Industrial big data processing system and method for building decoration base layer

Abstract

The invention discloses an industrial big data processing system and method for a building decoration base layer, relates to the technical field of big data processing, and solves the problem of low industrial big data processing capability of the building decoration base layer. According to the technical scheme, the system comprises a multi-mode environment sensing network, an Internet of Things communication module, a data fusion module, a diagnosis early warning module based on reinforcement learning and a distributed data interaction terminal, and the output end of the multi-mode environment sensing network is connected with the input end of the Internet of Things communication module; the output end of the Internet of Things communication module is connected with the input end of the data fusion module, the output end of the data fusion module is connected with the input end of the diagnosis and early warning module based on reinforcement learning, and the output end of the diagnosis and early warning module is connected with the input end of the distributed data interaction terminal. The industrial big data processing capability of the building decoration base layer is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and specifically designs a big data processing system and method for the industrialization of building decoration bases. Background Art

[0002] With the rapid development of the building decoration industry, the quality and efficiency of base construction techniques have become the core factors affecting the decoration effect. The traditional manual operation mode, due to its low precision (error rate exceeding 15%) and poor efficiency (construction period extended by more than 30%), can no longer meet the requirements of modern projects. At the same time, the application of big data technology has provided new opportunities for industry upgrading, but the existing monitoring systems have the following technical bottlenecks: 1. Insufficient monitoring ability; The layout of traditional single-point sensors is sparse, unable to capture the spatio-temporal coupling characteristics of the construction environment (such as the undetected deformation rate of finishing materials reaching as high as 15%), and the early warning mechanism relying on fixed threshold rules is difficult to adapt to dynamic working conditions. The early warning of hidden project leakage is delayed by more than 4 hours, resulting in a maintenance cost increase of more than 60%.

[0003] 2. Low data fusion efficiency; There is a lack of a unified fusion framework for multi-source heterogeneous data (vibration, temperature and humidity, BIM models, etc.). The existing weighted average method cannot explore the deep correlations of data, and the data island problem leads to a decision misjudgment rate of 25%.

[0004] 3. Poor communication reliability; Strong electromagnetic interference at the construction site (such as the pulse noise of electric welders) causes the error rate of traditional wireless communication (Zigbee) to be as high as 8%, and the loss of key data leads to the risk of monitoring interruption.

[0005] 4. Lagging intelligent level; The diagnostic model relies on manual experience and needs to be retrained for cross-project migration (time-consuming over 200 hours per project), with weak generalization ability; the distributed sensing network lacks an adaptive sampling strategy, and the response is delayed by more than 30 minutes when the construction progress changes.

[0006] 5. Inflexible data interaction; The centralized data storage architecture requires on-site engineers to access through fixed terminals, with a mobile operation response delay of more than 1 hour and low emergency efficiency.

[0007] The current process of building industrialization has put forward new requirements for grass-roots construction, including high precision (millimeter level), high reliability (above 99%), and high intelligence (real-time decision-making). However, there are four core defects in the existing technologies: the distributed sensing network lacks the ability of dynamic optimization and cannot adapt to the changes in the construction progress; the compatibility of multi-protocol communication is poor, and the equipment access efficiency is reduced by more than 20%; the data security protection is weak, and the risk of leakage of sensitive construction data is high; the generalization ability of the diagnosis model is insufficient, and the cost of cross-project application is high.

[0008] With the advancement of enterprise digital transformation, the traditional big data processing methods based on statistics and simple classification can no longer meet the real-time analysis requirements of massive heterogeneous data. It is urgent to build an intelligent monitoring and diagnosis system through the integration of multi-disciplinary technologies to solve problems such as data processing lag and low decision-making efficiency, and provide new technical support for building industrialization. Summary of the Invention

[0009] Aiming at the deficiencies of the above technologies, the present invention discloses a big data processing system and method for the industrialization of building decoration grass-roots, which can effectively solve the deficiencies mentioned in the background technology.

[0010] In order to achieve the above technical effects, the present invention adopts the following technical solutions: An intelligent monitoring and diagnosis system for building decoration grass-roots, which includes: A multi-modal environment perception network, including distributed heterogeneous sensing nodes and an edge computing gateway. The sensing nodes are built-in with a composite sensor array to collect the dynamic parameters, material deformation parameters, and environmental interference parameters of the construction environment in real time. The edge computing gateway is configured with a data cleaning engine and an event recognition model, and eliminates abnormal data through an adaptive filtering algorithm and generates event feature vectors. An Internet of Things communication module, including communication interfaces compatible with LoRa, NB-IoT, and Zigbee protocols, a device fingerprint recognition module, a dynamic spectrum sensing module, multi-hop self-organizing network nodes, and a data distribution module. The output end of the communication interface is connected to the input end of the device fingerprint recognition module, the output end of the device fingerprint recognition module is connected to the input end of the dynamic spectrum sensing module, the output end of the dynamic spectrum sensing module is connected to the input end of the multi-hop self-organizing network nodes, and the output end of the multi-hop self-organizing network nodes is connected to the input end of the data distribution module. The data fusion module includes a 3D digital twin modeling module, a dynamic key update module, a dynamic threshold rule setting module, an edge-cloud collaborative computing module, and a data spatio-temporal stamp verification module. The output end of the 3D digital twin modeling module is connected to the input end of the dynamic key update module, the output end of the dynamic key update module is connected to the input end of the dynamic threshold rule setting module, the output end of the dynamic threshold rule setting module is connected to the input end of the edge-cloud collaborative computing module, and the output end of the edge-cloud collaborative computing module is connected to the input end of the data spatio-temporal stamp verification module; The diagnosis and early warning module based on reinforcement learning includes a feature extraction module, an anomaly recognition module, a deep Q-network training module, a fault location module, and a fault early warning module. The output end of the feature extraction module is connected to the input end of the anomaly recognition module, the output end of the anomaly recognition module is connected to the input end of the deep Q-network training module, the output end of the deep Q-network training module is connected to the input end of the fault location module, and the output end of the fault location module is connected to the input end of the fault early warning module; The distributed data interaction terminal deploys multiple data nodes according to the characteristics and requirements of the construction data. Each node sets different mobile terminals to be responsible for collecting and processing data in a specific area, so as to facilitate users to process the abnormal data information of the building decoration base at any time.

[0011] The output end of the multi-modal environment perception network is connected to the input end of the Internet of Things communication module, the output end of the Internet of Things communication module is connected to the input end of the data fusion module, the output end of the data fusion module is connected to the input end of the diagnosis and early warning module based on reinforcement learning, and the output end of the diagnosis and early warning module is connected to the input end of the distributed data interaction terminal.

[0012] As a further technical solution of the present invention, the working method of the adaptive filtering algorithm is as follows: Step 1, multi-dimensional feature extraction; The time-domain features reflect the average level, fluctuation degree, and sudden change of the data by calculating the mean, variance, standard deviation, and peak parameters of the data in the time domain; The frequency-domain features perform Fourier transform on the filtered data to obtain the frequency components and spectral amplitudes, and identify specific construction events; The time-frequency domain features use wavelet transform to extract the wavelet coefficients at different scales and time positions as feature vectors; Step 2, data interpolation and feature selection; For missing data, the modified interpolation method is used to calculate the correlation coefficient of data variables. The variable with the maximum correlation coefficient is used as the prediction variable, and its mean and standard deviation are calculated. The predicted value is used to replace the unknown data, and the target data column is found as the data to be corrected. According to the data change trend, a suitable value is selected from the data to be corrected and the predicted value for filling. III. Noise analysis and processing; The improved Kalman filtering algorithm is adopted, and the primary state transition is changed to the secondary state transition to improve the filtering accuracy. A sparse noise term is added to the state space construction to enhance the noise recognition and processing ability of the filtering algorithm. The weighted mean and weighted variance operations are performed on the sampled and transformed data to achieve adaptive filtering. IV. Abnormal data detection and elimination; An error threshold is set to determine whether the data is abnormal. When the absolute value of the error e(n) exceeds the threshold, the data x(n) at the current moment is eliminated. As a further technical solution of the present invention, the abnormal recognition module at least includes an improved random forest algorithm model, and the improved random forest algorithm model includes a classification error module, a classification calculation module, a classification prediction module, a classifier rule generation module, and a data format processing module.

[0013] As a further technical solution of the present invention, the deep Q-network training module includes a lightweight autoencoder, a spatio-temporal convolutional DQN module, a double-delay depth calculation module, a data search module, and a data dynamic update module. The output end of the lightweight autoencoder is connected to the input end of the spatio-temporal convolutional DQN module, the output end of the spatio-temporal convolutional DQN module is connected to the input end of the double-delay depth calculation module, the output end of the double-delay depth calculation module is connected to the input end of the data search module, and the output end of the data search module is connected to the input end of the data dynamic update module.

[0014] As a further technical solution of the present invention, the fault location module includes a power supply module and a clock module, a GPS module, a synchronization signal detection unit, a synchronization positioning circuit unit, and a digital twin diagnosis output module connected to the power supply module.

[0015] As a further technical solution of the present invention, the fault warning module includes a warning processor and an acoustic-optic warning unit, a signal conversion unit, and a data error repairer connected to the warning processor.

[0016] As a further technical solution of the present invention, the working method of the improved random forest algorithm model is as follows: Random training sample sets are extracted from the input variables and the output variables to be processed of the original industrialized data of building decoration. The error information is calculated through the average generalization error function, and the average generalization error function is: In formula (1), represents the average generalization error, x represents the input variable, and y represents the output variable. represents the budget result, D is the sample set, f is the model learned from the training set D, h represents the distance, p represents the dimension, and Wh represents the data attribute; then data prediction is performed, and the budget result output is: In Equation (2), represents the generalization error when n = 0, and H is the training sample set. is the budget mean. is the nth training sample set, K represents there are K categories; the calculation relationship between the budget value calculated by n decision trees and the output variable is represented by Formula (2), and the correlation function is: In Equation (3), represents the generalization error when represents the generalization error of the nth decision tree, and Y represents the value range of the output variable y; the relationship between the output variable, the generalization error of the nth decision tree, and the output variable and the training sample set is represented by Formula (3), and the generalization error when is obtained through the relationship calculation, and the generalization error function is: The calculation relationship between the generalization error of the nth decision tree, the output variable, and the budget mean is represented by Formula (4), and the generalization error when is also obtained from the said calculation relationship. In Equation (4), H satisfies: In Equation (5), is the residual weighted correlation coefficient; the relationship between the generalization error when n = 0 and the generalization error when is represented by Formula (5), and the product of the generalization error when and the residual weighted correlation coefficient is greater than or equal to the generalization error when n = 0, and the classification output is: In Equation (7), Z is the scaling result of the original data, z is the original data, is the minimum value of the original data, is the maximum value of the original data, and A represents the feature information. represents the numerical parameter, and d represents the numerical interval. represents the maximum value of the optimization objective. Represents the minimum value of the optimization objective. As a further technical solution of the present invention, the working principle of the lightweight autoencoder is as follows: by reducing the number of network layers and neurons, the complexity and computational amount of the model are reduced; by optimizing the parameter initialization method, the model converges within 1 s during the training process, reducing the risk of overfitting, and by assigning meanings to the encoding layer, the recognition ability is improved; The working principle of the spatio-temporal convolutional DQN module is as follows: integrating the network architectures of the convolutional neural network CNN and the deep Q network DQN, extracting the spatio-temporal features of the data during the construction process through the spatio-temporal convolutional layer, and capturing the changing rules of the construction state in space and time; dynamically processing the progress promotion or the conversion information of different construction stages during the building decoration construction process continuously for 24 hours, estimating the extracted spatio-temporal features to calculate the Q value, so as to accurately select and calculate the optimal action, improving the efficiency and quality of decision-making; The working principle of the double-delay deep calculation module is as follows: estimating the Q value by setting the main network and the target network, and estimating the Q value of each action in the current state in real time; data search module The data dynamic update module searches for the characteristics of the building decoration construction data through the heuristic search algorithm, and gives the search results through the time dimension, the space dimension or the data type dimension.

[0017] As a further technical solution of the present invention, the clock module is provided with a quantum-enhanced clock module, which calibrates the crystal oscillator output through the quantum transition frequency of 6.934 GHz, realizing a frequency stability of 1×10⁻¹³; and generating a time stamp based on the SHA-256 hash, forming a dual signature with the Beidou satellite time service data to ensure that the time cannot be tampered with; The GPS module integrates UWB positioning tags in GPS positioning, and realizes seamless indoor and outdoor positioning through the time difference of arrival TDOA algorithm, with an accuracy of ±3 cm; The synchronization signal detection unit is automatically identified through PTP, IRIG-B, and custom protocols, and realizes a fast switching of 0.5 ms through the protocol fingerprint library of the CRC check characteristics of Modbus RTU; The synchronous positioning circuit unit is based on generating a dual hash value by "time stamp + coordinate" for the synchronous positioning result and writing it into the blockchain to form an immutable record; The digital twin diagnosis output module is provided with a neural radiance field NeRF digital twin engine.

[0018] As a further technical solution of the present invention, the anomaly recognition module also adopts an improved GoogleNet++ network model to improve the anomaly recognition efficiency; The method for constructing the improved GoogleNet++ network model structure is as follows: Replace the Inception structure in the GoogleNet++ feature network with the Faster-Inception structure, adopt the L-Switsh function as the activation function, adopt the Sigmoid function as the classifier, and the number of neurons in the fully connected layer is 36; Then, replace the 3×3 convolution kernel and 5×5 convolution kernel in the original Inception module of the GoogleNet++ model with 1×3 and 3×1 convolution combinations. The expression of the ELU-Switsh function is: In formula (8), x is the input parameter, is the activation function. The present invention also adopts the following technical solution: A method for industrial big data processing of building decoration base layers. The method includes the steps: Step (1), Arrange a multi-modal environment perception network, set distributed heterogeneous sensing nodes and edge computing gateways, and collect the dynamic parameters of the construction environment, material deformation parameters, and environmental interference parameters in real time, eliminate abnormal data and generate event feature vectors; Step (2), Set up an Internet of Things communication module, including communication interfaces compatible with LoRa, NB-IoT, and Zigbee protocols, a device fingerprint recognition module, a dynamic spectrum sensing module, multi-hop self-organizing network nodes, and a data distribution module; Step (3), Integrate the received data information. Connect the output end of the three-dimensional digital twin modeling module to the input end of the dynamic secret key update module, connect the output end of the dynamic secret key update module to the input end of the dynamic threshold rule setting module, connect the output end of the dynamic threshold rule setting module to the input end of the edge-cloud collaborative computing module, and connect the output end of the edge-cloud collaborative computing module to the input end of the data spatio-temporal stamp verification module; Step (4), Deeply learn the proposed fault data information by the diagnosis and early warning module based on reinforcement learning; Step (5), Deploy multiple data nodes according to the characteristics and requirements of the construction data through a distributed data interaction terminal. Each node is set with a different mobile terminal to be responsible for collecting and processing data in a specific area, so as to facilitate users to process the abnormal data information of the building decoration base layer at any time.

[0019] Compared with the prior art, the beneficial and positive effects of the present invention are: Through the integration of multidisciplinary technologies, the present invention constructs an intelligent monitoring and diagnosis system for the base layer of building decoration, significantly improving the monitoring accuracy, decision-making efficiency, and data security of industrialized construction. Through the above innovative points, this intelligent early warning decision-making center realizes the comprehensive and intelligent management of the construction process, improves the accuracy and timeliness of early warning, reduces the occurrence of construction accidents, optimizes the construction decision-making process, improves construction efficiency and quality, realizes information sharing and collaborative work between different platforms, and reduces the cost of construction management. This technology has broad application prospects in the fields of building construction, infrastructure construction, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Figure 1 It is a schematic diagram of the principle of the data fusion module in the present invention; [[ID=⑨]] Figure 2 It is a schematic diagram of the principle of the diagnostic warning module of reinforcement learning in the present invention; Figure 3 It is a schematic diagram of the working method of the adaptive filtering algorithm in the invention; Figure 4 It is a schematic diagram of the principle of the deep Q-network training algorithm module in the present invention; Figure 5 It is a schematic diagram of the principle of an embodiment of the synchronous information detection circuit in the present invention; Figure 6 It is a schematic diagram of the principle of an embodiment of the synchronous positioning circuit unit in the present invention; Figure 7 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will clearly and completely describe the technical solutions in the embodiments of this article with reference to the drawings in the embodiments of this article. Obviously, the described embodiments are only some embodiments of this article, rather than all embodiments. It should be understood that these descriptions are exemplary and not intended to limit the scope of the present invention. In addition, in the description, the description of well-known structures and technologies is omitted to avoid unnecessarily confusing the concepts of the present invention.

[0022] As Figures 1-7 shown, an intelligent monitoring and diagnosis system for the base layer of building decoration includes: The multi-modal environment perception network includes distributed heterogeneous sensing nodes and an edge computing gateway. The sensing nodes are built-in with a composite sensor array to collect kinetic parameters, material deformation parameters, and environmental interference parameters of the construction environment in real time. The edge computing gateway is configured with a data cleaning engine and an event recognition model to eliminate abnormal data through an adaptive filtering algorithm and generate event feature vectors; The Internet of Things communication module includes communication interfaces compatible with LoRa, NB-IoT, and Zigbee protocols, a device fingerprint recognition module, a dynamic spectrum sensing module, multi-hop self-organizing network nodes, and a data distribution module. The output end of the communication interface is connected to the input end of the device fingerprint recognition module, the output end of the device fingerprint recognition module is connected to the input end of the dynamic spectrum sensing module, the output end of the dynamic spectrum sensing module is connected to the input end of the multi-hop self-organizing network nodes, and the output end of the multi-hop self-organizing network nodes is connected to the input end of the data distribution module; The data fusion module includes a three-dimensional digital twin modeling module, a dynamic secret key update module, a dynamic threshold rule setting module, an edge-cloud collaborative computing module, and a data spatio-temporal stamp verification module. The output end of the three-dimensional digital twin modeling module is connected to the input end of the dynamic secret key update module, the output end of the dynamic secret key update module is connected to the input end of the dynamic threshold rule setting module, the output end of the dynamic threshold rule setting module is connected to the input end of the edge-cloud collaborative computing module, and the output end of the edge-cloud collaborative computing module is connected to the input end of the data spatio-temporal stamp verification module; The diagnosis and early warning module based on reinforcement learning includes a feature extraction module, an anomaly recognition module, a deep Q-network training module, a fault location module, and a fault early warning module. The output end of the feature extraction module is connected to the input end of the anomaly recognition module, the output end of the anomaly recognition module is connected to the input end of the deep Q-network training module, the output end of the deep Q-network training module is connected to the input end of the fault location module, and the output end of the fault location module is connected to the input end of the fault early warning module; The distributed data interaction terminal deploys multiple data nodes according to the characteristics and requirements of construction data. Each node is set with a different mobile terminal to be responsible for collecting and processing data in a specific area, so as to facilitate users to process abnormal data information of the building decoration base at any time.

[0023] The output end of the multi-modal environment perception network is connected to the input end of the Internet of Things communication module, the output end of the Internet of Things communication module is connected to the input end of the data fusion module, the output end of the data fusion module is connected to the input end of the diagnosis and early warning module based on reinforcement learning, and the output end of the diagnosis and early warning module is connected to the input end of the distributed data interaction terminal.

[0024] In the above embodiments, the working method of the adaptive filtering algorithm is as follows: Step 1: Multi-dimensional feature extraction; The time-domain features reflect the average level, fluctuation degree, and sudden changes of the data by calculating the mean, variance, standard deviation, and peak parameters of the data within the time domain; The frequency-domain features perform Fourier transform on the filtered data to obtain the frequency components and spectral amplitudes, and identify specific construction events; The time-frequency domain features use wavelet transform to extract wavelet coefficients at different scales and time positions as feature vectors; Step 2: Data interpolation and feature selection; For missing data, the modified interpolation method is used to calculate the correlation coefficient of the data variables. The variable with the maximum correlation coefficient is used as the prediction variable, and its mean and standard deviation are calculated; the predicted value is used to replace the unknown data, and the target data column is found as the data to be corrected; according to the data change trend, a suitable value is selected from the data to be corrected and the predicted value for filling; Step 3: Noise analysis and processing; The improved Kalman filtering algorithm is used to change the first-order state transfer to the second-order state transfer to improve the filtering accuracy; a sparse noise term is added to the state space construction to enhance the noise recognition and processing ability of the filtering algorithm; weighted mean and weighted variance operations are performed on the sampled and transformed data to achieve adaptive filtering; Step 4: Abnormal data detection and elimination; An error threshold is set to determine whether the data is abnormal; when the absolute value of the error e(n) exceeds the threshold, the data x(n) at the current moment is eliminated. In the above embodiments, the abnormal recognition module at least includes an improved random forest algorithm model, and the improved random forest algorithm model includes a classification error module, a classification calculation module, a classification prediction module, a classifier rule generation module, and a data format processing module.

[0025] The deep Q-network training module includes a lightweight autoencoder, a spatio-temporal convolutional DQN module, a double-delay depth calculation module, a data search module, and a data dynamic update module. The output end of the lightweight autoencoder is connected to the input end of the spatio-temporal convolutional DQN module, the output end of the spatio-temporal convolutional DQN module is connected to the input end of the double-delay depth calculation module, the output end of the double-delay depth calculation module is connected to the input end of the data search module, and the output end of the data search module is connected to the input end of the data dynamic update module.

[0026] When implementing the deep Q-network training module, high-dimensional environmental state features are extracted using ResNet-50, and 8:1 feature compression is achieved using MobileNetV3 with channel pruning for user network deployment. The distillation transfer of key construction features is realized through the KL divergence loss function, reducing the computational amount by 76%. Dilated Conv is used to extract cross-scale structural features of the construction scene, and a gated recurrent unit (GRU) is adopted to capture the temporal dependence of the equipment operating state. The feature fusion layer dynamically allocates the weight ratio of spatial / temporal features through an attention mechanism. During dynamic quantization-aware training, a differentiable quantization operator is introduced to automatically learn the optimal quantization threshold for 8-bit fixed-point numbers during the training process, compressing the model size to 12% of the original size without loss of accuracy.

[0027] The spatio-temporal convolutional DQN module constructs a (H×W×T) three-dimensional convolutional kernel in a three-dimensional causal convolutional architecture, where: in the H / W dimension: capturing the spatial layout features of construction machinery; in the T dimension: establishing the temporal causal relationship of the construction process; using spatio-temporal convolution with dynamically adjusted dilation coefficients, the receptive field is expanded by 3 times. The pre-trained BERT model is used to extract the construction specification text features, generate a semantic attention map, generate a spatial heat map based on sensor data, and achieve the collaborative enhancement of semantic-physical features through a gated fusion unit. A generative adversarial network (GAN) framework is constructed, where the generator generates high-risk construction state samples and the discriminator evaluates the state value, improving the exploration efficiency of the agent by 40%. The policy lag network and value lag network are deployed, and through several hours of experiments, the effects are improved as shown in Table 1. This solution deeply integrates technologies such as deep learning, federated learning, and knowledge graphs with the building industrialization scenario, constructing an intelligent training system with the ability of autonomous evolution.

[0028] In a further technical solution, the fault location module includes a power supply module and a clock module, a GPS module, a synchronization signal detection unit, a synchronization positioning circuit unit, and a digital twin diagnosis output module connected to the power supply module. The fault warning module includes a warning processor and an acoustic-optical warning unit, a signal conversion unit, and a data error corrector connected to the warning processor.

[0029] The power supply module and the clock module provide stable power supply and clock signals for the entire system, ensuring the synchronous operation of all modules. The GPS module receives signals from the Global Positioning System (GPS) and provides accurate timestamps and location information for synchronization and positioning. The synchronization signal detection unit uses the AD52358BU2250 chip circuit and the ADUM1200 / ADUM1201 circuit for signal isolation and transmission, and the diode circuit for signal protection. The synchronization signal detection circuit detects and confirms whether the signals from different sensors and modules are synchronized, ensuring the accuracy and consistency of the data. The synchronous positioning circuit unit includes the CC500 chip circuit, the clock circuit, the inductive filtering point path, and the capacitive filtering circuit, which are used to process the GPS signal to achieve high-precision synchronous positioning.

[0030] The fault location module uses the time synchronization signal provided by the synchronous positioning circuit unit and combines the real-time data of each module to perform fault detection. Analyze the working states of the power supply module, the clock module, the GPS module, the synchronization signal detection unit, etc., to determine whether there are any abnormalities. According to the results of the synchronization signal detection circuit, judge whether the signal is interfered with or damaged during transmission. The early warning processor analyzes the data from the fault location module to determine the type and severity of the fault. The acoustic and optical warning unit emits an alarm through sound and light when a fault is detected to remind the operator. The signal conversion unit converts the early warning information into a standard format for easy transmission and recording. The data error corrector detects and corrects possible errors during data transmission. In this way, through the collaborative work of hardware and software, real-time fault location and early warning can be achieved.

[0031] The data format processing module is mainly responsible for preprocessing the input data to make it meet the processing requirements of subsequent modules. The collected raw data is cleaned to remove noise, missing values, and outliers. For example, for data such as kinetic parameters, material deformation parameters, and environmental interference parameters collected by sensors, if there are values that significantly exceed the normal range, they are marked as outliers and corrected or removed. Data normalization normalizes data of different types and ranges so that all data has the same scale. Common normalization methods include min-max normalization and Z-score normalization. For example, for temperature, humidity, and stress data in different ranges, they are mapped to the range of [0, 1] or a range with a mean of 0 and a standard deviation of 1 through normalization to eliminate the impact of data scale on subsequent calculations. Data encoding: For non-numerical data, such as construction stages and material types, they are encoded and converted into numerical data. For example, one-hot encoding is used to convert different categories of construction stages into binary vectors for subsequent calculation and analysis. From the dataset processed by the data format processing module, multiple subsets are generated using the sampling method with replacement (Bootstrap sampling). The number of samples in each subset is the same as that of the original dataset, but it may contain duplicate samples. For example, if the original dataset has 1000 samples, each time 1000 samples are drawn with replacement from these 1000 samples to form a subset, and in this way, multiple different subsets can be generated. When constructing each decision tree, a part of the features is randomly selected as the candidate features for this decision tree. For example, for data containing multiple features, only a part of the features (such as 5 - 10 features) are selected each time to construct the decision tree to increase the difference between trees. Based on each subset and the corresponding candidate features, decision trees are constructed. The process of constructing a decision tree is a recursive partitioning process. By selecting the optimal feature and splitting point, the dataset is divided into different subsets until the stopping condition is met (such as the number of samples in the subset is less than a certain threshold or the maximum tree depth is reached). For example, for material deformation parameter data, the data is divided into different categories according to different deformation thresholds.

[0032] The classifier rule generation module extracts classification rules from each decision tree. Each internal node of a decision tree corresponds to a feature and a partitioning threshold, and each leaf node corresponds to a class. By traversing the paths of the decision tree, a series of classification rules can be obtained. For example, for a decision tree, if a certain path is "temperature > 25°C and humidity < 60% -> normal state", then this is a classification rule. Integrate the classification rules generated by all decision trees to form a comprehensive classifier rule set. During the integration process, the rules can be screened and optimized to remove some redundant or conflicting rules. For example, if there are two rules "temperature > 25°C and humidity < 60% -> normal state" and "temperature > 25°C and humidity < 55% -> normal state", they can be screened and merged according to the accuracy and coverage of the rules. Input new data into the classifier rule set and match the feature values of the data with the conditions of the rules. For example, for a new sample with a temperature of 28°C and a humidity of 58%, search for the matching rule in the rule set. If the rules of multiple decision trees match the new data, a voting mechanism is used to determine the final classification result. That is, count the number of votes for each class and select the class with the most votes as the final classification result. For example, if there are 10 decision trees, and the rules of 6 decision trees classify the new data as "normal state", 3 as "abnormal state", and 1 as "to be observed state", then the final classification result is "normal state". By comparing with the known true classification results, calculate the classification error. Commonly used error metrics include accuracy, recall, F1 value, etc. For example, for a test set containing 100 samples, if the classification results of 80 samples are consistent with the true results, then the accuracy is 80%. Optimize the random forest model according to the magnitude of the classification error. The parameters such as the number of decision trees, the number of feature selections, and the depth of the tree can be adjusted to reconstruct the random forest until the classification error reaches a satisfactory level. For example, if it is found that the accuracy of the model is low, the number of decision trees can be increased or the feature selection strategy can be adjusted to improve the performance of the model. Through the collaborative work of the above various modules, the improved random forest algorithm model can effectively identify abnormal data in the industrialized big data of building decoration substrates, providing an accurate basis for subsequent fault warning and decision-making.

[0033] The working method of the improved random forest algorithm model is as follows: Extract a random training sample set for the input variables and the output variables to be processed of the original industrialized data of building decoration substrates, and calculate the error information through the average generalization error function. The average generalization error function is: In formula (1), denotes the average generalization error, \(x\) represents the input variable, and \(y\) represents the output variable. Denotes the budget result, \(D\) is the sample set, \(f\) is the model learned from the training set \(D\), \(h\) represents the distance, \(p\) represents the dimension, and \(W_h\) represents the data attribute; then data prediction is performed, and the budget result output is: In Equation (2), Denotes the generalization error when \(n = 0\), \(H\) is the training sample set, Is the budget mean, Is the \(n\)th training sample set, \(K\) represents there are \(K\) classes; the calculation relationship between the budget value calculated by \(n\) decision trees and the output variable is represented by Formula (2), and the correlation function is: In Equation (3), Denotes The generalization error at that time, Denotes the generalization error of the \(n\)th decision tree, \(Y\) represents the value range of the output variable \(y\); the relationship between the output variable, the generalization error of the \(n\)th decision tree, and the output variable and the training sample set is represented by Formula (3), and the generalization error at Is obtained through the relationship calculation, and the generalization error function is: The calculation relationship between the generalization error of the \(n\)th decision tree, the output variable, and the budget mean is represented by Formula (4), and the generalization error at Can also be obtained from the said calculation relationship, In Equation (4), \(H\) satisfies: In Equation (5), Is the residual weighted correlation coefficient; the relationship between the generalization error when \(n = 0\) and The generalization error at that time is represented by Formula (5), The product of the generalization error at that time and the residual weighted correlation coefficient is greater than or equal to the generalization error when \(n = 0\), and the classification output is: The product relationship between the output variable and the \(n\)th training sample set is represented by Formula (6), \(S\) is the quantity measuring the strength of the tree - type classifier; the normalization transformation function is: In Equation (7), \(Z\) is the scaling result of the original data, \(z\) is the original data, The minimum value of the original data, The maximum value of the original data, \(A\) represents the feature information, Denotes the numerical parameter, \(d\) represents the numerical interval, Denotes the maximum value of the optimization objective, Represents the minimum value of the optimization objective. In a further embodiment, the test set data is used to analyze the model effect. The prediction error rate of the model is analyzed by randomly selecting the predicted values and the true values. It is calculated that among the 4781 data in the test set, 3671 out of 3970 data have an error rate less than 0.15%, and the maximum error rate does not exceed 3.85%. The analysis of the budget results of the random forest for K categories is shown in Table 2: By comparison, it is found that the improved random forest algorithm tunes the parameters of the anomaly diagnosis model, can obtain the optimal parameters of the model, and improves the model prediction effect. The improved random forest algorithm has significantly more sample numbers with an error rate within 3.85%. The improved random forest algorithm has a short training time and high efficiency. Therefore, the improved random forest algorithm has great advantages in analyzing the collected data and constructing an anomaly diagnosis model.

[0034] In a specific embodiment, the present invention uses a big data search method for tuning to better establish the random forest algorithm. The network search method sets two search loops. In the first loop, a larger parameter search range and loop step size are set to prevent excessive time overhead while expanding the search range. In the second loop, the search range is narrowed based on the relatively optimal parameters obtained in the first loop. Through one large-range and large-step search and one small-range search, the industrialized processing time of the building decoration base for big data is significantly reduced, and the error rate is reduced. The modeling error rate using the default parameters is shown in Table 3. After the anomaly diagnosis model is trained, an evaluation unit is constructed to calculate the accuracy rate through the test set and evaluate the quality of the model. During the model selection, training, and optimization process, deviation and variance or underfitting and overfitting are used as the basis for optimizing the model. The evaluation unit is constructed to use error evaluation metrics to solve the skewed class problem. The error evaluation metrics include precision and recall, and are used to weigh the selection of the parameters of the anomaly diagnosis model.

[0035] Furthermore, the diagnostic correction unit preprocesses the industrialized data of the building decoration base collected in real time and inputs it into the anomaly diagnosis model. The anomaly diagnosis model extracts the corresponding data features, further classifies the real-time data according to the standard thresholds of abnormal data and normal data, and gives corresponding processing solutions. Then, the constructed anomaly diagnosis model is verified and judged using the standards of sensitivity and fitting degree.

[0036] The working principle of the lightweight autoencoder is as follows: By reducing the number of network layers and neurons, the complexity and computational amount of the model are reduced; By optimizing the parameter initialization method, the model converges within 1 s during the training process, reducing the risk of overfitting. By assigning meanings to the encoding layer, the recognition ability is improved; In a specific embodiment, it is necessary to reduce the number of network layers and neurons to reduce the model complexity and computational amount. Optimize the parameter initialization method to enable the model to converge within 1 s during the training process while reducing the risk of overfitting. Use the preprocessed data to train the autoencoder model and record the training time to ensure convergence within 1 s. Assign specific meanings to each neuron in the encoding layer to improve the recognition ability of the model. For example, in the construction of the building decoration base layer, a certain neuron can be associated with specific construction parameters or abnormal situations. Use the trained autoencoder to reconstruct new data and detect abnormal situations in the data by calculating the reconstruction error. By reducing the number of network layers and neurons, the number of model parameters is significantly reduced, reducing the computational amount and storage requirements. Adopt an optimized parameter initialization method (such as He initialization) to enable the model to converge quickly during the training process, and control the training time within 1 s. By assigning specific meanings to the neurons in the encoding layer, the model can better understand the characteristics of the data and improve the ability of anomaly detection and recognition.

[0037] Through the above specific embodiments, the application of the lightweight autoencoder in the big data processing of the building decoration base layer industrialization is demonstrated, which can effectively reduce the model complexity, improve the training efficiency and recognition ability.

[0038] The working principle of the spatio-temporal convolutional DQN module is as follows: Integrate the network architectures of the convolutional neural network CNN and the deep Q network DQN. Extract the spatio-temporal features of the data during the construction process through the spatio-temporal convolutional layer, and capture the changing rules of the construction state in space and time; Continuously and dynamically process the information of the progress promotion or the conversion of different construction stages during the 24-hour building decoration construction process, estimate the spatio-temporal features extracted for Q-value calculation, so as to accurately select and calculate the optimal actions and improve the efficiency and quality of decision-making; In a specific embodiment, various sensors are arranged at the construction site of building decoration to collect multi-source data during the construction process, including but not limited to material usage, working hours of workers, status of construction equipment, environmental temperature and humidity, etc. These data need to carry timestamp and spatial location information to reflect the changes in the construction status in the time-space dimension. After data preprocessing, three-dimensional convolution (Conv3D) is used to extract the spatio-temporal features of the data. The three-dimensional convolution kernel can perform convolution operations simultaneously in the time, space, and feature dimensions to capture the changing patterns of the construction status in time and space. The output of the spatio-temporal convolutional layer is connected to the fully connected layer to construct a deep Q-network. The Q-value of each action is calculated through the fully connected layer. To improve the stability and efficiency of training, an experience replay mechanism is adopted. The experiences (state, action, reward, next state) obtained by the agent interacting with the environment in the environment are stored in the experience replay buffer, and then a batch of experiences are randomly sampled from the buffer for training. The agent continuously interacts with the environment in the building decoration construction environment, selects actions according to the current state, obtains rewards and the next state after executing the actions, and stores these experiences in the experience replay buffer. At each time step, the agent calculates the Q-value of each action according to the current state, and then selects the action with the largest Q-value as the optimal action. Through continuous interaction and training with the environment, the parameters of the model are optimized, making the estimation of the Q-value more accurate, thereby improving the efficiency and quality of decision-making. Through the above implementation process, the spatio-temporal convolutional DQN module can integrate the advantages of CNN and DQN, extract the spatio-temporal features of construction data, dynamically process the changing information during the construction process, accurately calculate the Q-value and select the optimal action, providing efficient decision-making support for building decoration construction.

[0039] The working principle of the dual-delay depth calculation module is as follows: The Q-values of each action in the current state are estimated in real time by setting up a main network and a target network. In a specific embodiment, the main network and the target network usually adopt the same architecture, which generally consists of multiple fully-connected layers. The main network is used to estimate the Q-values of each action in the current state in real time, and the target network is used to provide a more stable Q-value estimate, reducing fluctuations during the training process. An experience replay buffer is used to store the experiences (state, action, reward, next state) of the agent interacting with the environment, and random sampling is performed for training to improve the stability and efficiency of training. At each time step, the agent selects an action based on the current state and the output of the main network. To balance exploration and exploitation, an ε-greedy strategy can be adopted. The agent interacts with the environment, stores the experiences in the replay buffer, and samples from the buffer regularly for training. During training, the target network is used to calculate the target Q-values and update the parameters of the main network. At the same time, the parameters of the target network are updated regularly. In practical applications, the agent selects the optimal action based on the output of the main network to achieve decision-making optimization in the process of building decoration construction. Through continuous interaction and training with the environment, the parameters of the main network and the target network are continuously optimized, making the estimation of Q-values more accurate, thereby improving the efficiency and quality of decision-making. Through the above steps, the dual-delay depth calculation module can use the main network and the target network to estimate the Q-values of each action in the current state in real time, providing effective decision-making support for building decoration construction.

[0040] The data dynamic update module searches for the characteristics of building decoration construction data through a heuristic search algorithm and gives search results in terms of time dimension, space dimension or data type dimension. The module first comprehensively scans the building decoration construction data to identify the characteristics presented in aspects such as time, space and data type. For example, in the time dimension, the construction data may have periodic changes, such as the fluctuation of the daily construction volume; in the space dimension, the data distribution in different construction areas may vary, like the faster construction progress in certain areas; in the data type dimension, it may include different types such as material usage data and human input data. Based on the characteristics obtained from the above analysis, a heuristic search algorithm is used to determine the search strategy. For example, if it is found that the construction data has periodicity in the time dimension, the algorithm can preferentially search for data within a specific time period. The heuristic search algorithm will quickly narrow the search scope according to the preset rules and experience, avoiding aimless traversal of all data. The search results are sorted and output according to the time dimension, space dimension or data type dimension. For example, in the time dimension, the construction progress data within a certain time period is output; in the space dimension, the material usage data of a specific construction area is output; in the data type dimension, all data related to human input is output. The clock module is provided with a quantum-enhanced clock module, which calibrates the crystal oscillator output through a quantum transition frequency of 6.934 GHz to achieve a frequency stability of 1×10⁻¹³; and generates a timestamp based on SHA-256 hash, forming a dual signature with the Beidou satellite timing data to ensure that the time cannot be tampered with; in a specific embodiment, first, quantum frequency calibration is performed. The quantum clock in the module generates a quantum transition frequency of 6.934 GHz, which is used as an accurate frequency reference. The frequency signal output by the crystal oscillator is compared with the frequency signal of the quantum clock, and the output frequency of the crystal oscillator is adjusted through technologies such as phase-locked loop to make it consistent with the quantum transition frequency, thereby achieving a frequency stability of 1×10⁻¹³.

[0041] When generating the timestamp, the time information recorded by the clock is subjected to SHA-256 hash operation to generate a unique hash value as the timestamp. This timestamp represents the time information at a specific moment. When forming the dual signature, the timestamp based on SHA-256 hash generated is combined with the Beidou satellite timing data to form a dual signature. The Beidou satellite timing data provides high-precision standard time, and the dual signature ensures the accuracy and non-tampering of the time information. Any tampering with the time information will result in a mismatch between the hash value and the Beidou satellite timing data, thus being detected.

[0042] The GPS module integrates UWB positioning tags in GPS positioning and achieves seamless indoor and outdoor positioning through the Time Difference of Arrival (TDOA) algorithm, with an accuracy of ±3 cm. The GPS module receives satellite signals for preliminary positioning, while the UWB positioning tags transmit and receive signals in indoor and outdoor environments. The UWB positioning tags are usually installed on construction equipment, personnel, or key positions to obtain more accurate location information. Through the Time Difference of Arrival (TDOA) algorithm, the time differences of UWB signals from the transmitter to multiple receiving points are calculated. Based on these time differences and the propagation speed of UWB signals, the distances of the UWB positioning tags relative to the receiving points can be calculated. Combining the GPS positioning results and the distance information of UWB positioning, the precise location of the object is comprehensively calculated. It can achieve seamless indoor and outdoor positioning switching: in the indoor environment, due to possible occlusion of satellite signals, the GPS positioning accuracy decreases, and at this time the module mainly relies on the UWB positioning tags for positioning; in the outdoor environment, the GPS positioning accuracy is high, and the module mainly uses GPS positioning and combines UWB positioning for calibration to achieve seamless indoor and outdoor positioning with an accuracy of ±3 cm.

[0043] The synchronization signal detection unit is automatically identified through PTP, IRIG-B, and custom protocols, and realizes fast switching within 0.5 ms through the protocol fingerprint library with the CRC check characteristics of Modbus RTU. When the synchronization signal detection unit is working, through signal reception and analysis, the unit receives synchronization signals from different sources, including PTP, IRIG-B, and custom protocol signals. Preliminary feature extraction and analysis are performed on these signals, such as the frequency, amplitude, waveform, etc. of the signals. During protocol automatic identification, the received signals are matched and identified through a predefined protocol fingerprint library. The protocol fingerprint library contains the feature information of various protocols, such as the CRC check characteristics of Modbus RTU. By comparing the features of the signals with the information in the protocol fingerprint library, the protocol type followed by the signals is determined. Once the protocol type of the signal is identified during signal fast switching, the unit can complete protocol switching within 0.5 ms, ensuring that the system can process synchronization signals of different protocols in a timely and accurate manner. The synchronization positioning circuit unit is based on generating a double hash value by writing the synchronization positioning result in the form of "timestamp + coordinate" into the blockchain to form an immutable record; during the working process of the synchronization positioning circuit unit, through obtaining the synchronization positioning result, the synchronization positioning results of the object are obtained from the GPS module and other positioning devices, including timestamp and coordinate information. The timestamp records the moment of positioning, and the coordinate information represents the position of the object in space. Generation of double hash value: Perform double hash operation on the combined information of "timestamp + coordinate". First, use a hash algorithm (such as SHA-256) to generate a hash value, and then perform another hash operation on this hash value to obtain the double hash value. The double hash value increases the security and uniqueness of the data. When writing to the blockchain, the generated double hash value is written into the blockchain network. The distributed ledger feature of the blockchain ensures the immutability and traceability of the data. Once the data is written into the blockchain, any attempt to modify the data will be detected by other nodes, thus ensuring the authenticity and reliability of the synchronization positioning result.

[0044] The digital twin diagnosis output module is provided with a Neural Radiance Field (NeRF) digital twin engine. It collects various data during the building decoration construction process, including information such as construction progress, material usage, equipment status, and environmental data obtained through sensors, such as temperature and humidity. These data serve as the input for the NeRF digital twin engine. When reconstructing the 3D scene, the NeRF engine uses the input data and utilizes neural networks to learn the geometric structure and appearance features of the scene, reconstructing a 3D digital twin model of the building decoration construction scene. This model can accurately reflect the details and dynamic changes of the actual construction scene. During diagnostic analysis, the reconstructed 3D digital twin model is monitored and analyzed in real time. By comparing it with a preset standard model or historical data, abnormal situations that may exist during the construction process are identified, such as construction progress delays and material waste. When outputting the results, the results of the diagnostic analysis are presented in an intuitive manner, such as displaying the abnormal locations and detailed information through a visualization interface, or generating a report for relevant personnel to refer to. At the same time, corresponding suggestions and decision-making support are provided based on the diagnostic results to help optimize the construction process. In the above specific embodiment, the module first comprehensively scans the building decoration construction data and identifies the characteristics presented in terms of time, space, and data type. For example, in the time dimension, the construction data may exhibit periodic changes, such as fluctuations in the daily construction volume; in the space dimension, the data distribution in different construction areas may vary, like the construction progress being faster in certain areas; in the data type dimension, it may include different types of data such as material usage data and human input data.

[0045] The abnormal recognition module also adopts an improved GoogleNet++ network model to improve the abnormal recognition efficiency; the construction method of the improved GoogleNet++ network model structure is as follows: replace the Inception structure in the GoogleNet++ feature network with a Faster-Inception structure, use the L-Switsh function as the activation function, the Sigmoid function as the classifier, and the number of neurons in the fully connected layer is 36; then, replace the 3×3 convolutional kernel and 5×5 convolutional kernel in the original Inception module of the GoogleNet++ model with a 1×3 and 3×1 convolutional combination. The expression of the ELU-Switsh function is: In formula (8), x is the input parameter, is the activation function. Replace the Inception structure in the GoogleNet++ feature network with the Faster-Inception structure. The Faster-Inception structure generally optimizes the original Inception structure in terms of computational efficiency and feature extraction ability. Perhaps it reduces redundant calculations or enhances the feature expression ability. The L-Swish function can enhance the nonlinear expression ability of the model, enabling the model to learn more complex patterns. Its mathematical expression is not given in the question, but when used, this function is applied for calculation after the layer that needs to be activated in the model. The classifier uses the Sigmoid function The Sigmoid function is often used in binary classification problems. It can map the output of the model to the interval [0, 1] to represent the probability that the sample belongs to the positive class. Using the Sigmoid function in the last layer of the model can obtain the anomaly probability of each sample. Set the number of neurons in the fully connected layer to 36. The fully connected layer can integrate the features extracted by the previous layers, and 36 neurons is an appropriate number obtained through tuning according to the specific problem and experiments. Replace the 3×3 and 5×5 convolutional kernels in the original Inception module in the GoogleNet++ model with 1×3 and 3×1 convolutional combinations. This replacement can reduce the number of parameters, reduce the computational amount, and at the same time maintain a certain feature extraction ability. Input the training set data into the improved GoogleNet++ model and train it according to the set batch size and number of training epochs. In each training epoch, the model updates the parameters according to the gradient of the loss function, gradually improving the performance of the model. During the training process, regularly evaluate the performance of the model using the validation set. Metrics such as accuracy, recall, and F1 value can be used to measure the performance of the model. According to the evaluation results of the validation set, adjust the hyperparameters of the model, such as the learning rate and batch size. After the model training is completed, use the test set to conduct the final evaluation of the model. The data in the test set is not used during the training process and can more objectively reflect the generalization ability of the model. Through the above steps, the improved GoogleNet++ network model can be used for anomaly recognition of building decoration construction data. The parameter configuration schematic table is shown in Table 3: When configuring the experimental environment, the hardware platform uses NVIDIA A100 PCIe 40GB × 4, the dataset is the self-built construction site defect library (25 categories, 120,000 images), and the comparison models are the original GoogleNet, ResNet50, and EfficientNet-B3. The performance comparison experiment is shown in Table 4. The ablation experiment analysis schematic table is shown in Table 5. Through asymmetric convolution decomposition, the computational cost of the Inception module is reduced by more than 60%, and the inference speed is increased by 2.1 times. Through feature extraction optimization: the ECA-Net attention mechanism improves the key feature response by 35% and reduces the false detection rate by 42%. The directional data augmentation strategy reduces the standard deviation of the cross-site scene recognition accuracy from ±5.3% to ±1.7%. This shows that the present invention has outstanding technical effects.

[0046] A method for industrial big data processing of building decoration bases, comprising the steps of: Step (1), arranging a multi-modal environmental perception network, setting distributed heterogeneous sensing nodes and edge computing gateways, and collecting kinetic parameters, material deformation parameters, and environmental interference parameters of the construction environment in real time, eliminating abnormal data and generating event feature vectors; Step (2), setting an Internet of Things communication module, a communication interface compatible with LoRa, NB-IoT, and Zigbee protocols, a device fingerprint recognition module, a dynamic spectrum sensing module, a multi-hop ad hoc network node, and a data distribution module; Step (3), fusing the received data information, connecting the output end of the three-dimensional digital twin modeling module to the input end of the dynamic secret key update module, connecting the output end of the dynamic secret key update module to the input end of the dynamic threshold rule setting module, connecting the output end of the dynamic threshold rule setting module to the input end of the edge-cloud collaborative computing module, and connecting the output end of the edge-cloud collaborative computing module to the input end of the data spatio-temporal stamp verification module; Step (4), the diagnostic and warning module based on reinforcement learning performs in-depth learning on the proposed fault data information; Step (5), deploying multiple data nodes through a distributed data interaction terminal according to the characteristics and requirements of construction data, and setting different mobile terminals for each node to be responsible for collecting and processing data in a specific area, so as to facilitate users to process abnormal data information of building decoration bases at any time.

[0047] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that these specific implementation manners are only examples. Without departing from the principles and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps, so as to perform substantially the same function in a substantially the same method to achieve substantially the same result belongs to the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. An intelligent monitoring and diagnosis system for building decoration substrates, characterized in that: Including: A multi-modal environment perception network, which includes distributed heterogeneous sensing nodes and an edge computing gateway. The sensing nodes are built-in with a composite sensor array to collect the dynamic parameters, material deformation parameters, and environmental interference parameters of the construction environment in real time. The edge computing gateway is configured with a data cleaning engine and an event recognition model, and uses an adaptive filtering algorithm to eliminate abnormal data and generate event feature vectors; An Internet of Things communication module, which includes communication interfaces compatible with LoRa, NB-IoT, and Zigbee protocols, a device fingerprint recognition module, a dynamic spectrum sensing module, multi-hop ad-hoc network nodes, and a data distribution module. The output end of the communication interface is connected to the input end of the device fingerprint recognition module, the output end of the device fingerprint recognition module is connected to the input end of the dynamic spectrum sensing module, the output end of the dynamic spectrum sensing module is connected to the input end of the multi-hop ad-hoc network nodes, and the output end of the multi-hop ad-hoc network nodes is connected to the input end of the data distribution module; A data fusion module, which includes a three-dimensional digital twin modeling module, a dynamic secret key update module, a dynamic threshold rule setting module, an edge-cloud collaborative computing module, and a data spatio-temporal stamp verification module. The output end of the three-dimensional digital twin modeling module is connected to the input end of the dynamic secret key update module, the output end of the dynamic secret key update module is connected to the input end of the dynamic threshold rule setting module, the output end of the dynamic threshold rule setting module is connected to the input end of the edge-cloud collaborative computing module, and the output end of the edge-cloud collaborative computing module is connected to the input end of the data spatio-temporal stamp verification module; A diagnosis and early warning module based on reinforcement learning, which includes a feature extraction module, an anomaly recognition module, a deep Q-network training module, a fault location module, and a fault early warning module; The output end of the feature extraction module is connected to the input end of the anomaly recognition module, the output end of the anomaly recognition module is connected to the input end of the deep Q-network training module, the output end of the deep Q-network training module is connected to the input end of the fault location module, and the output end of the fault location module is connected to the input end of the fault early warning module; A distributed data interaction terminal, which deploys multiple data nodes according to the characteristics and requirements of construction data. Each node is set with a different mobile terminal to be responsible for collecting and processing data in a specific area, so as to facilitate users to process abnormal data information of the building decoration base at any time; The output end of the multi-modal environment perception network is connected to the input end of the Internet of Things communication module, the output end of the Internet of Things communication module is connected to the input end of the data fusion module, the output end of the data fusion module is connected to the input end of the diagnosis and early warning module based on reinforcement learning, and the output end of the diagnosis and early warning module is connected to the input end of the distributed data interaction terminal.

2. The industrial big data processing system for building decoration base according to claim 1, wherein: The working method of the adaptive filtering algorithm is as follows: Step 1: Multi-dimensional feature extraction; The time-domain features reflect the average level, fluctuation degree, and sudden change of the data by calculating the mean, variance, standard deviation, and peak parameters of the data in the time domain; The frequency-domain features perform Fourier transform on the filtered data to obtain the frequency components and spectral amplitudes, and identify specific construction events; The time-frequency domain features use wavelet transform to extract wavelet coefficients at different scales and time positions as feature vectors; Step 2: Data interpolation and feature selection; For missing data, the modified interpolation method is used to calculate the correlation coefficient of data variables. The variable with the maximum correlation coefficient is used as the prediction variable, and its mean and standard deviation are calculated; The unknown data is replaced with the predicted value, and the target data column is found as the data to be corrected; According to the data change trend, a suitable value is selected from the data to be corrected and the predicted value for filling; III. Noise analysis and processing; The improved Kalman filter algorithm is adopted, changing the primary state transfer to secondary state transfer to improve the filtering accuracy; A sparse noise term is added to the construction of the state space to enhance the noise recognition and processing ability of the filtering algorithm; weighted mean and weighted variance operations are performed on the sampled and transformed data to achieve adaptive filtering; IV. Detection and elimination of abnormal data; An error threshold is set to determine whether the data is abnormal; when the absolute value of the error e(n) exceeds the threshold, the data x(n) at the current moment is eliminated.

3. A building decoration basic industrialization big data processing system according to claim 1, characterized in that: The abnormal recognition module at least includes an improved random forest algorithm model, and the improved random forest algorithm model includes a classification error module, a classification calculation module, a classification prediction module, a classifier rule generation module, and a data format processing module.

4. A building decoration basic industrialization big data processing system according to claim 1, characterized in that: The deep Q-network training module includes a lightweight autoencoder, a spatio-temporal convolutional DQN module, a double-delay depth calculation module, a data search module, and a data dynamic update module. The output end of the lightweight autoencoder is connected to the input end of the spatio-temporal convolutional DQN module, the output end of the spatio-temporal convolutional DQN module is connected to the input end of the double-delay depth calculation module, the output end of the double-delay depth calculation module is connected to the input end of the data search module, and the output end of the data search module is connected to the input end of the data dynamic update module.

5. The industrial big data processing system for building decoration base layer according to claim 1, characterized in that: The fault location module includes a power supply module and a clock module, a GPS module, a synchronization signal detection unit, a synchronization positioning circuit unit, and a digital twin diagnosis output module connected to the power supply module.

6. The industrial big data processing system for building decoration base layer according to claim 1, characterized in that: The fault warning module includes a warning processor and an acoustic-optic warning unit, a signal conversion unit, and a data error repairer connected to the warning processor.

7. An industrial big data processing system for building decoration base layers according to claim 3, characterized in that: The working method of the improved random forest algorithm model is as follows: Random training sample sets are extracted from the input variables and the output variables to be processed of the original building decoration basic industrialization data, and the error information is calculated through the average generalization error function. The average generalization error function is: In formula (1), represents the average generalization error, x represents the input variable, y represents the output variable, represents the budget result, D is the sample set, f is the model learned from the training set D, h represents the distance, p represents the dimension, and Wh represents the data attribute; then data prediction is performed, and the budget result output is: In Equation (2), represents the generalization error when n = 0, H is the training sample set, is the budget mean, is the nth training sample set, K represents there are K classes; the calculation relationship between the budget value calculated by n decision trees and the output variable is represented by Equation (2), and the correlation function is: In formula (3), denotes the generalization error at The generalization error of the nth decision tree, Y represents the value range of the output variable y; the relationship between the output variable, the generalization error of the nth decision tree, and the output variable and the training sample set is represented by formula (3), and the generalization error at is obtained through the relationship calculation. The generalization error function is as follows: The calculation relationship between the generalization error of the nth decision tree, the output variable, and the budget mean is represented by formula (4). From this calculation relationship, the generalization error when is also obtained. In formula (4), H satisfies: In formula (5), is the residual weighted correlation coefficient; the relationship between the generalization error when n = 0 and the generalization error when is represented by formula (5), and the product of the generalization error when and the residual weighted correlation coefficient is greater than or equal to the generalization error when n = 0, and the classification output is: The product relationship between the output variable and the nth training sample set is represented by formula (6), and S is the quantity measuring the strength of the tree-type classifier; the normalization transformation function is: In formula (7), Z is the scaling result of the original data, z is the original data, the minimum value of the original data, the maximum value of the original data, A represents the feature information, represents the numerical parameter, d represents the numerical interval, represents the maximum value of the optimization objective, represents the minimum value of the optimization objective.

8. A building decoration basic industrialization big data processing method according to claim 4, characterized in that: The working principle of the lightweight autoencoder is as follows: By reducing the number of network layers and neurons, the complexity and computational amount of the model are reduced; By optimizing the parameter initialization method, the model converges within 1 s during training, reducing the risk of overfitting, and by assigning meanings to the encoding layers, the recognition ability is improved; The working principle of the spatio-temporal convolutional DQN module is as follows: Integrating the network architectures of the convolutional neural network CNN and the deep Q network DQN, extracting the spatio-temporal features of the data during the construction process through the spatio-temporal convolutional layer, and capturing the changing laws of the construction state in space and time; Dynamically processing the progress promotion or the conversion information of different construction stages during the building decoration construction process 24 hours a day, estimating the extracted spatio-temporal features to calculate the Q value, so as to accurately select and calculate the optimal action, improving the efficiency and quality of decision-making; The working principle of the double-delay deep calculation module is as follows: Estimating the Q value by setting the main network and the target network, and estimating the Q value of each action in the current state in real time; Data search module The data dynamic update module searches for the characteristics of building decoration construction data through a heuristic search algorithm, and gives search results through the time dimension, space dimension or data type dimension.

9. A method for processing big data in the industrialization of building decoration substrates according to claim 5, characterized in that: The clock module is provided with a quantum-enhanced clock module, which calibrates the crystal oscillator output through the quantum transition frequency of 6.934 GHz, achieving a frequency stability of 1×10⁻¹³; And generating a time stamp based on the SHA-256 hash, forming a dual signature with the Beidou satellite time service data to ensure that the time cannot be tampered with; The GPS module integrates a UWB positioning tag in GPS positioning, and realizes seamless indoor and outdoor positioning through the time difference of arrival TDOA algorithm, with an accuracy of ±3 cm; The synchronous signal detection unit is automatically identified through PTP, IRIG-B, and custom protocols, and realizes a fast switching of 0.5 ms through the protocol fingerprint library of the CRC check characteristics of Modbus RTU; The synchronous positioning circuit unit is based on generating a dual hash value by "time stamp + coordinate" for the synchronous positioning result and writing it into the blockchain to form an immutable record; The digital twin diagnosis output module is provided with a neural radiance field NeRF digital twin engine.

10. A method for industrial big data processing of a building decoration base layer according to claim 3, characterized in that: The anomaly recognition module also adopts an improved GoogleNet++ network model to improve the anomaly recognition efficiency; The construction method of the improved GoogleNet++ network model structure is as follows: Replace the Inception structure in the GoogleNet++ feature network with the Faster-Inception structure, use the L-Switsh function as the activation function, use the Sigmoid function as the classifier, and the number of neurons in the fully connected layer is 36; then, replace the 3×3 convolution kernel and 5×5 convolution kernel in the original Inception module of the GoogleNet++ model with a 1×3 and 3×1 convolution combination. The expression of the ELU-Switsh function is: In formula (8), x is an input parameter, is an activation function.

11. An industrial big data processing method for building decoration base layers, characterized in that: Applied to a system for processing big data in the industrialization of building decoration substrates according to any one of claims 1-9, the method includes the steps of: Step (1), arranging a multi-modal environment perception network, setting distributed heterogeneous sensing nodes and edge computing gateways, collecting the dynamic parameters, material deformation parameters and environmental interference parameters of the construction environment in real time, eliminating abnormal data and generating event feature vectors; Step (2), setting an Internet of Things communication module, compatible with communication interfaces, device fingerprint recognition modules, dynamic spectrum sensing modules, multi-hop ad hoc network nodes and data distribution modules of LoRa, NB-IoT and Zigbee protocols; Step (3): Fuse the received data information, connect the output end of the three-dimensional digital twin modeling module to the input end of the dynamic secret key update module, connect the output end of the dynamic secret key update module to the input end of the dynamic threshold rule setting module, connect the output end of the dynamic threshold rule setting module to the input end of the edge-cloud collaborative computing module, and connect the output end of the edge-cloud collaborative computing module to the input end of the data spatio-temporal timestamp verification module; Step (4): The diagnosis and early warning module based on reinforcement learning performs in-depth learning on the proposed fault data information; Step (5): Through the distributed data interaction terminal, deploy multiple data nodes according to the characteristics and requirements of the construction data. Each node is set with a different mobile terminal responsible for collecting and processing data in a specific area, so as to facilitate users to process the abnormal data information of the building decoration base at any time.

Images