BIM data processing and building construction quality evaluation method based on artificial intelligence

Abstract

The invention discloses a BIM data processing and building construction quality evaluation method based on artificial intelligence, and relates to the technical field of building electrical digital data processing. The method comprises the following steps of: acquiring building data, manually marking, uniformly preprocessing the building data, expanding the building construction data, extracting features of the building construction data based on a neural network of nonlinear halo optimization, evaluating construction quality, feeding back and optimizing. According to the method, the authenticity and task adaptability of the generated data are realized through feature matching and task correlation loss, the complexity of potential features is enhanced in combination with a variational auto-encoder, and the problem of insufficient training samples in building construction tasks is solved. By adopting a neural network based on nonlinear halo optimization and a feedback mechanism based on neighborhood parameters, the breadth and depth of feature extraction are improved, the optimization capability of the network is enhanced by using chaotic disturbance, the problem of local optimum is avoided, and the modeling requirements of high-dimensional and nonlinear characteristics of construction data are met.

Description

Technical Field

[0001] The present invention relates to the technical field of building electrical digital data processing, and particularly to a BIM data processing and building construction quality evaluation method based on artificial intelligence. Background Art

[0002] The building construction process is complex and diverse, involving multiple data types and information sources, including BIM models, Internet of Things sensor data, and construction document records. These data are heterogeneous in terms of source, format, and content. At the same time, the construction site environment changes dynamically, and problems such as incomplete data and complex feature associations occur frequently. Traditional construction quality evaluation methods usually rely on empirical judgment or analysis based on a small amount of historical data, and cannot make full use of multi-source real-time data during the construction process, resulting in insufficient evaluation accuracy. In addition, the cost of data acquisition and annotation required for model training is relatively high, and insufficient training samples will directly affect the generalization ability and prediction reliability of the model.

[0003] In the prior art, a Chinese invention patent with the publication number CN119167792A discloses a method, system, device and medium for adaptively calculating building cooling and heating loads, belonging to the technical field of load prediction, including: collecting image data and electricity consumption data of a building, and after preprocessing, extracting time series features from the electricity consumption data; using edge detection, feature point recognition and semantic segmentation technologies to extract image features from the image data; through multi-view geometry technology and surface reconstruction method, using the image features to generate a three-dimensional point cloud, and using Revit software to convert the three-dimensional model into a BIM model; calculating solar radiation and internal heat sources based on the electricity consumption data and the power of internal equipment in the building, and using random forest and neural network algorithms to train a cooling and heating load prediction model; adopting online learning and grid search technologies to dynamically adjust and optimize the cooling and heating load prediction model, and inputting real-time electricity consumption data and environmental parameters into the cooling and heating load prediction model to generate a building cooling and heating load prediction result. A Chinese invention patent with the publication number CN119128752A discloses a method for analyzing tunnel monitoring data based on long-term monitoring data, belonging to the field of construction engineering; this method first obtains the long-term monitoring data of the target tunnel in history, and divides the long-term monitoring data into: continuous data and classified data; uses the interpolation algorithm adopted to generate a continuous data change curve corresponding to the continuous data; the prediction of continuous data includes methods of time series analysis, regression analysis and machine learning; the analysis of classified data is divided into observing whether there are abnormalities in the classified data and the prediction of classified data; verifying the collected classified data through neural network technology; predicting the classified data through deep learning technology; dividing the tunnel monitoring data into continuous data and classified data, and respectively giving the analysis methods of continuous data and classified data, and giving the methods of processing different tunnel monitoring data by using different interpolation methods and machine learning and other methods. A Chinese invention patent with the publication number CN118761859A discloses a building energy consumption prediction and control system based on machine learning, specifically related to the technical field of building energy consumption management. Through a data integration and processing module, building and environmental data are comprehensively collected and standardized to ensure the quality and consistency of the input data. Through a deep feature analysis module, key variables affecting energy consumption can be effectively identified and extracted. Using multi-dimensional scaling and principal component analysis technologies, deep-level features in the data are fully mined. The key variable network module introduces a graph neural network to construct a dynamic interaction network between key variables, realizing the modeling and analysis of complex interaction relationships between variables.

[0004] The following drawbacks exist in the prior art: 1. In the tasks of BIM data processing and building construction quality assessment, the prior art has limited capabilities in processing multi-source data and cannot effectively support the unified processing of BIM models, Internet of Things data, and construction documents, resulting in low data utilization; 2. In the tasks of BIM data processing and building construction quality assessment, traditional models rely on limited real construction data and cannot solve the problems of insufficient model accuracy and task adaptability caused by insufficient training data; 3. In the tasks of BIM data processing and building construction quality assessment, existing neural network methods are difficult to dynamically capture the complex interactions between the characteristics of building construction data, and the feature extraction is not comprehensive enough, which limits the accurate assessment of construction quality; 4. In the tasks of BIM data processing and building construction quality assessment, many existing methods only achieve static assessment of construction quality, fail to provide the ability to optimize the plan and update the BIM model, and cannot form an efficient feedback mechanism for the construction process. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the above-mentioned drawbacks of the prior art and provide a BIM data processing and building construction quality assessment method based on artificial intelligence.

[0006] The technical solution adopted to solve the above technical problem is: A BIM data processing and building construction quality assessment method based on artificial intelligence, including the following steps:

[0007] S1. Collect building data and manually label the collected building data. The labeling categories include: excellent construction quality, construction quality to be improved, construction abnormal events, construction delay risks, and unqualified materials;

[0008] S2. Perform unified preprocessing on the building data;

[0009] S3. Generate samples based on a generative adversarial network for feature matching to achieve the expansion of building construction data;

[0010] S4. Extract features of building construction data based on a neural network optimized by non-linear halation;

[0011] S5. Generate a comprehensive quality assessment report through the predicted categories output by the building construction quality assessment model, identify the deviations in the building construction process, quantitatively analyze the deviations, and judge their severity;

[0012] S6. Automatically generate an optimization plan according to the evaluation results, including suggestions such as construction process adjustment and resource reallocation, provide detailed improvement priorities and implementation steps, and feedback the problems identified in the evaluation to the BIM model to update the model to reflect the current construction status.

[0013] Further, the attributes of the building data in S1 include the spatial relationship feature R in the BIM model a , the construction stage timestamp D a , the temperature data T collected by the sensor a , the humidity data H collected by the sensor a , the acceleration V collected by the vibration monitoring device a , the text content feature C in the construction record a , the material type identifier M a , the semantic label L of the construction log a , the quality inspection result Q a , the geographical location coordinates S of the construction area a .

[0014] Further, the sources of the building data in S1 include: BIM model input, Internet of Things data collection, and construction document integration.

[0015] Further, the preprocessing in S2 includes data cleaning and formatting, data completion and fusion, and time series data processing.

[0016] Further, the method for generating samples based on the feature-matching generative adversarial network in S3 includes the following steps:

[0017] S301, initialize the generator and discriminator of the generative adversarial network. The generator converts the random noise and task features into building construction data samples through multi-layer non-linear mapping. The discriminator discriminates the authenticity and task feature matching degree of the input samples, and optimizes their weight parameters through interactive confrontation, so as to gradually improve the quality and diversity of the generated building construction data. In order to ensure the diversity of the generated building construction data, the noise vector is processed in the way of variational autoencoder, and then a more complex and task-demand-compliant latent representation is obtained, expressed as:

[0018]

[0019] In the formula, Z c is the noise vector encoded by the variational autoencoder, and f VAE () is the encoding function of the variational autoencoder, which characterizes how to map the real building construction data samples into vectors in the latent space, is the real building construction data sample, Θ VAE is the parameter of the variational autoencoder, and is the training parameter, which is obtained by training through the gradient descent method;

[0020] The generator generates building construction data according to the noise vector encoded by the variational autoencoder and the task features, expressed as:

[0021]

[0022] In the formula, is the generated building construction data sample, G c () is the mapping function of the generator, T c is the task feature information, representing the associated features of the task, is the weight of the generator, representing the parameters that are continuously updated during network training;

[0023] S302. Define the loss function during the training process of the generator. The adversarial loss part measures the game between the generator and the discriminator in terms of authenticity discrimination. The feature matching loss part measures the difference between the generated samples and the real samples in the multi-layer feature space, ensuring that the generated building construction data is consistent with the real building construction data samples in terms of fine-grained features. The loss function of the generator is used to maximize the misjudgment probability of the discriminator for the generated building construction data. Finally, considering both the adversarial loss and the feature matching loss as the total loss function during the training process of the generator, it is expressed as:

[0024]

[0025] In the formula, is the adversarial loss of the generator, is the expected value, Z c is the noise vector, T c is the task feature, represents the expected value when the input is the noise vector and the task feature, D c () is the mapping function of the discriminator, representing the estimated probability that the input sample is real, is the function of the discriminator;

[0026] Use a multi-layer feature extractor to measure the difference between the generated samples and the real samples. The calculation method of the feature matching loss is expressed as:

[0027]

[0028] In the formula, L fm is the feature matching loss, L cea is the number of feature extraction layers of the generator, Φ l () is the l-th layer feature extraction function of the generator. The feature extraction function can be any one of the Sigmoid activation function, the convolution kernel function, or the ReLU activation function. ∥∥ is the L2 norm, representing the difference between the two in the feature space of this layer;

[0029] Considering both the adversarial loss and the feature matching loss, the calculation method of the total loss function of the generator is expressed as:

[0030]

[0031] In the formula, is the total loss of the generator, and λ f is the weight hyperparameter for feature matching, which characterizes the degree of attention given to different loss terms in the comprehensive optimization. λ f is set to 0.2;

[0032] S303. Define the loss function in the discriminator training process. The discriminator judges the authenticity of the samples and the matching of task features, and optimizes its weights by maximizing the discrimination between judging the real construction data samples as real and the generated construction data as fake. A dual discriminative loss is adopted. The discriminative adversarial loss ensures that the discriminator can distinguish real and generated samples, and the task relevance discriminative loss guides the discriminator to pay attention to task features, making the generated construction data more in line with actual needs. In the dual adversarial loss of the discriminator, the calculation method of the discriminative adversarial loss is expressed as:

[0033]

[0034] In the formula, is the discriminative adversarial loss, represents the expected value when the input is a real construction data sample, represents the expected value when the input is a generated construction data sample;

[0035] The calculation method of the task relevance discriminative loss is expressed as:

[0036]

[0037] In the formula, is the task relevance discriminative loss, and Φ er () is a preset classification function, and the preset classification function adopts any one of a pre-trained decision tree, random forest, or support vector machine;

[0038] The calculation method of the total loss function of the discriminator is expressed as:

[0039]

[0040] In the formula, is the total loss of the discriminator, and λ t is the weight hyperparameter of the task relevance loss, and λ t is set to 0.3;

[0041] S304. During the iterative training process, after receiving the noise vector and task feature information, the generator generates samples with real distribution characteristics and meeting the task requirements through multi-layer non-linear mapping of the network, and uses the feature matching loss to improve the similarity between the generated samples and the real construction data samples in the high-dimensional space, so as to ensure the accuracy and diversity of the construction data; the discriminator discriminates the authenticity and feature matching of the input real samples and generated samples, provides feedback signals for the generator, and through the adversarial loss and task relevance discrimination loss, the discriminator gradually improves its discrimination ability in the optimization, which in turn drives the generator to iteratively update in the direction of deceiving the discriminator and meeting the task feature requirements;

[0042] S305. During the iterative training process, calculate the respective losses of the generator and the discriminator according to the results output by the discriminator, and update the network parameters in the way of backpropagation. The generator continuously improves the sample diversity and task adaptability by minimizing its total loss, and the discriminator continuously enhances its discrimination ability by maximizing its discrimination degree of the real construction data samples and the generated construction data. The update method of the generator weights is expressed as:

[0043]

[0044] In the formula, is the weight of the generator at the (t + 1)-th iteration, is the weight of the generator at the t-th iteration, η is the learning rate of the generator g is set to 0.01, is the gradient of the total loss of the generator with respect to its weight;

[0045] The update method of the discriminator weights is expressed as:

[0046]

[0047] In the formula, is the weight of the discriminator at the (t + 1)-th iteration, is the weight of the discriminator at the t-th iteration, η is the learning rate of the discriminator d is set to 0.05, is the gradient of the total loss of the discriminator with respect to its weight;

[0048] S306. During the iterative training process, the generator adaptively adjusts the generation strategy during training, continuously optimizes the diversity and feature matching degree of the generated samples through the feedback of the discriminator to meet the comprehensive requirements of the distribution and label requirements of the construction data in practical applications. Each newly generated sample will be evaluated by the discriminator, and the generator updates its own parameters accordingly to ensure that the quality and diversity of the generated construction data are gradually improved in subsequent iterations;

[0049] In S307, the above process is iterated repeatedly until the parameters of the generator and the discriminator tend to be stable, the difference in the performance of the construction data generated by the generator and the real construction data samples in the discriminator is reduced to the extent that meets the application requirements, and they are consistent with the real construction data samples at the feature and label levels. Then, the generator can output high-quality construction training data that is rich, diverse, and task-adaptive. The preset stop iteration condition is that the total loss of the generator and the total loss of the discriminator reach the convergence state simultaneously.

[0050] Furthermore, the training process of the neural network algorithm based on non-linear halo optimization in S4 includes the following steps:

[0051] In S401, initialize the weights and biases of the neural network, expressed as:

[0052]

[0053] In the formula, is the initial weight of the neural network, is the initial bias of the neural network, ~ follows a specific distribution, and U(-0.5, 0.5) represents a random variable uniformly distributed within the interval [-0.5, 0.5];

[0054] Adopt a non-linear perturbation model to achieve the randomness of the initial perturbation amplitude, and the calculation method is expressed as:

[0055]

[0056] In the formula, is the initial perturbation amplitude, σ p is the perturbation amplitude scaling factor, σ p is set to 0.1, α p is the perturbation adjustment parameter, α p is set to 5;

[0057] In S402, in each iteration process, according to the current parameter state of the neural network, calculate the halo effect feedback of each parameter. The halo effect affects the update of the current parameter through the weighted sum of the neighborhood parameters. The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change of the loss function, enhancing the ability to model the feature relationship in the construction data during the feature extraction process, making the extracted features more practical. The calculation method of the halo effect feedback is expressed as:

[0058]

[0059] In the formula, is the halo effect feedback in the t-th iteration, N p represents the neighborhood of the current parameter, λ pqThe influence coefficient of the q-th parameter in the neighborhood on the current parameter's halo effect is the weight within the neighborhood at the t-th iteration is the dynamic feedback coefficient of the parameter at the t-th iteration. The dynamic feedback coefficient is adjusted according to the loss change sensitivity to adapt to the high-dimensional non-linear characteristics in the task of extracting construction data features, α pq is the interaction coefficient between the current parameter and the q-th parameter is the bias within the neighborhood at the t-th iteration

[0060] The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change of the loss function, and the calculation method is expressed as:

[0061]

[0062] In the formula, κ p is the sensitivity parameter of the feedback coefficient, κ p takes values between 0.5 and 1.5 is the change amount of the loss function of the neural network at the t-th iteration

[0063] The halo effect influence coefficient and the interaction coefficient control the interaction between different neurons, and can be dynamically adjusted according to the structural distance between neurons and the current state. To enhance the adaptability of the model, the calculation methods of the halo effect influence coefficient and the interaction coefficient are expressed as:

[0064]

[0065] In the formula, δ pq is the first non-linear function, δ′ pq is the second non-linear function, p ner is the index of the current neuron, q ner is the index of the q-th neuron, d max is the distance between the farthest neurons in the neural network, that is, the number of neurons in the neural network minus 1

[0066] S403. Based on the parameter state at the current iteration, a perturbation term is generated using a chaotic map. Each parameter is dynamically adjusted by adopting the perturbation term to enhance the diversity of optimization and push the neural network out of the local optimum. The chaotic map enhances the diversity of optimization and avoids the neural network falling into the local optimum. For the complex non-linear characteristics existing in the construction scenario, chaotic optimization provides a powerful exploration ability. The calculation method of the perturbation term is expressed as:

[0067]

[0068] In the formula, is the perturbation term of the neural network weight at the t-th iteration, αpfg is the first coefficient for controlling the intensity of chaotic perturbation, and f chaos () is the chaotic mapping function, and f chaos (, t) is the chaotic mapping function at the t-th iteration, and f chaos (, t - 1) is the chaotic mapping function at the (t - 1)-th iteration, is the weight of the neural network at the (t - 1)-th iteration, and int(t - 1) represents the index of the number of iterations at the (t - 1)-th iteration, is the perturbation term of the bias of the neural network at the t-th iteration, and β pfg is the second coefficient for controlling the intensity of chaotic perturbation, is the bias of the neural network at the (t - 1)-th iteration;

[0069] To enhance the complexity and nonlinearity of the perturbation, the chaotic mapping function is defined as a combined model based on the hyperbolic tangent function and the sine function, which is expressed as:

[0070]

[0071] In the formula, is the dynamic feedback coefficient of the parameter at the t-th iteration, and ω p is the frequency coefficient, and ω p is set to 5, int(t) represents the current number of iterations, and int(t - 1) represents the number of iterations in the previous time;

[0072] S404. After performing the halo effect feedback and chaotic perturbation, calculate the update increment of each parameter and add it to the current value of the parameter to form the update amount of the neural network parameter. The calculation method is expressed as:

[0073]

[0074]

[0075] In the formula, is the update amount of the weight of the neural network at the t-th iteration, and λ pgv is the feedback coefficient for weight update, and λ pgv is set to 0.1, is the perturbation amplitude at the t-th iteration, is the update amount of the bias of the neural network at the t-th iteration, and λ bev is the feedback coefficient for bias update, and λ bev is set to 0.3;

[0076] S405. Adaptively adjust the perturbation amplitude according to the change of the loss function in each iteration, so that the optimization process can adjust the intensity of the perturbation according to the convergence state of the loss, avoiding instability caused by excessive perturbation or inefficient optimization caused by too small perturbation. The adjustment calculation method of the perturbation amplitude is expressed as:

[0077]

[0078] In the formula, is the perturbation amplitude of the (t + 1)-th iteration, and γ ped is the perturbation amplitude adjustment factor, and γ ped is set to 0.95. is the change amount of the loss function of the neural network in the t-th iteration, represents the maximum value of the change amounts of the loss functions of the neural network in all previous iterations;

[0079] S406. Update the weights and biases of the neural network, and the update method is expressed as:

[0080]

[0081] In the formula, is the weight of the neural network in the t-th iteration, is the weight of the neural network in the (t + 1)-th iteration, is the bias of the neural network in the t-th iteration, is the bias of the neural network in the (t + 1)-th iteration;

[0082] S407. Repeat the above steps iteratively until the preset iteration stop condition is satisfied, which means the model training is completed.

[0083] The beneficial effects of the present invention are as follows: (1) In the BIM data processing and building construction quality assessment tasks of the present invention, by supporting the acquisition of various BIM formats and real-time sensor data, the problems of heterogeneous data sources and diverse formats in the building construction scenario are solved, and the distributed storage in JSON format meets the storage requirements of structured and unstructured data.

[0084] (2) In the BIM data processing and building construction quality assessment tasks of the present invention, the loss function is improved on the basis of the traditional generative adversarial network. The authenticity and task adaptability of the generated data are realized through feature matching and task relevance loss, and the complexity of the latent features is enhanced by combining the variational autoencoder, solving the problem of insufficient training samples in the building construction task.

[0085] (3) In the BIM data processing and building construction quality assessment tasks of the present invention, a neural network based on non-linear halo optimization is adopted. A feedback mechanism based on neighborhood parameters is used to enhance the breadth and depth of feature extraction, and chaotic perturbation is utilized to enhance the optimization ability of the network, avoiding the problem of falling into local optima and meeting the modeling requirements of the high-dimensional and non-linear characteristics of construction data.

[0086] (4) In the BIM data processing and building construction quality assessment tasks of the present invention, a quality assessment report is generated through the prediction results of the model, realizing quantitative analysis of construction process deviations. An automatic construction optimization plan is generated based on the assessment results, and the BIM model is dynamically updated to form a closed loop of data and construction tasks. Description of the Drawings

[0087] Figure 1 It is a comparison chart of the training losses of the halo optimization neural network and the conventional neural network of the present invention. Detailed Embodiments

[0088] In order to make the objectives, technical solutions and advantages of the present invention clearer, the following embodiments further elaborate on the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0089] The method for BIM data processing and building construction quality assessment based on artificial intelligence in this embodiment includes the following steps:

[0090] S1. Collect building data and perform manual annotation on the collected building data. The annotation categories include: excellent construction quality, construction quality to be improved, construction abnormal events, construction delay risks, and unqualified materials.

[0091] In this embodiment, the data of 5 categories are shown in the following table.

[0092]

[0093] Vectorize the data attributes in text format. The present invention uses the Word2Vec algorithm to perform vectorization processing on the text. The Word2Vec algorithm is a commonly used vectorization algorithm in the art. According to a preset large-scale corpus, these texts to be vectorized are scanned, and each word is represented as a one-hot encoded vector, and the dimension of the one-hot encoded vector is equal to the size of the vocabulary in the corpus.

[0094] The attributes of the building data include the spatial relationship feature R in the BIM model a , the construction stage timestamp D a , the temperature data T collected by the sensor a , the humidity data H collected by the sensor a , the acceleration V collected by the vibration monitoring devicea 、Text content feature C in construction records a 、Material type identifier M a 、Semantic tag L of construction log a 、Quality inspection result Q a 、Geographical location coordinates S of construction area a 。

[0095] The sources of building data include: BIM model input: The BIM model is Building Information Modeling, which supports common BIM formats such as Industry Foundation Classes (IFC), Revit, etc., and directly reads the BIM model files in the design and construction stages.

[0096] Internet of Things data collection: Collect real-time construction data through devices such as temperature and humidity sensors and vibration monitoring devices deployed at the construction site.

[0097] Construction document integration: Integrate structured and unstructured data such as construction logs, quality inspection records, and material acceptance forms to form a multi-modal data source.

[0098] The collected data is stored using a distributed storage system, and the format is a lightweight object storage based on JSON (JavaScript Object Notation), which ensures fast access to structured data and compatibility with unstructured data.

[0099] S2. Uniformly preprocess the building data.

[0100] The preprocessing includes data cleaning and formatting: Remove redundant information, correct outliers, and standardize the data format to meet the requirements of subsequent analysis.

[0101] Data completion and fusion: Use artificial intelligence techniques (such as Bayesian inference or KNN algorithm) to fill in missing data to ensure data integrity. Integrate data from BIM models, sensors, and documents to provide a panoramic view for analysis.

[0102] Time series data processing: Process the time series in sensor data, including smoothing, denoising, and anomaly detection, to ensure the temporal consistency of the data.

[0103] S3. Generate samples based on a feature-matching generative adversarial network to achieve building construction data augmentation.

[0104] Based on the traditional generative adversarial network, the loss function is improved so that the generated construction data can not only be authentic enough to "deceive" the discriminator, but also be highly consistent with the real construction data samples in terms of feature distribution and task relevance, forming diverse construction data with stronger adaptability to specific task requirements.

[0105] The method for sample generation based on the feature-matching generative adversarial network includes the following steps:

[0106] S301, Initialize the generator and discriminator of the generative adversarial network. The generator converts random noise and task features into construction data samples through multi-layer non-linear mapping. The discriminator discriminates the authenticity and task feature matching degree of the input samples, and optimizes their weight parameters through interactive confrontation, thereby gradually improving the quality and diversity of the generated construction data. To ensure the diversity of the generated construction data, the noise vector is processed using a variational autoencoder to obtain a more complex and task-demand-compliant latent representation, expressed as:

[0107]

[0108] In the formula, Z c is the noise vector encoded by the variational autoencoder, and f VAE () is the encoding function of the variational autoencoder, which characterizes how to map real construction data samples into vectors in the latent space. is the real construction data sample. The attributes of the construction data sample include the spatial relationship feature R a in the BIM model, the construction stage timestamp D a the temperature data T collected by the sensor, a the humidity data H collected by the sensor, a the acceleration V collected by the vibration monitoring device, a the text content feature C in the construction record, a the material type identifier M, a the semantic label L of the construction log, a the quality inspection result Q, a the geographical location coordinates S of the construction area, a . Θ VAE is the parameter of the variational autoencoder and is a training parameter obtained by training through the gradient descent method.

[0109] The generator generates construction data based on the noise vector encoded by the variational autoencoder and the task features, expressed as:

[0110]

[0111] In the formula, For the generated construction data sample, G c () is the mapping function of the generator, T c is the task feature information, representing the associated features of the task, such as the operating status and utilization rate of construction equipment, the role division and experience level of construction personnel, the safety risk level and protection measures at the construction site, the consumption of energy and resources such as electricity and water usage, the cost budget and material procurement records related to the construction task, the environmental features such as noise and air quality generated during the construction process, as well as the completion success rate of historical tasks, common problem records, customer demand characteristics, project priorities, weather conditions such as wind speed, precipitation probability, and geological conditions, etc. is the weight of the generator, representing the parameters that are continuously updated during the network training process.

[0112] S302, define the loss function during the generator training process. The adversarial loss part measures the game between the generator and the discriminator in terms of authenticity discrimination. The feature matching loss part measures the difference between the generated samples and the real samples in the multi-layer feature space, ensuring that the generated construction data is consistent with the real construction data samples in terms of fine-grained features. The loss function of the generator is used to maximize the misjudgment probability of the discriminator for the generated construction data. Finally, considering both the adversarial loss and the feature matching loss as the total loss function during the generator training process, which is expressed as:

[0113]

[0114] In the formula, is the adversarial loss of the generator, is the expected value, Z c is the noise vector, T c is the task feature, represents the expected value when the input is the noise vector and the task feature, D c () is the mapping function of the discriminator, representing the estimated probability that the input sample is real, is the function of the discriminator.

[0115] Use a multi-layer feature extractor to measure the difference between the generated samples and the real samples. The calculation method of the feature matching loss is expressed as:

[0116]

[0117] In the formula, L fm is the feature matching loss, L cea is the number of feature extraction layers of the generator, Φ l() is the feature extraction function of the l-th layer of the generator. The feature extraction function can adopt any one of the Sigmoid activation function, the convolution kernel function, or the ReLU activation function. ∥∥ is the L2 norm, representing the difference between the two in the feature space of this layer.

[0118] Taking into account both the adversarial loss and the feature matching loss comprehensively, the calculation method of the total loss function of the generator is expressed as:

[0119]

[0120] In the formula, is the total loss of the generator, and λ f is the weight hyperparameter of feature matching, representing the degree of attention given to different loss terms in the comprehensive optimization. λ f is set to 0.2.

[0121] The feature matching loss (measuring the difference of the generated samples in the multi-layer feature space) and the task relevance loss (guiding the generated data to be more in line with the requirements of the building construction task) effectively solve the problem of insufficient sample authenticity and task relevance during data augmentation.

[0122] S303. Define the loss function during the discriminator training process. The discriminator judges the authenticity of the samples and the matching of task features, and optimizes its weights by maximizing the discrimination between judging real building construction data samples as real and generated building construction data samples as fake. Adopting the dual discriminant loss, the discriminant adversarial loss ensures that the discriminator can distinguish real and generated samples, and the task relevance discriminant loss guides the discriminator to focus on task features, making the generated building construction data more in line with the actual requirements. In the dual adversarial loss of the discriminator, the calculation method of the discriminant adversarial loss is expressed as:

[0123]

[0124] In the formula, is the discriminant adversarial loss, represents the expected value when the input is a real building construction data sample, represents the expected value when the input is a generated building construction data sample.

[0125] The calculation method of the task relevance discriminant loss is expressed as:

[0126]

[0127] In the formula, is the task relevance discriminant loss, and Φ er () is a preset classification function, and the preset classification function can adopt any one of the pre-trained decision tree, random forest, or support vector machine.

[0128] The calculation method of the total loss function of the discriminator is expressed as:

[0129]

[0130] In the formula, is the total loss of the discriminator, and λ t is the weight hyperparameter of the task relevance loss. λ t is set to 0.3.

[0131] S304. During the iterative training process, after receiving the noise vector and task feature information, the generator generates samples with real distribution characteristics and task requirements through the multi-layer non-linear mapping of the network, and uses the feature matching loss to improve the similarity between the generated samples and the real building construction data samples in the high-dimensional space, so as to ensure the accuracy and diversity of the building construction data. The discriminator discriminates the authenticity and feature matching of the input real samples and generated samples, provides feedback signals for the generator, and through the adversarial loss and task relevance discrimination loss, the discriminator gradually improves its discrimination ability during optimization, and in turn promotes the generator to iteratively update in the direction of deceiving the discriminator and meeting the task feature requirements.

[0132] S305. During the iterative training process, calculate the respective losses of the generator and the discriminator according to the results output by the discriminator, and update the network parameters in the way of backpropagation. The generator continuously improves the sample diversity and task adaptability by minimizing its total loss, and the discriminator continuously enhances its discrimination ability by maximizing its discrimination degree of the real building construction data samples and the generated building construction data. The update method of the generator weights is expressed as:

[0133]

[0134] In the formula, is the weight of the generator at the (t + 1)-th iteration, is the weight of the generator at the t-th iteration, and η is the learning rate of the generator. g is set to 0.01, is the gradient of the total loss of the generator with respect to its weight.

[0135] The update method of the discriminator weights is expressed as:

[0136]

[0137] In the formula, is the weight of the discriminator at the (t + 1)-th iteration, is the weight of the discriminator at the t-th iteration, and η is the learning rate of the discriminator. d is set to 0.05, is the gradient of the total loss of the discriminator with respect to its weight.

[0138] S306. During the iterative training process, the generator adaptively adjusts the generation strategy during training, continuously optimizing the diversity and feature matching degree of the generated samples through the feedback of the discriminator to meet the comprehensive requirements for the distribution and labels of construction data in practical applications. Each newly generated sample will be evaluated by the discriminator, and the generator updates its own parameters accordingly to ensure that the quality and diversity of the generated construction data gradually improve in subsequent iterations.

[0139] S307. Repeat the above process iteratively until the parameters of the generator and the discriminator tend to be stable, and the performance difference between the construction data generated by the generator and the real construction data samples in the discriminator is reduced to the extent that meets the application requirements, and is consistent with the real construction data samples at the feature and label levels. At this time, the generator can output high-quality construction training data that is rich, diverse, and task-adaptive enough. The preset stop iteration condition is that the total loss of the generator and the total loss of the discriminator reach the convergence state simultaneously.

[0140] S4. Extract the features of construction data based on the neural network optimized by the non-linear halo.

[0141] The non-linear halo optimization method is inspired by the halo effect. The halo effect usually means that when an observer evaluates a certain thing, it will be affected by other construction data features of the thing, thus changing the evaluation of the construction data features of the thing itself. During the optimization process of the neural network, the update of the weight and bias parameters of the neural network is not only affected by the current state, but also affected by the neighboring states. The update process of each parameter is adjusted by the parameters in its surrounding neighborhood, thus forming a feedback mechanism similar to the halo effect, enabling each optimization to dynamically adjust the neighboring parameters of each neuron in the neural network, thereby accelerating convergence and enhancing the global search ability of the neural network in the parameter space. Through the halo effect mechanism, the update of the current neural network parameters is associated with the neighborhood parameters to achieve dynamic feedback optimization, simulating the correlation between construction data features, such as the interactive influence of different features (such as temperature, humidity, vibration data), so as to more comprehensively extract the complex features in the construction scene.

[0142] The training process of the neural network algorithm optimized by the non-linear halo includes the following steps:

[0143] S401. Initialize the weights and biases of the neural network, expressed as:

[0144]

[0145] In the formula, is the initial weight of the neural network, is the initial bias of the neural network, ~ follows a specific distribution, and U(-0.5, 0.5) represents a random variable uniformly distributed within the interval [-0.5, 0.5].

[0146] The randomness of the initial perturbation amplitude is achieved by using a non-linear perturbation model, and the calculation method is expressed as:

[0147]

[0148] In the formula, is the initial perturbation amplitude, σ p is the perturbation amplitude scaling factor, σ p is set to 0.1, α p is the perturbation adjustment parameter, α p is set to 5.

[0149] S402. In each iteration process, according to the current parameter state of the neural network, calculate the feedback of the halo effect for each parameter. The halo effect affects the update of the current parameter through the weighted sum of neighboring parameters. The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change amount of the loss function, enhancing the modeling ability of the relationship between construction data features in the feature extraction process, making the extracted features more practically significant. The calculation method of the halo effect feedback is expressed as:

[0150]

[0151] In the formula, is the feedback of the halo effect in the t-th iteration, N p represents the neighborhood of the current parameter, λ pq is the influence coefficient of the q-th parameter in the neighborhood on the halo effect of the current parameter, is the weight within the neighborhood in the t-th iteration, is the dynamic feedback coefficient of the parameter in the t-th iteration. The dynamic feedback coefficient is adjusted according to the loss change sensitivity to adapt to the high-dimensional non-linear characteristics in the construction data feature extraction task. α pq is the interaction coefficient between the current parameter and the q-th parameter, is the bias within the neighborhood in the t-th iteration.

[0152] The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change of the loss function, and the calculation method is expressed as:

[0153]

[0154] In the formula, κ p is the sensitivity parameter of the feedback coefficient, κ p takes values between 0.5 and 1.5, is the change amount of the loss function of the neural network in the t-th iteration.

[0155] The loss function of the neural network adopts cross-entropy loss, and the calculation of cross-entropy loss is obtained by calculating the true label matrix and the predicted label matrix of the samples. The predicted labels in the predicted label matrix are calculated by applying the preset Softmax to the feature vectors extracted from the construction data features of the neural network.

[0156] The halo effect influence coefficient and the interaction coefficient control the interaction between different neurons, and can be dynamically adjusted according to the structural distance between neurons (such as neuron layer, topological structure, etc.) and the current state. To enhance the adaptability of the model, the calculation methods of the halo effect influence coefficient and the interaction coefficient are expressed as:

[0157]

[0158] In the formula, δ pq is the first non-linear function, δ′ pq is the second non-linear function, p ner is the index of the current neuron, q ner is the index of the q-th neuron, d max is the distance between the farthest neurons in the neural network, that is, the number of neurons in the neural network minus 1.

[0159] S403. Based on the parameter state at the current iteration, a perturbation term is generated using a chaotic map. Each parameter is dynamically adjusted by adopting the perturbation term to enhance the diversity of optimization and push the neural network out of the local optimum. The chaotic map enhances the diversity of optimization and prevents the neural network from falling into the local optimum. For the complex non-linear features existing in the construction scenario (such as the correlation between the operating state of construction equipment and environmental noise), chaotic optimization provides a powerful exploration ability. The calculation method of the perturbation term is expressed as:

[0160]

[0161] In the formula, is the perturbation term of the neural network weight at the t-th iteration, α pfg is the first coefficient controlling the chaotic perturbation intensity, f chaos () is the chaotic map function, f chaos (,t) is the chaotic map function at the t-th iteration, f chaos (,t - 1) is the chaotic map function at the (t - 1)-th iteration, is the weight of the neural network at the (t - 1)-th iteration, int(t - 1) represents the index of the (t - 1)-th iteration number, is the perturbation term of the neural network bias at the t-th iteration, β pfg is the second coefficient controlling the chaotic perturbation intensity, is the bias of the neural network for the (t - 1)-th iteration.

[0162] To enhance the complexity and non-linearity of the perturbation, the chaotic mapping function is defined as a combined model based on the hyperbolic tangent function and the sine function, expressed as:

[0163]

[0164] where is the dynamic feedback coefficient of the parameter for the t-th iteration, ω p is the frequency coefficient, ω p is set to 5, int(t) represents the current iteration number, and int(t - 1) represents the previous iteration number.

[0165] S404, After performing the halo effect feedback and chaotic perturbation, calculate the update increment of each parameter and add it to the current value of the parameter to form the update amount of the neural network parameters. The calculation method is expressed as:

[0166]

[0167] where is the update amount of the neural network weight for the t-th iteration, λ pgv is the feedback coefficient for weight update, λ pgv is set to 0.1, is the perturbation amplitude for the t-th iteration, is the update amount of the neural network bias for the t-th iteration, λ bev is the feedback coefficient for bias update, λ bev is set to 0.3.

[0168] S405, According to the change of the loss function in each iteration, adaptively adjust the perturbation amplitude so that the optimization process can adjust the intensity of the perturbation according to the convergence state of the loss, avoiding instability caused by excessive perturbation or inefficient optimization caused by too small perturbation. The adjustment calculation method of the perturbation amplitude is expressed as:

[0169]

[0170] where is the perturbation amplitude for the (t + 1)-th iteration, γ ped is the perturbation amplitude adjustment factor, γ ped is set to 0.95, is the change amount of the loss function of the neural network for the t-th iteration, represents the maximum value of the change amounts of the loss functions in all previous iterations of the neural network.

[0171] S406, Perform the update of the weights and biases of the neural network. The update method is expressed as:

[0172]

[0173] In the formula, is the weight of the neural network at the t-th iteration, is the weight of the neural network at the (t + 1)-th iteration, is the bias of the neural network at the t-th iteration, is the bias of the neural network at the (t + 1)-th iteration.

[0174] S407. Repeat the above steps iteratively until the preset iteration stop condition is met. The preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times, which means the model training is completed. As Figure 1 shown, the loss changes of the halo-optimized neural network and the conventional neural network in 100 iterations. The loss of the halo-optimized neural network drops faster and relatively smoothly, showing strong convergence ability.

[0175] After the neural network training is completed, use the trained neural network to extract features from the building construction data, input the data after feature extraction into the preset Softmax function to calculate the class probabilities, and take the class with the maximum class probability as the category for building construction quality assessment. For example, the categories include excellent construction quality, construction quality to be improved, construction abnormal events, construction delay risks, and unqualified materials, a total of 5 categories.

[0176] S5. Generate a comprehensive quality assessment report through the predicted category output by the building construction quality assessment model, identify the deviations in the building construction process, and conduct quantitative analysis on the deviations to judge their severity.

[0177] S6. Automatically generate an optimization plan according to the evaluation results, including suggestions such as construction process adjustment and resource reallocation, provide detailed improvement priorities and implementation steps, and feedback the problems identified in the evaluation to the BIM model to update the model to reflect the current construction status.

[0178] Meanwhile, this embodiment can also achieve flexible operation and expansion through a highly compatible and user-friendly interaction interface. In addition, an open API (Application Programming Interface) is provided to facilitate integration with project management systems, enterprise resource planning systems, and other building management platforms.

[0179] The above is only a preferred embodiment of the present invention and is not used to limit the protection scope of the present invention.

Claims

1. An artificial intelligence-based BIM data processing and building construction quality assessment method, characterized in that It includes the following steps: S1. Collect building data, and manually label the collected building data. The labeling categories include: excellent construction quality, construction quality to be improved, construction abnormal events, construction delay risks, and unqualified materials; S2. Conduct unified preprocessing on the building data; S3. Generate samples based on the feature-matching generative adversarial network to achieve the expansion of building construction data; S4. Extract features of building construction data based on the neural network optimized by non-linear halos; S5. Generate a comprehensive quality assessment report through the predicted categories output by the building construction quality assessment model, identify deviations in the building construction process, quantitatively analyze the deviations, and judge their severity; S6. Automatically generate an optimization plan according to the evaluation results, including suggestions such as construction process adjustment and resource reallocation, provide detailed improvement priorities and implementation steps, and feedback the problems identified in the evaluation to the BIM model to update the model to reflect the current construction status.

2. The BIM data processing and building construction quality assessment method based on artificial intelligence according to claim 1, characterized in that: The attributes of the construction data in S1 include the spatial relationship feature R in the BIM model a , the construction stage timestamp D a , the temperature data T collected by the sensor a , the humidity data H collected by the sensor a , the acceleration V collected by the vibration monitoring device a , the text content feature C in the construction record a , the material type identifier M a , the semantic label L of the construction log a , the quality inspection result Q a , the geographical location coordinates S of the construction area a .

3. The method for BIM data processing and building construction quality assessment based on artificial intelligence according to claim 1, characterized in that: In the above S1, the sources of building data include: BIM model input, Internet of Things data collection, and construction document integration.

4. The method for BIM data processing and building construction quality assessment based on artificial intelligence according to claim 1, wherein: In the above S2, the preprocessing includes data cleaning and formatting, data completion and fusion, and time series data processing.

5. The BIM data processing and building construction quality assessment method based on artificial intelligence according to claim 1, characterized in that, The method of generating samples based on the feature-matching generative adversarial network in the above S3 includes the following steps: S301. Initialize the generator and discriminator of the generative adversarial network. The generator converts random noise and task features into building construction data samples through multi-layer non-linear mapping. The discriminator discriminates the authenticity and task feature matching degree of the input samples, and optimizes their weight parameters through interactive confrontation, so as to gradually improve the quality and diversity of the generated building construction data. In order to ensure the diversity of the generated building construction data, the noise vector is processed in the way of a variational autoencoder, and then a more complex and task-demand-fitting latent representation is obtained, which is expressed as: where Z c is the noise vector encoded by the variational autoencoder, and f VAE () is the encoding function of the variational autoencoder, which characterizes how to map the real building construction data samples into vectors in the latent space, is the real building construction data sample, and Θ VAE are the parameters of the variational autoencoder and are training parameters obtained by training through the gradient descent method; The generator generates building construction data according to the noise vector encoded by the variational autoencoder and the task features, which is expressed as: In the formula, is the generated building construction data sample, and G c () is the mapping function of the generator, and T c is the task feature information, representing the associated features of the task, is the weight of the generator, representing the parameters that are continuously updated during the network training process; S302. Define the loss function in the training process of the generator. The adversarial loss part measures the game between the generator and the discriminator in terms of authenticity discrimination. The feature matching loss part measures the difference between the generated samples and the real samples in the multi-layer feature space, ensuring that the generated building construction data is consistent with the real building construction data samples in terms of fine-grained features. The loss function of the generator is used to maximize the misjudgment probability of the discriminator for the generated building construction data. Finally, the comprehensive consideration of the adversarial loss and the feature matching loss is used as the total loss function in the training process of the generator, which is expressed as: In the formula, is the adversarial loss of the generator, is the expected value, Z c is the noise vector, T c is the task feature, represents the expected value under the input of the noise vector and the task feature, D c () is the mapping function of the discriminator, representing the estimated probability that the input sample is real, is the function of the discriminator; A multi-layer feature extractor is used to measure the difference between the generated samples and the real samples. The calculation method of the feature matching loss is expressed as: Where, L fm is the feature matching loss, and L cea is the number of feature extraction layers of the generator. Φ l () is the l-th layer feature extraction function of the generator. The feature extraction function can be any one of the Sigmoid activation function, the convolution kernel function, or the ReLU activation function. ∥∥ is the L2 norm, representing the difference between the two in the feature space of this layer; Comprehensively considering the adversarial loss and the feature matching loss, the calculation method of the total loss function of the generator is expressed as: In the formula, is the total loss of the generator, and λ f is the weight hyperparameter for feature matching, which characterizes the degree of attention given to different loss terms in the comprehensive optimization; S303. Define the loss function during the discriminator training process. The discriminator judges the authenticity of the samples and the matching of task features, and optimizes its weights by maximizing the discrimination between correctly classifying real construction data samples as real and generated construction data samples as fake. A dual discriminant loss is adopted. The discriminative adversarial loss ensures that the discriminator can distinguish between real and generated samples, and the task-related discriminant loss guides the discriminator to focus on task features, making the generated construction data more in line with actual requirements. In the dual adversarial loss of the discriminator, the calculation method of the discriminative adversarial loss is expressed as: In the formula, is the discriminative adversarial loss, represents the expected value when the input is a real building construction data sample, represents the expected value when the input is a generated building construction data sample; The calculation method of the task-related discriminant loss is expressed as: In the formula, is the task relevance discrimination loss, and Φ er () is a preset classification function, and the preset classification function adopts any one of a pre-trained decision tree, a random forest, or a support vector machine; The calculation method of the total loss function of the discriminator is expressed as: In the formula, is the total discriminator loss, and λ t is the weight hyperparameter of the task relevance loss; S304. During the iterative training process, after receiving the noise vector and task feature information, the generator generates samples with real distribution characteristics and task requirements through the multi-layer non-linear mapping of the network, and uses the feature matching loss to improve the similarity between the generated samples and the real construction data samples in the high-dimensional space, thereby ensuring the accuracy and diversity of the construction data; the discriminator discriminates the authenticity and feature matching of the input real samples and generated samples, and provides feedback signals for the generator. Through the adversarial loss and the task-related discriminant loss, the discriminator gradually improves its discrimination ability during optimization, which in turn drives the generator to iteratively update in the direction of deceiving the discriminator and meeting the task feature requirements; S305. During the iterative training process, calculate the respective losses of the generator and the discriminator according to the results output by the discriminator, and update the network parameters in the way of backpropagation. The generator continuously improves the sample diversity and task adaptability by minimizing its total loss, and the discriminator continuously enhances its discrimination ability by maximizing the discrimination between real construction data samples and generated construction data. The update method of the generator weights is expressed as: Wherein, is the weight of the generator at the (t + 1)-th iteration, is the weight of the generator at the t-th iteration, and η is the learning rate of the generator, g is set to 0.01, is the gradient of the total loss of the generator with respect to its weight; The update method of the discriminator weights is expressed as: Wherein, is the weight of the discriminator at the (t + 1)-th iteration, is the weight of the discriminator at the t-th iteration, and η is the learning rate of the discriminator d is set to 0.05, is the gradient of the total loss of the discriminator with respect to its weight; S306. During the iterative training process, the generator adaptively adjusts the generation strategy during training, and continuously optimizes the diversity and feature matching degree of the generated samples through the feedback of the discriminator to meet the comprehensive requirements of the distribution and label requirements of construction data in practical applications. Each newly generated sample will be evaluated by the discriminator, and the generator updates its own parameters accordingly to ensure that the quality and diversity of the generated construction data are gradually improved in subsequent iterations; S307. Iterate the above process repeatedly until the parameters of the generator and the discriminator tend to be stable. The performance difference between the construction data generated by the generator and the real construction data samples in the discriminator is reduced to the extent that meets the application requirements, and is consistent with the real construction data samples at the feature and label levels. The generator can then output high-quality construction training data that is rich, diverse, and task-adaptive. The preset stop iteration condition is that the total losses of the generator and the discriminator both reach the convergence state.

6. The method for BIM data processing and building construction quality assessment based on artificial intelligence according to claim 1, wherein: The training process of the neural network algorithm based on non-linear halo optimization in the described S4 includes the following steps: S401. Initialize the weights and biases of the neural network, which is expressed as: In the formula, is the initial weight of the neural network, is the initial bias of the neural network, and ~ follows a specific distribution. U(-0.5, 0.5) represents a random variable uniformly distributed within the interval [-0.5, 0.5]; The randomness of the initial perturbation amplitude is realized by using a non-linear perturbation model, and the calculation method is expressed as: In the formula, is the initial perturbation amplitude, and σ p is the perturbation amplitude scaling factor, and α p is the perturbation adjustment parameter; S402. In each iteration process, according to the current parameter state of the neural network, calculate the feedback of the halo effect for each parameter. The halo effect affects the update of the current parameter through the weighted sum of neighboring parameters. The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change amount of the loss function, enhancing the modeling ability of the feature relationship in the construction data during the feature extraction process, making the extracted features more practical. The calculation method of the halo effect feedback is expressed as: In the formula, is the halo effect feedback of the t-th iteration, N p represents the neighborhood of the current parameter, λ pq is the influence coefficient of the q-th parameter in the neighborhood on the halo effect of the current parameter, is the weight within the neighborhood of the t-th iteration, is the dynamic feedback coefficient of the parameter of the t-th iteration. The dynamic feedback coefficient is adjusted according to the loss change sensitivity to adapt to the high-dimensional non-linear characteristics in the task of building construction data feature extraction, α pq is the interaction coefficient between the current parameter and the q-th parameter, is the bias within the neighborhood of the t-th iteration; The dynamic feedback coefficient of the parameter is adaptively adjusted according to the change of the loss function, and the calculation method is expressed as: where κ p is the sensitivity parameter of the feedback coefficient, κ p takes values between 0.5 and 1.5, is the change in the loss function of the neural network at the t-th iteration; The halo effect influence coefficient and the interaction coefficient control the interaction between different neurons, and can be dynamically adjusted according to the structural distance between neurons and the current state. To enhance the adaptability of the model, the calculation methods of the halo effect influence coefficient and the interaction coefficient are expressed as: where δ pq is the first non-linear function, δ′ pq is the second non-linear function, p ner is the index of the current neuron, q ner is the index of the q-th neuron, d max is the distance between the farthest neurons in the neural network, that is, the number of neurons in the neural network minus 1; S403. Based on the parameter state at the current iteration, use the chaotic map to generate a perturbation term, and dynamically adjust each parameter by using the perturbation term, enhancing the diversity of optimization and promoting the neural network to jump out of the local optimum. The chaotic map enhances the diversity of optimization and avoids the neural network falling into the local optimum. For the complex non-linear features existing in the building construction scenario, chaotic optimization provides a powerful exploration ability. The calculation method of the perturbation term is expressed as: Wherein, is the perturbation term of the neural network weights at the t-th iteration, and α pfg is the first coefficient for controlling the chaotic perturbation intensity, and f chaos () is the chaotic mapping function, and f chaos (, t) is the chaotic mapping function at the t-th iteration, and f chaos (, t - 1) is the chaotic mapping function at the (t - 1)-th iteration, is the weight of the neural network at the (t - 1)-th iteration, and int(t - 1) represents the index of the number of iterations at the (t - 1)-th iteration, is the perturbation term of the neural network bias at the t-th iteration, and β pfg is the second coefficient for controlling the chaotic perturbation intensity, is the bias of the neural network at the (t - 1)-th iteration; To enhance the complexity and non-linearity of the perturbation, the chaotic map function is defined as a combined model based on the hyperbolic tangent function and the sine function, which is expressed as: wherein, is the dynamic feedback coefficient of the parameter for the t-th iteration, ω p is the frequency coefficient, int(t) represents the current iteration number, and int(t - 1) represents the previous iteration number; S404. After the halo effect feedback and chaotic perturbation, calculate the update increment of each parameter, and add it to the value of the current parameter to form the update amount of the neural network parameter. The calculation method is expressed as: In the formula, is the update amount of the neural network weights at the t-th iteration, and λ pgv is the feedback coefficient for weight update, is the perturbation amplitude at the t-th iteration, is the update amount of the neural network bias at the t-th iteration, and λ bev is the feedback coefficient for bias update; S405. According to the change of the loss function in each iteration, adaptively adjust the perturbation amplitude, so that the optimization process can adjust the intensity of the perturbation according to the convergence state of the loss, avoiding instability caused by excessive perturbation or inefficient optimization caused by too small perturbation. The calculation method of the adjustment of the perturbation amplitude is expressed as: In the formula, is the perturbation amplitude of the (t + 1)-th iteration, and γ ped is the perturbation amplitude adjustment factor. γ ped is set to 0.95, is the change in the loss function of the neural network at the t-th iteration, represents the maximum value of the changes in the loss function of the neural network in all previous iterations; S406. Update the weights and biases of the neural network, and the update method is expressed as: Wherein, is the weight of the neural network at the t-th iteration, is the weight of the neural network at the (t + 1)-th iteration, is the bias of the neural network at the t-th iteration, is the bias of the neural network at the (t + 1)-th iteration; S407. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which means the model training is completed.

Images