, the optimization system including an input layer, a hidden layer, and an output layer, the device comprising: The acquisition module is used by the input layer to acquire various heterogeneous data of the building project of the resource to be optimized, and to construct a state vector based on the various heterogeneous data; different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a non-linear correlation with another building physical quantity; The input module is used to input the state vector into the hidden layer and control the hidden layer to extract low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity. The fusion module is used to fuse the low-level resource features to obtain high-level resource features, and to filter out target high-level resource features from the high-level resource features to obtain multiple actions based on the target high-level resource features; all of the multiple actions have undergone compliance verification. An optimization module is used to receive multiple actions transmitted by the hidden layer from the output layer, and select the best action from the multiple actions based on the expected benefits, so as to achieve collaborative optimization of the resources to be optimized in the project to be built based on the best optimization strategy corresponding to the best action.

[0014] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of any one of the building resource collaborative optimization methods are performed.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any one of the building resource collaborative optimization methods.

[0016] This application provides a method for collaborative optimization of building resources, applicable to an optimization system. The optimization system includes an input layer, a hidden layer, and an output layer. The method first involves the input layer acquiring various heterogeneous data of the building project to be optimized, and constructing a state vector based on this data. Different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a non-linear correlation with another. Next, the state vector is input to the hidden layer, which extracts low-order resource features from the state vector based on the non-linear correlation. Each low-order resource feature corresponds to a single building physical quantity. Then, the low-order resource features are fused to obtain high-order resource features, and target high-order resource features are selected from these features to generate various actions. All actions undergo compliance verification. Finally, the output layer receives the various actions transmitted from the hidden layer and selects the optimal action based on expected returns. The optimal optimization strategy corresponding to the optimal action is then used to achieve collaborative optimization of the building project's resources. Based on the above methods, collaborative optimization of building resources and engineering projects is achieved in construction projects with resources to be optimized. The state space enables dynamic reflection of changes on the engineering site through state data, transforming design and resource data from static separation to real-time linkage. This provides a precise environmental perception foundation for subsequent intelligent decision-making. Furthermore, based on the hidden layer, it achieves the effect of efficiently exploring feasible solutions in the early stage of the project and stably executing the best optimization strategy in the middle and later stages, completely getting rid of the limitations of existing technologies that rely on manual decision-making and have poor adaptability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a collaborative optimization method for building resources provided in an embodiment of this application is shown. Figure 2 A schematic diagram of the process for obtaining target high-order resource features provided in an embodiment of this application is shown; Figure 3 A schematic diagram illustrating the process of selecting the optimal action provided in an embodiment of this application is shown; Figure 4 This paper shows a structural block diagram of a building resource collaborative optimization device provided in an embodiment of this application; Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0020] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0022] Existing technologies have limitations in the data interaction mechanisms used for collaborative optimization of construction projects and resources. This results in the system being unable to capture dynamic changes at the construction site, thus failing to achieve resource optimization and affecting project progress efficiency.

[0023] Based on this, the present application provides a method, apparatus, and electronic device for collaborative optimization of building resources, which are described below through embodiments.

[0024] Example 1 To facilitate understanding of this embodiment, a detailed description of a building resource collaborative optimization method disclosed in this application embodiment will be provided first. For example... Figure 1 The flowchart shown is a method for collaborative optimization of building resources. This application provides a method for collaborative optimization of building resources, applicable to optimization systems. The optimization system includes an input layer, a hidden layer, and an output layer. The method includes: S101, The input layer acquires various heterogeneous data of the building project of the resource to be optimized, and constructs a state vector based on the various heterogeneous data; different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a non-linear correlation with another building physical quantity; S102. Input the state vector to the hidden layer, and control the hidden layer to extract low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity; S103. The low-level resource features are fused to obtain high-level resource features, and target high-level resource features are selected from the high-level resource features to obtain multiple actions based on the target high-level resource features; all of the multiple actions have undergone compliance verification. S104. The output layer receives multiple actions transmitted by the hidden layer, and selects the best action from the multiple actions based on the expected benefit, so as to achieve the collaborative optimization of the resources to be optimized in the project to be built based on the best optimization strategy corresponding to the best action.

[0025] In this application, the optimization system is based on an improved DQN network to implement the optimization method. That is, the traditional DQN network is improved to obtain an improved DQN network that is more suitable for the construction project domain of the resource to be optimized. The optimization system includes an input layer, a hidden layer, and an output layer. In the engineering environment, the optimal strategy is learned through the quantitative interaction of state (S), action (A), and reward (R), forming a three-element closed loop of perception, action, and evaluation. The improved DQN network contains three core steps: state space, action space, and reward function. These three are the basic elements of reinforcement learning and are logically related to each other and indispensable. The relationship logic among the three is: the state space defines what the environment is, the action space defines what changes can be made, and the reward function defines whether the changes are good or bad. Together, they form the minimum closed loop of perceiving the environment, executing actions, and evaluating the effect, providing the improved DQN network with a standardized input that is perceptible, decision-making, and evaluable.

[0026] In the specific implementation of the optimization method of this application, one embodiment is as follows: the optimization system further iterates continuously based on the following steps: A1. Define the priority of multiple training samples in the training sample set, so as to control the training of the optimization system based on the priority; the training samples are dynamically added. A2. During the training process, the dynamic ε-greedy strategy and the soft update strategy are used in conjunction to iterate the optimization system.

[0027] In steps A1-A2, this application proposes an internally improved DQN decision engine, comprising two categories: network structure design and three major improvement mechanisms. The network structure serves as the basic architecture carrying core functions, while the improvement mechanisms adapt to the associated logic of engineering scenarios. The network structure is the fundamental carrier of decision-making, including an input layer, a first hidden layer, a second hidden layer, and an output layer. The three improvement mechanisms are scenario adaptation optimization methods: priority experience replay, dynamic ε-greedy, and soft update. These mechanisms improve the traditional DQN network to better adapt to the collaborative optimization of building engineering and building resources. The three improvement mechanisms are as follows: (1) Prioritized experience replay: Improving the utilization rate of effective samples: The optimization system of this application has a corresponding training sample set, which has the characteristic of dynamic addition, that is, the training samples in the training sample set are dynamically increased. Traditional DQN networks randomly sample samples, and the proportion of valid samples with compliant and high rewards in construction projects is less than 10%. This mechanism defines sample priority through TD error (temporal difference error) and prioritizes training high-value samples. A TD error parameter is set to measure the Q-value prediction deviation; the larger the deviation, the higher the sample value: δ_i=r_i+γ·max_aQ(s'_i,a;θ')-Q(s_i,a_i;θ), where r_i is the reward value of sample i, s'_i is the new state after the action is executed, θ is the main network parameter, θ' is the target network parameter, γ=0.9 is the discount factor, and the future reward weight in construction projects is lower than the current one; the sample priority assignment is set as: P(i)=|δ_i |+ε, where ε=0.01, to avoid a priority of 0 and ensure that all samples have a chance of being sampled; set the sampling probability calculation: P(i|D)=P(i)^α / Σ_jP(j)^α, α=0.6, to control the degree of influence of priority, when α=0, it degenerates into random sampling; set the importance sampling weight to correct the bias caused by priority sampling, w_i=[1 / (N·P(i|D))]^β / max_jw_j, β=0.4, β→1 in the later stage of training to reduce the influence of bias.

[0028] (2) Dynamic ε-greedy strategy: Balancing exploration and exploitation: Traditional DQN networks have a fixed exploration rate ε, such as ε=0.1. In construction projects, exploration is needed in the early stages and stable execution is needed in the later stages. Therefore, this application allows ε to dynamically decay with the project progress. Project progress is defined as: τ=current completed time / total time (τ∈[0,1]); ε decay formula: ε(τ)=0.1+0.7·max(0,1-τ / 0.5); early stage of the project (τ≤0.1): ε≈0.8 (80% probability of exploring new actions to find feasible solutions); mid-stage of the project (0.1<τ<0.5): ε decreases from 0.8 to 0.1 (gradually shifting to utilizing the optimal strategy); late stage of the project (τ≥0.5): ε=0.1 (only 10% probability of exploration to avoid design changes affecting the finalization).

[0029] (3) Target network soft update: Ensure stable parameter updates: Traditional DQN networks undergo "hard updates" every N steps (θ'←θ), which can easily lead to sudden parameter changes and decision fluctuations. Therefore, this application adopts "soft updates" to gradually synchronize the main network parameters to the target network. The soft update formula is: θ'=τ·θ+(1-τ)·θ', where τ=0.001 is the update rate, which controls the speed at which the target network follows the main network. The smaller τ is, the more stable the network. The loss function is defined to minimize the prediction error of the main network: L(θ)=E[w_i·(y_i-Q(s_i,a_i;θ))²], where y_i=r_i+γ·max_aQ(s'_i,a;θ') is the target Q value, and w_i is the importance sampling weight.

[0030] In step S101, since the state is a digital snapshot of the construction environment, it needs to cover three types of heterogeneous data: design parameters, resources, and schedule. Based on the construction project of the resource to be optimized, this application sets a state space in the input layer. This state space is used to normalize and model various heterogeneous data. The input layer includes 8 neurons, corresponding to the 8 dimensions of the state vector S, and has no activation function. The input layer acquires various heterogeneous data of the construction project of the resource to be optimized. All of this heterogeneous data is collected from the construction project of the resource to be optimized, including data from equipment sensors, RFID material tags, and on-site work reports. The system, etc., includes various heterogeneous data, including design, resource, and schedule data, where schedule refers to construction progress. Different heterogeneous data correspond to different building physical quantities, and the number of building physical quantities corresponding to each type of heterogeneous data is different. For example, heterogeneous data of design parameters includes three building physical quantities: beam cross-sectional area, pipeline spacing, and concrete strength grade; heterogeneous data of resources includes three building physical quantities: worker load rate, tower crane utilization rate, and steel bar inventory rate; and heterogeneous data of schedule includes two building physical quantities: process completion rate and schedule deviation rate. Based on these various heterogeneous data, a state vector is constructed, which is represented as S∈ (8-dimensional, covering 3 types of core parameters), the expression is: S=[s1,s2,s3,s4,s5,s6,s7,s8], where: S1-S3 (design parameters), S4-S6 (resource status), S7-S8 (progress), and at least one of the above building physical quantities has a non-linear correlation with another building physical quantity. That is, the beam cross-sectional area, pipeline spacing, concrete strength grade, worker load rate, tower crane utilization rate, and steel bar inventory rate will more or less have a non-linear impact on the two building physical quantities of process completion rate and schedule deviation rate.

[0031] The data for beam cross-sectional area comes from the BIM model, reflecting the structural design safety (the larger the area, the stronger the load-bearing capacity); the data for pipeline spacing comes from the digital twin model, reflecting the risk of mechanical and electrical installation conflicts (the smaller the spacing, the higher the risk of conflict); the data for concrete strength grade comes from material testing reports + BIM model, reflecting the material performance compliance (the higher the strength, the better the durability); the data for worker load rate comes from IoT personnel wristbands, reflecting the efficiency of human resource utilization (excessive load rate leads to fatigue); the data for tower crane utilization rate comes from equipment sensors, reflecting the utilization efficiency of large equipment (excessive utilization rate leads to idleness); the data for rebar inventory rate comes from RFID material tags, reflecting the material supply guarantee (excessive inventory rate leads to supply disruption); the data for process completion rate comes from the on-site work reporting system, reflecting the progress of construction (low completion rate indicates delays); and the data for schedule deviation rate comes from the digital twin schedule module, reflecting schedule risk warnings (high deviation rate indicates high risk).

[0032] In a specific implementation of step S101, one embodiment is as follows: constructing a state vector based on the various heterogeneous data includes: S1011. Call the normalization network pre-set for each type of building physical quantity in the state space to normalize the data of the corresponding building physical quantity; S1012. Collect the normalized data of each type of building physical quantity to construct the state vector corresponding to the various heterogeneous data.

[0033] In steps S1011-S1012, the normalization network pre-set for each type of building physical quantity in the state space is invoked to normalize the corresponding building physical quantity data. The normalization networks corresponding to beam cross-sectional area, pipeline spacing, concrete strength grade, worker load rate, tower crane utilization rate, steel bar inventory rate, process completion rate, and construction period deviation rate are represented by formulas (1)-(8): S1=(A-A_min) / (A_max-A_min)(1); S2=(d-d_min) / (d_max-d_min)(2); S3=(f_c-f_cmin) / (f_cmax-f_cmin)(3); S4 = L / 100 (L is the actual load percentage) (4); S5 = U / 100 (U is the actual utilization percentage) (5); S6=min(K / 100,1.0) (K is the inventory adequacy ratio) (6); S7 = P / 100 (P is the actual percentage completed) (7); S8=min(ΔT / 50,1.0) (ΔT is the percentage of deviation) (8); Where S represents the normalized building physical quantity, and A, d, f_c, L, U, K, and P are the data corresponding to the building physical quantities other than the schedule deviation rate, respectively. The state vector corresponding to the various heterogeneous data is then constructed by collecting the normalized data for each type of building physical quantity, which is S∈ , S=[s1,s2,s3,s4,s5,s6,s7,s8].

[0034] Before normalization, it is necessary to ensure that there are no missing data for the above eight building physical quantities. If there are missing data, they are filled in according to the data completion rules. If a sensor failure causes a certain dimension of data to be missing, such as when the tower crane sensor is offline and the S5 cannot collect data, the exponential smoothing method with the formula x_t=α·x_{t-1}+(1-α)·x_{t-2} is used to fill in the missing data, where α=0.7 (recent data has a higher weight to ensure that the filled value is close to the actual value).

[0035] In step S102, the input layer of this application will assign the state vector S∈ S = [s1, s2, s3, s4, s5, s6, s7, s8] is input into the hidden layer. Since actions are adjustment instructions or optimization strategies for the agent, the hidden layer categorizes the actions into two types—design adjustment and resource allocation—through a pre-set action space, corresponding to a 6-dimensional discrete vector A = [a1, a2, a3, a4, a5, a6] ∈ {0, 1}. 6(One-hot encoding), a1, a2, a3, a4, a5, and a6 represent three types of design adjustment actions and three types of resource scheduling actions, respectively. The action space is strongly correlated with the state space; that is, the state space covers design, resources, and schedule, while the action space corresponds to design adjustments and resource allocation. This ensures that actions can accurately change key dimensions in the state, such as changing "tower crane utilization" through "tower crane scheduling." The hidden layer includes a first hidden layer and a second hidden layer. The first hidden layer contains 32 neurons, which are the core computational units. The ReLU activation function is used to handle nonlinear relationships, simulating the non-proportional correlation between state and action in engineering. The weight matrix W1∈ ^(8×32), where 8 corresponds to the 8-dimensional state vector of the input layer (S=[s1-s8], covering three types of parameters: design, resources, and schedule), 32 corresponds to the 32 neurons of the first hidden layer, used to quantify the signal transmission strength between the input layer and the first hidden layer, and the bias vector b1∈ ^32, each element corresponds to one neuron in the first hidden layer, used to adjust the activation threshold of the neuron to avoid calculation deviation caused by input data offset. The first hidden layer controls the hidden layer to extract low-order resource features from the state vector based on the nonlinear correlation. The low-order resource features correspond to a single building physical quantity. The 32 neurons avoid insufficient feature extraction due to too few neurons, such as the inability to capture the implicit correlation between tower crane utilization and pipeline layout, and also avoid computational redundancy due to too many neurons.

[0036] In a specific implementation of step S102, one embodiment is as follows: the hidden layer extracts low-order resource features from the state vector based on the nonlinear correlation, including: S1021. The state vector is linearly weighted by the target neuron set in the first hidden layer to obtain a linear weighted result; S1022. Call a preset function network to process the linear weighted result to obtain an activation value, and obtain the low-order resource feature based on the activation value.

[0037] In steps S1021-S1022, the state vector is linearly weighted using the target neuron set in the first hidden layer to obtain a linearly weighted result. Specifically, a weight matrix W1∈ Each element W1(i,j) in the ^(8×32) matrix, where i=1-8, corresponds to the 8 dimensions of the state space; j=1-32, corresponds to the 32 neurons in the first hidden layer, representing the influence weight of the i-th state dimension of the state space on the j-th neuron of the first hidden layer. For example, W1(5,10)=0.8, meaning the weight of crane utilization S5 on neuron 10 is much greater than W1(3,10)=0.1, meaning the weight of concrete strength S3 on neuron 10. This indicates that neuron 10 focuses more on extracting features related to crane utilization, adapting to the decision-making needs of resource actions such as crane scheduling a5. In subsequent network iterations and optimizations, the weight matrix W1 will be updated according to the training results of the training samples. For example, when the crane utilization is low and the reward value of samples with delayed processes is high, the weight of the corresponding dimension and neuron will be increased, so that the feature extraction accuracy gradually improves with the accumulation of engineering data. The preset function network processes the linearly weighted result to obtain the activation value, i.e., the ReLU activation function ReLU(x)=max(0,x). That is, when the input x≤0, the output is 0 and the neuron is not activated; when x>0, the output is equal to x and the neuron is activated according to the input intensity. If a certain state vector, such as concrete strength grade S3, is not related to the current action, such as tower crane scheduling a5, the ReLU function will set the calculation result corresponding to the state vector to 0 to avoid invalid features interfering with the decision. For example, the change in concrete strength has no effect on the Q value of tower crane scheduling and does not need to be included in the calculation, so low-order resource features are obtained. If a certain state vector has a non-linear relationship with the action, such as when the tower crane utilization rate S5 increases from 60% to 85%, the Q value of tower crane scheduling a5 first increases rapidly and then tends to level off. The ReLU function can simulate this trend by changing the activation intensity to ensure that the calculation result fits the actual law of engineering.

[0038] Where the bias vector b1∈ Each element b1(j) of ^32, j=1-32, corresponds to 32 neurons in the first hidden layer. It sets the activation threshold for each neuron, addressing the data bias problem in construction engineering. For example, the normalized value of worker load rate S4 in engineering is usually concentrated between 0.6 and 0.9, meaning that workers are mostly in a medium to full load state. If calculated solely by the weight matrix W1, the neurons may not respond adequately to special scenarios such as low load (S4<0.5). By setting b1(j)=-0.2, the bias of neuron j, the activation condition of the neuron can be adjusted. Originally, the sum of W1(i,j)×s_i ≥ 0 was required for activation, but now only the sum ≥ 0.2 is needed for activation. This ensures that features of special states such as low load can be effectively extracted, avoiding decision blind spots caused by data distribution bias.

[0039] In step S103, the first hidden layer inputs multiple low-order resource features into the second hidden layer. The second hidden layer comprises 16 neurons, which are the core computational units of the second hidden layer. It also employs the ReLU activation function to further fuse higher-order features, deepening the extraction of multi-dimensional state relationships in the project, such as the correlation between tower crane idleness and process delays. The weight matrix W2∈ ^(32×16), where 32 corresponds to the low-order resource feature a1 output by the 32 neurons of the first hidden layer after activation by the ReLU activation function, 16 corresponds to the 16 neurons of the second hidden layer, used to quantify the intensity of high-order feature transmission between the first and second hidden layers, and the bias vector b2∈ ^16, meaning each element corresponds to one neuron in the second hidden layer, is used to adjust the activation threshold of neurons during the extraction of high-order resource features. This avoids deviations in high-order resource association extraction caused by the offset of low-order resource features. The 16 neurons in the second hidden layer fuse the low-order resource features to obtain high-order resource features. The number of 16 neurons is less than the 32 in the first hidden layer. Essentially, it is a refined fusion of 32-dimensional low-order resource features. This avoids redundancy in high-order resource features due to an excessive number of neurons, such as repeatedly extracting similar associations and increasing decision-making time. It also avoids omission of association features due to a small number of neurons, such as failing to capture the indirect association between low steel inventory and concrete strength adjustment. This ensures that high-order features are fully extracted under the efficiency requirement of decision-making time ≤ 10 minutes. Target high-order resource features are then selected from these features to obtain multiple candidate actions. That is, multiple candidate actions for building resource collaborative optimization are obtained through the first and second hidden layers included in the hidden layer, and these multiple candidate actions are input to the output layer.

[0040] Each action is bound to engineering specifications through mathematical formulas to ensure compliance. That is, all actions obtained based on the first and second hidden layers of the hidden layer have been verified for compliance. For example, if pipeline spacing optimization is required, then A=[0,1,0,0,0,0]. The system automatically reads the constraint formula corresponding to a2 to verify compliance. As shown in Table 1, the operation variables, constraint formulas and specification basis of each action are as follows: if the current pipeline spacing d0=180mm, then Δd≥200-180=20mm. The system automatically adjusts the pipeline spacing to 200mm (minimum compliance value) to avoid violations. The action space is placed in the hidden layer.

[0041] Table 1. Operational variables, constraint formulas, and standard references for actions.

[0042] The high-order resource features are not manually defined, such as the combination of idle tower cranes and delayed processes, but are automatically learned by the network during training. The core principle is to adjust the hidden layer weights through feedback from the reward function, making neurons sensitive to feature combinations that enhance rewards. The specific process is as follows: Initial stage: Weight matrices W1 and W2 are randomly assigned values. The responses of hidden layer neurons to any combination of features are irregular. For example, the responses to idle tower crane + advanced process and idle tower crane + delayed process are the same.

[0043] Training Phase: When the DQN network outputs a crane dispatching action, if the reward value R increases (e.g., from -2.28 to -1.0), it indicates that the action is effective under the "current state combination". Through backpropagation, the network automatically adjusts the weights: increasing the weights of the neurons corresponding to crane idleness (s5) and process lag (s7) in the first hidden layer to make them easier to activate; increasing the connection weights between the crane idleness neurons and process lag neurons in the second hidden layer to make their combination easier to capture.

[0044] Convergence Phase: After multiple rounds of training, specific neurons in the second hidden layer will strongly activate the combination of idle tower cranes and delayed processes, i.e., output high values, while the activation of the combination of idle tower cranes and advanced processes is weaker. This selective response to effective feature combinations is the high-order feature automatically determined by the DQN network.

[0045] For example, when S5=0.6 (tower crane idle) and S7=0.7 (process delayed) in the state vector, the neuron corresponding to the higher-order resource feature in the second hidden layer outputs 0.9248 (as in the previous example), which drives the output layer to assign a higher Q value to the tower crane scheduling action, thereby achieving a decision-making effect of accurate response to multi-dimensional correlation.

[0046] In the specific implementation of step S103, one embodiment is as follows: Figure 2 As shown, the process of fusing the low-level resource features to obtain high-level resource features, and then filtering out target high-level resource features from the high-level resource features, includes: S1031. Based on the type of the higher-order resource features, set the corresponding weight matrix and bias vector to extract the target higher-order resource features; S1032. Combining the sample-driven update mechanism and the project duration adaptation strategy, optimize the target high-order resource features.

[0047] In steps S1031-S1032, since some high-order resource features obtained from combinations of low-order resource features have no practical decision-making value, the second hidden layer sets corresponding weight matrices and bias vectors based on the type of the high-order resource features to extract target high-order resource features. The types are those with practical decision-making value and those without. For example, the direct correlation between beam cross-sectional area and rebar inventory rate has no impact on the expected return value Q of pipeline spacing optimization. The ReLU activation function will set the calculation results of such invalid high-order correlations to 0 to avoid interfering with effective correlation features, such as the correlation between tower crane utilization rate and process completion rate. Furthermore, the relationship between multi-state correlations and action returns is often non-linear. For example, when tower crane utilization rate decreases (i.e., idle time intensifies) and process completion rate decreases (i.e., delays intensify), the increase in Q value of tower crane scheduling actions is not the sum of single-state changes, but rather an exponential increase. In this case, the return on scheduling tower cranes is significantly reduced. Furthermore, the ReLU activation function can simulate this pattern through the nonlinear change of activation intensity, ensuring the matching degree between high-order correlation features and action rewards. Combined with the sample-driven update mechanism and the project duration adaptation strategy, the target high-order resource features are optimized. Specifically, when an action corresponding to a certain type of target high-order resource feature is executed and a high reward is obtained, such as the risk of project delay decreasing after scheduling tower cranes, the sample containing state S, action A, and reward R will be given high priority (calculated through TD error) and will be used to update W2 in the network iterative optimization, enhancing the extraction sensitivity of this feature. Combined with the dynamic ε-greedy strategy, in the early stage of construction project (τ≤0.1, exploration stage), the extraction range of target high-order resource features will be expanded to ensure coverage of more potential correlations; in the later stage of project (τ≥0.5, utilization stage), the focus will be on core features such as steel bar inventory-beam cross section and concrete strength-project duration, improving decision stability.

[0048] In the specific implementation of step S1032, one embodiment is as follows: setting the corresponding weight matrix and bias vector to extract the target high-order resource features includes; S10321. Assign corresponding weight values ​​to neurons in the second hidden layer that depend on the target high-order resource features; S10322. Adjust the bias value of the bias vector based on the activation difficulty of the neuron, so as to extract the target high-order resource features based on the bias value and the weight value.

[0049] In steps S10321-S10322, the weight matrix set in the second hidden layer serves to precisely define the priority of multi-state associations. The weight of the l-th neuron in the second hidden layer is W2(5,l)=0.7, which corresponds to the low-order resource feature of the tower crane utilization rate in the first hidden layer. The weight of the l-th neuron in the second hidden layer is also W2(7,l)=0.6, which corresponds to the low-order resource feature of the process completion rate in the first hidden layer. This weight is significantly higher than the weights of other low-order resource features, such as W2(1,l)=0.1, which corresponds to the weight of the beam cross-sectional area. This indicates that the neuron specifically extracts the high-order resource features associated with tower crane idleness and process delays. Therefore, a higher weight is assigned to the corresponding neuron from this perspective, providing a core basis for calculating the Q-value of actions such as tower crane scheduling and personnel dispatch. During network iterative optimization, the weight matrix W2 in the second hidden layer is updated based on sample feedback. For example, after executing a tower crane scheduling action on a sample with tower crane idleness and process delays, if the R-value significantly increases (e.g., the risk of project delay decreases by 50%), the system will increase the weight of the tower crane utilization rate in W2. The weights of neurons associated with low-order and high-order resource features, such as utilization rate and process completion rate, are used to make the extraction of high-order resource features in similar scenarios more accurate. The bias value of the bias vector is adjusted based on the activation difficulty of the neurons. For example, in construction engineering, the tower crane utilization rate shows that the tower crane is in a medium utilization state, and the process completion rate shows that the corresponding process is in a normal progress state. If only W2×a1 is used for calculation, the second hidden layer may not be able to activate the neurons in calculating high-order resource features in special scenarios such as tower crane idleness + process delay due to the input value being too low. By setting the bias b2(l)=-0.3 of the l-th neuron of the second hidden layer, the activation threshold of the neuron can be reduced, ensuring that the target high-order resource features such as tower crane idleness + process delay, which are low probability but high impact, can be effectively extracted. This avoids the risk of not being able to extract due to the activation threshold being too high, such as the risk of tower crane idleness aggravating process delay and failure to schedule tower crane in time. Therefore, the target high-order resource features are extracted based on the bias value and the weight value, thereby ensuring the accuracy of obtaining the target high-order resource features.

[0050] This application takes a slow-paced engineering project (process cycle 1-3 days) and high-risk project (violations can easily lead to safety accidents) as an example. In the state vector, S5=0.6 represents the idle degree of the tower crane, and S7=0.7 represents the process lag degree. The first hidden layer uses the ReLU activation function to transform the linear relationship in the original state vector into a non-linear expression. For example, the tower crane utilization rate is not necessarily better the higher it is; exceeding 85% will lead to the risk of equipment overload. This lays the foundation for subsequent feature combination. The weight matrix W1∈ ^(8×32), bias vector b1∈ ^32. The second hidden layer establishes the correlation between tower crane idleness and process lag characteristics. Engineering decision-making often relies on this multi-dimensional correlation, rather than a single dimension. For example, knowing only that the tower crane is idle (s5=0.6) or only that the process is lag-prone (s7=0.7) is insufficient for decision-making. However, when both conditions are met simultaneously, 'scheduling the tower crane' is the optimal solution. The ReLU activation function is also used to further integrate higher-order resource features (such as the correlation between tower crane idleness and process lag). The weight matrix W2∈ ^(32×16), bias vector b2∈ ^16.

[0051] In step S104, after receiving the multiple actions, the output layer calculates the expected reward value Q of the multiple actions. The target neuron set in the output layer selects the best action from the multiple actions based on the expected reward value Q. According to the best action, the optimal optimization strategy is generated according to the strategy generation template. The optimal optimization strategy can fully inherit the corresponding effect of the best action and has applicability, thereby ensuring that the optimal optimization strategy can be directly executed to achieve the collaborative optimization of the resources to be optimized in the project to be built. After execution, the DQN network is iterated and optimized according to the best action to ensure the accuracy of the DQN network.

[0052] In the specific implementation of step S104, one embodiment is as follows: Figure 3 As shown, selecting the optimal action from the multiple actions based on expected returns includes: S1041. Based on the multiple neurons in the output layer, calculate the expected revenue value corresponding to each of the various actions; S1042. Based on the revenue value, determine the degree of impact of the various actions on the construction project, and select the optimal action.

[0053] In steps S1041-S1042, the optimization system, after determining that multiple actions are compliant based on various constraints in the action space, uses multiple neurons in the output layer (specifically six, directly determined by the dimension of the action space, i.e., six discrete actions). Each neuron outputs the Q-value of one action to ensure the DQN network can provide an independent expected return assessment for each candidate action. It can also select the optimal action based on a dynamic ε-greedy strategy. Since no activation function is set in the output layer, the obtained Q-value can be positive or negative. Based on the expected return value, the impact of the multiple actions on the construction project is determined, and the optimal action is selected. If the Q-value is positive, it represents the expected return of the action, such as reducing idle resources through resource scheduling; if the Q-value is negative, it represents the expected loss. The weight matrix W3∈ ^(16×6), bias vector b3∈ ^6.

[0054] The reward function in the neurons of the output layer aims to minimize design change costs, resource idle rate, and project delay risk. It is converted into maximizing reward through a negative sign. The weight of each objective is determined by the project manager using the Analytic Hierarchy Process (AHP). The total weight is set to 1. For example, the weight of the project time limit for key projects is ω3=0.5, and for cost-sensitive projects it is ω1=0.4. The calculation of the reward function depends on the changes in the state space, such as the change in the project time deviation rate before and after the action is executed. At the same time, the reward result will guide the selection of the action space. For example, tower crane scheduling actions with high reward values ​​will be preferentially selected by the improved DQN network. The target values ​​corresponding to multiple objectives are quantified. The quantification formulas, physical meanings and calculation examples of each indicator are shown in Table 2, with a total project budget of 10 million yuan and total resources of 10 people × 10 days = 100 person-days as an example. The weights and target values ​​of multiple targets are fused according to the pre-set fusion method, i.e., the reward formula: R = -(ω1·Cchange + ω2·Ridle + ω3·Rdelay), to calculate the R value of the various actions, where the R value is the reward value. ω = [ω1, ω2, ω3] = [0.6, 0.3, 0.1] is the weight vector. Through AHP consistency verification: consistency ratio CR = 0.033 < 0.1, ensuring reasonable weights. For example, substituting the weights ω = [0.6, 0.3, 0.1], the baseline state before action execution is R = -0.6 × 0.5 - 0.3 × 6 - 0.1 × 1.8 = -2.28; the optimization effect after actual action execution is Cchange = 0.2, Ridle = 3, Rdelay = 0.8, then R = -0.6 × 0.2 - 0.3 × 3 - 0.1 × 0.8 = -1.0. Comparing the dynamic... An increase in the reward value R before and after execution indicates a better action effect. Based on a pre-set optimization objective containing multiple goals, the R values ​​of the various actions are calculated. Based on the R values, the output layer calculates the Q value of the corresponding action, Q = W3·a2 + b3 (Q is a 6-dimensional vector, each element corresponds to the expected benefit of one action, the larger the value, the better the action). Thus, the best action is selected based on the Q value. For example, if the state vector S = [0.45, 0.33, 0.43, 0.8, 0.6, 0.5, 0.7, 0.2] (S5 = 0.6: tower crane idle; S7 = 0.7: process lag), the calculated Q = [0.2, 0.5, 0.3, 0.8, 1.5872, 0.6], among which "tower crane scheduling (a5)" has the highest Q value (1.5872), which is the current best action.

[0055] Table 2. Quantitative formulas, physical meanings, and calculation diagrams for each indicator.

[0056] The optimal optimization strategy is a compliant and optimal engineering action execution plan. This is the most direct output of the closed-loop process, that is, the specific operational plan implemented on the construction site after state perception and decision calculation. It is divided into two categories: Resource optimization solutions include: shifting tower cranes from non-critical processes to main structure pouring processes (utilization rate controlled at 80%), assigning 3 idle workers to rebar tying processes (load rate not exceeding 90%), and prioritizing rebar supply for the No. 1 building pouring process (meeting 2 days' construction needs). All of these solutions clearly define the operation targets, execution parameters, and compliance boundaries. Design adjustment schemes: such as optimizing pipeline spacing from 180mm to 200mm (complying with GB50981 standard), and slightly adjusting beam cross-section width by 10% (from 250mm to 275mm, with load-bearing capacity meeting the standard), simultaneously generating BIM model change documents and construction handover drawings.

[0057] The core features of these solutions are intelligent decision-making and engineering compliance. They select the optimal solution based on the Q-value evaluation of DQN, and ensure that there are no safety and compliance risks through the constraint verification of the action space, which can directly guide on-site construction.

[0058] In a specific implementation of step S1042, one embodiment is as follows: after selecting the optimal action, the process includes: S10421. Feedback on the action result generated after the optimal action is executed, and quantify multiple core indicator data in the action result; S10422. Calculate the reward value of the best action based on the core indicator data, and generate training samples based on the reward value to update the optimization system.

[0059] In steps S10421-S10422, after the optimal optimization strategy is executed, the optimization system collects the action results within 24 hours after the execution of the optimal action, and quantifies multiple core indicator data in the action results, including design change cost Cchange, resource idle rate Ridle, and schedule delay risk Rdelay. The system calculates the design change cost Cchange, resource idle rate Ridle, and schedule delay risk Rdelay according to a preset weighted formula to obtain the reward value Rt for the optimal action. Based on the reward value, training samples (St, At, Rt, S{t+1}) are generated and stored in the priority experience replay pool. High-priority samples are periodically extracted from the experience replay pool, and the network parameters (weight matrices W1-W3, bias vectors b1-b3) of the improved DQN decision engine are updated using gradient descent to improve the accuracy of subsequent action decisions. The closed-loop process of state perception, action decision-making, and reward function in this application generates precise quantitative feedback on the action result. Specifically, it calculates the required core indicator data based on a pre-defined reward function to quantify the action result corresponding to the core indicator, including: Design change cost data: such as material waste of 0.8 million yuan due to pipeline optimization and labor adjustment cost of 0.2 million yuan, which is ultimately quantified as Cchange = 0.2; Resource utilization effect data: For example, after tower crane scheduling, the idle rate decreased from 6% to 3%, and the personnel load rate was optimized from 75% to 85%, which is quantified as R_idle = 3; Data on the impact on construction schedule: For example, if the delay time of the main process is reduced from 3 days to 1 day, the delay risk level is reduced from level 3 to level 2, which is quantified as R_delay = 0.8; These data not only serve as a scoring basis for the effectiveness of the best action, but also form a traceable engineering effect ledger, providing data support for subsequent project management. This application will also combine the state vector, the best action, and the action result across the entire chain of data to obtain training samples, and encapsulate them into standardized training samples for reinforcement learning (S_t, A_t, R_t, S_{t+1}). The value of the training samples is defined based on the core indicator data, thereby defining the priority of the training samples, and the training samples are added to the training sample set. The standardized training sample (S_t, A_t, R_t, S_{t+1}) includes the state vector S_t before the action is executed, such as S5=0.6 (tower crane idle), S7=0.7 (process delayed), the executed action A_t (e.g., [0,0,0,0,1,0] tower crane scheduling), the quantified reward value R_t=-1.0, and the new state vector S_{t+1} after execution, such as S5=0.8 (tower crane utilization improved), S7=0.85 (process progress). These training samples are the core fuel for DQN network iteration, especially high-value samples after priority experience replay screening, such as samples with significant reward value improvement, which can accurately optimize network weights and make the model better understand the decision logic of engineering scenarios.

[0060] The optimization system of this application forms a complete closed loop by connecting state perception, action decision-making, constraint verification, reward calculation, and network optimization through a time axis. The specific implementation process of the optimization system, using the scenario of tower crane idleness and process delay as an example, is shown in Table 3. The data flow is as follows: State vector S_t → Action A_t → New state S_{t+1} → Reward R_t → Sample (S_t, A_t, R_t, S_{t+1}) → Network parameter update → Calculation of the optimized Q-value, forming a preliminary data closed loop. The reward R_t determines the TD error δ, δ determines the sample priority, and the priority affects the sampling probability. The sampled samples drive the update of weight matrices W1, W2, and W3 through the loss function L(θ), ultimately making the Q-value more accurately reflect the action reward, forming a complete mathematical closed loop.

[0061] From the perspective of data collaboration, this application transforms three types of heterogeneous data into an 8-dimensional vector within a unified range through state space normalization modeling. This not only eliminates data format barriers but also ensures that state data can dynamically reflect changes on the engineering site through real-time acquisition by IoT devices and a missing data completion mechanism. This transforms design and resource data from "static separation" to "real-time linkage," providing a precise environmental perception foundation for subsequent intelligent decision-making.

[0062] From the perspective of decision-making compliance and adaptability, this application binds mathematical constraints to the action space, such as beam cross-section adjustment range ≤ ±10% and tower crane utilization rate not exceeding 85%, so that the action generation has compliance verification by default, and violations can be avoided without manual intervention. At the same time, the improved DQN decision engine, through dynamic ε-greedy strategy (adjusting the exploration rate with the progress of the project), priority experience replay (prioritizing the learning of high-value samples), and target network soft update (avoiding parameter mutation), enables decision-making to efficiently explore feasible solutions in the early stage of the project and stably execute the optimal strategy in the middle and later stages, completely getting rid of the limitations of existing technologies that "decision-making depends on manual intervention and has poor adaptability".

[0063] From the perspective of continuous optimization, the virtual-real linkage closed loop constructed in this application calculates reward values ​​by collecting cost, idle time, and schedule data after action execution 24 hours a day. It stores the "state-action-reward-new state" sample into the experience pool and drives the parameter iteration of the DQN network, enabling the system to continuously learn from actual engineering results. For example, after handling the scenario of "material shortage causing schedule delay" for the first time, when encountering similar problems in the future, the decision will more accurately balance "design adjustment" and "resource scheduling", realizing the transformation from "passive response" to "proactive adaptation". This is a continuous optimization capability that existing technologies cannot achieve due to the lack of feedback links.

[0064] Table 3 shows the implementation process of the optimization system, taking the scenarios of tower crane idleness and process delays as an example.

[0065] Example 2 This application also provides a building resource collaborative optimization device, such as... Figure 4 The diagram shows a block diagram of a building resource collaborative optimization device. The function implemented by this device corresponds to the steps of executing a building resource collaborative optimization method on a terminal device as described above. This device can be understood as a server component including a processor. The building resource collaborative optimization device described in this application is suitable for optimization systems, which include an input layer, a hidden layer, and an output layer. The device includes: The acquisition module 401 is used to acquire various heterogeneous data of the building project of the resource to be optimized by the input layer, and construct a state vector based on the various heterogeneous data; different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a non-linear correlation with another building physical quantity; Input module 402 is used to input the state vector to the hidden layer and control the hidden layer to extract low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity; The fusion module 403 is used to fuse the low-level resource features to obtain high-level resource features, and to filter out target high-level resource features from the high-level resource features to obtain multiple actions based on the target high-level resource features; all of the multiple actions have undergone compliance verification. The optimization module 404 is used to receive multiple actions transmitted by the hidden layer from the output layer, and select the best action from the multiple actions based on the expected benefits, so as to achieve the collaborative optimization of the resources to be optimized in the project to be built based on the best optimization strategy corresponding to the best action.

[0066] In one feasible implementation, the device further includes: A definition module is used to define the priority of multiple training samples in the training sample set, so as to control the training of the optimization system based on the priority; the training samples are dynamically added. The update module is used to collect the training results generated by the training samples and control the dynamic ε-greedy strategy and the soft update strategy to iterate the optimization system.

[0067] In one feasible implementation, the fusion module includes: The setting module is used to set the corresponding weight matrix and bias vector based on the type of the higher-order resource feature in order to extract the target higher-order resource feature. The first optimization module is used to optimize the target high-order resource features by combining the sample-driven update mechanism and the project duration adaptation strategy.

[0068] In one feasible implementation, the fusion module also includes: The allocation module is used to allocate corresponding weight values ​​to neurons in the second hidden layer that depend on the target high-order resource features. An extraction module is used to adjust the bias value of the bias vector based on the activation difficulty of the neuron, so as to extract the target high-order resource features based on the bias value and the weight value.

[0069] In one feasible implementation, the input module includes: The weighting module is used to linearly weight the state vector through the target neuron set in the first hidden layer to obtain a linear weighted result; The first calling module is used to call a preset function network to process the linear weighted result to obtain an activation value, and obtain the low-order resource feature based on the activation value.

[0070] In one feasible implementation, the optimization module includes: The calculation module is used to calculate the expected reward value corresponding to each of the various actions based on multiple neurons in the output layer; The selection module is used to determine the impact of the various actions on the construction project based on the benefit value, and select the best action.

[0071] In one feasible implementation, the optimization module further includes: The quantization module is used to provide feedback on the action result data generated after the optimal action is executed, and to quantify multiple core indicator data in the action result data. The combination module is used to calculate the reward value of the best action based on the core indicator data, and generate training samples based on the reward value to update the optimization system.

[0072] In one feasible implementation, the acquisition module includes: The second calling module is used to call the normalization network pre-set for each type of building physical quantity in the state space to normalize the data of the corresponding building physical quantity. The construction module is used to collect the normalized data of each type of building physical quantity in order to construct the state vector corresponding to the various heterogeneous data.

[0073] Example 3 This application also provides an electronic device, such as Figure 5 As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the building resource collaborative optimization methods described above are executed.

[0074] Example 4 This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the building resource collaborative optimization methods described herein.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0076] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0078] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0079] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A building resource collaborative optimization method, characterized in that, The method is suitable for an optimization system comprising an input layer, a hidden layer and an output layer, and the method comprises: The input layer acquires multiple heterogeneous data of a construction project of a resource to be optimized, and constructs a state vector based on the multiple heterogeneous data; different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a nonlinear correlation with another building physical quantity; The state vector is input to the hidden layer, and the hidden layer extracts low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity; High-order resource features are obtained by fusing the low-order resource features, and target high-order resource features are screened from the high-order resource features, so as to obtain multiple actions based on the target high-order resource features; the multiple actions have all been verified for compliance; The output layer receives the multiple actions transmitted by the hidden layer, and selects an optimal action from the multiple actions based on expected returns, so as to realize the collaborative optimization of the resource to be optimized of the construction project based on an optimal optimization strategy corresponding to the optimal action.

2. The method of claim 1, wherein, The optimization system also iterates based on the following steps: Priorities of multiple training samples in a training sample set are defined, so as to control the training samples to train the optimization system based on the priorities; the training samples are dynamically increased; In the process of training, a dynamic ε-greedy strategy and a soft update strategy are controlled to iteratively train the optimization system.

3. The method of claim 1, wherein, The fusion of the low-order resource features to obtain high-order resource features, and the screening of target high-order resource features from the high-order resource features, comprise: Based on the types of the high-order resource features, corresponding weight matrices and bias vectors are set to extract target high-order resource features; In combination with a sample-driven update mechanism and a construction period stage adaptation strategy, the target high-order resource features are optimized.

4. The method of claim 3, wherein, The setting of corresponding weight matrices and bias vectors to extract target high-order resource features comprises: For neurons in a second hidden layer on which the target high-order resource features depend, corresponding weight values are assigned; Based on the activation difficulty of the neurons, the bias values of the bias vectors are adjusted, so as to extract target high-order resource features based on the bias values and the weight values.

5. The method of claim 1, wherein, The hidden layer extracts low-order resource features from the state vector based on the nonlinear correlation, comprising: Linear weighting of the state vector is performed by target neurons set in the first hidden layer to obtain a linear weighting result; A preset function network processes the linear weighting result to obtain an activation value, and the low-order resource features are obtained based on the activation value.

6. The method of claim 1, wherein, The selection of an optimal action from the multiple actions based on expected returns comprises: Based on multiple neurons in the output layer, expected return values corresponding to the multiple actions are calculated; Based on the expected return values, the influence degrees of the multiple actions on the construction project are determined, and an optimal action is selected.

7. The method of claim 1, wherein, After the optimal action is selected, comprising: An action result generated after the optimal action is executed is fed back, and multiple core index data in the action result are quantified; The reward value of the optimal action is calculated based on the core index data, and a training sample is generated based on the reward value to update the optimization system.

8. The method of claim 1, wherein, The state vector is constructed based on the plurality of heterogeneous data, including: The normalization network set in advance for each building physical quantity in the state space is called to normalize the data of the corresponding building physical quantity; The normalized data of each building physical quantity is collected to construct the state vector corresponding to the plurality of heterogeneous data.

9. A building resource collaborative optimization apparatus, characterized by, The optimization system includes an input layer, a hidden layer and an output layer, and the device includes: The acquisition module acquires the plurality of heterogeneous data of the building project of the to-be-optimized resource in the input layer, and constructs a state vector based on the plurality of heterogeneous data; different heterogeneous data correspond to different building physical quantities; at least one building physical quantity has a nonlinear correlation with another building physical quantity; The input module inputs the state vector to the hidden layer, and controls the hidden layer to extract low-order resource features from the state vector based on the nonlinear correlation; the low-order resource features correspond to a single building physical quantity; The fusion module fuses the low-order resource features to obtain high-order resource features, and screens target high-order resource features from the high-order resource features to obtain a plurality of actions based on the target high-order resource features; the plurality of actions have been verified for compliance; The optimization module receives the plurality of actions transmitted by the hidden layer in the output layer, selects an optimal action based on the expected return from the plurality of actions, and realizes the collaborative optimization of the to-be-optimized resource of the to-be-built project based on the optimal optimization strategy corresponding to the optimal action.

10. An electronic device, comprising: It includes: A processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions are executed by the processor to execute the steps of the building resource collaborative optimization method in any one of claims 1 to 8.

Images