A gas pipeline construction quality information management system

Abstract

The application discloses a kind of gas pipeline construction quality information management system, belong to gas pipeline construction quality monitoring technical field.The system includes: data perception module, for gathering construction operation log text, pipeline welding time sequence parameter, environmental sensor reading and material quality index and so on multi-modal heterogeneous data;Heterogeneous graph construction module is used to construct dynamic heterogeneous graph according to construction process node, geographical position node and material batch node;Dual-channel decoupling representation module is used to decompose node attribute into domain invariant feature and domain-specific feature by variational autoencoder;Contrast prediction module is used to predict the quality index deviation of each construction process node by time series contrast learning;Causal tracing module is used to map the deviation to a causal graph, and identify root cause procedures and material defects using intervention analysis;Adaptive control module is used to generate construction parameter adjustment instructions based on root cause and issue them to field control terminals.

Description

Technical Field

[0001] This invention relates to the field of gas pipeline construction quality monitoring technology, specifically to an information management system for gas pipeline construction quality. Background Technology

[0002] Quality control in gas pipeline construction involves the comprehensive processing of multi-source heterogeneous data, including multimodal information such as construction operation records, welding process timing parameters, on-site environmental monitoring data, and material quality indicators. Existing pipeline construction quality management solutions generally rely on manual records and single threshold judgment rules, evaluating overall construction compliance by randomly inspecting the quality of some welds after welding is completed. This post-inspection model has significant limitations in data utilization; text logs can only serve as traceability evidence and cannot quantify operational deviations. The processing of timing parameters such as welding current and voltage is mostly limited to monitoring statistical mean and extreme value ranges, lacking in-depth characterization of signal waveform distortion and frequency domain energy distribution. Monitoring data from different sources differ significantly in acquisition frequency, numerical dimensions, and semantic granularity. Existing systems struggle to organically integrate process flow, spatial location, and material correlations into a unified structured representation, resulting in prominent data silos and the inability to effectively capture complex interactions between processes. In the quality anomaly identification stage, conventional methods use empirical formulas or fixed thresholds to determine whether welding parameters exceed standards, only detecting single-point anomalies and failing to trace the deviation propagation path along the process chain. More importantly, existing methods lack the ability to infer the underlying causal mechanisms from observed deviations. Anomalies in welding parameters are often accompanied by confounding effects from various potential contributing factors, such as material defects, changes in environmental temperature and humidity, or errors in previous processes. Correlation analysis alone is insufficient to pinpoint the true root cause process and defect source. Once quality deviations are identified, on-site adjustments often rely on the individual experience of construction personnel for tentative parameter corrections, lacking an adaptive decision-making mechanism oriented towards dynamic environments and causal diagnosis. Based on these shortcomings, there is an urgent need to address the problem of heterogeneous representation and fusion of multimodal construction data, as well as the causal tracing and adaptive control of quality deviations. Summary of the Invention

[0003] The purpose of this invention is to provide an information management system for the construction quality of gas pipelines, which solves the problem of heterogeneous representation and fusion of multimodal construction data in the prior art, as well as the problem of causal tracing and adaptive control of quality deviations.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an information management system for the construction quality of gas pipelines, including a data perception module, a heterogeneous graph construction module, a dual-channel decoupled representation module, a comparison and prediction module, a causal tracing module, and an adaptive control module. This system organically integrates multimodal heterogeneous construction data into a dynamic heterogeneous graph, and utilizes decoupled representation learning and time-series comparison and prediction to achieve accurate prediction, root cause tracing, and adaptive adjustment of construction quality deviations, forming a closed-loop control from perception to decision-making to execution.

[0005] The data perception module collects multimodal heterogeneous data during gas pipeline construction and performs targeted preprocessing on each modality. For construction operation log text, the module performs word segmentation and part-of-speech tagging, extracting operation description tuples composed of actions, objects, and construction locations. For pipeline welding timing parameters, the module segments the data according to welding passes and uses a sliding window to divide the welding current and voltage sequences. The window length adaptively changes according to the welding speed. Within the window, statistical features such as skewness, kurtosis, and envelope spectral entropy are extracted to obtain a structured representation that reflects the stability of the welding process. Simultaneously, the module calculates the current stability and voltage fluctuation rate within each segment. For environmental sensor readings, the module removes outliers and constructs an environmental time series, extracting soil moisture content change trends and geothermal gradient characteristics. For material quality indicators, the module matches them with a pre-defined batch standard library to generate material compliance tags. Through these processes, the originally heterogeneous and sparse field data is transformed into high-quality, computable, standardized features, laying the foundation for subsequent graph construction and analysis.

[0006] The heterogeneous graph construction module builds a dynamic heterogeneous graph based on construction process nodes, geographical location nodes, material batch nodes, and multimodal heterogeneous data, mapping data from different modalities to node attributes. During construction, the module adds directed edges to construction process nodes according to the logical order in the construction plan, forming a process skeleton graph; it establishes spatial association edges between each construction process node and its corresponding geographical location node, with the geographical location node carrying encoded coordinate segment information; and it connects material batch nodes to construction process nodes using that batch of material, forming material-process edges. Simultaneously, the module embeds dynamic features within a time window into the time decay weights of the corresponding edges based on the timestamps of the multimodal data, enabling the graph structure to reflect the dynamic evolution of the construction process. For text modalities, the module uses a pre-trained lightweight word vector model to encode operation description tuples into dense vectors. For numerical time-series modalities, it extracts the energy proportion of each frequency band through multi-layer wavelet packet decomposition to form frequency domain attribute vectors, and transforms the spatiotemporal coordinates into low-dimensional embeddings based on geohashing encoding. This heterogeneous graph construction method effectively preserves multi-dimensional interactive information such as process logic, spatial relationships, and material traceability, enhancing the semantic richness and accuracy of subsequent analysis.

[0007] A dual-channel decoupled representation module performs decoupled representation learning on dynamic heterogeneous graphs. The module constructs a dedicated variational autoencoder for each node, outputting a latent variable distribution. From this distribution, domain-invariant features and domain-specific features are sampled and separated. The domain-invariant features follow a standard Gaussian prior, capturing quality impact patterns applicable across processes and environments. Domain-specific features, constrained by modal conditions, retain personalized information specific to particular construction scenarios. By decomposing node attributes into these two types of features, the system maintains stable quality analysis capabilities even when facing changes in the construction environment or differences in sensor configurations, significantly improving the model's generalization and robustness.

[0008] The contrastive prediction module predicts the deviation of quality indicators for each construction process node based on domain-invariant and domain-specific features through time-series contrastive learning. The module selects domain-invariant features of the same construction process node at consecutive time steps as positive sample pairs and domain-invariant features of different construction process nodes as negative sample pairs, constructing a contrastive loss function to enhance the consistency and causality of domain-invariant features over time. Domain-specific features are mapped to quality influence factors through a fully connected layer. The module fuses the domain-invariant features constrained by the contrastive loss with the quality influence factors, inputting the data into a time-recursive decoder to generate predicted quality indicator values ​​for future moments, and then calculates the deviation between the predicted and actual collected values. This design allows deviation prediction to not only rely on historical trends but also effectively integrate the immediate impact of specific modalities, improving the ability to capture minor quality fluctuations and sudden anomalies.

[0009] The causal attribution module maps quality indicator deviations to a predefined causal directed acyclic graph (DAG), using intervention analysis to identify the root causes of deviations in processes and material defects. The DAG template loaded by the module is based on structured information from historical construction quality accident reports. An initial causal framework is learned through a PC algorithm and then verified and corrected by domain experts, containing pre-assumed causal relationships between process nodes, material nodes, and quality indicator nodes. The module uses quality indicator deviations as an intervention variable and updates the conditional distribution of each node in the causal graph using an additive noise model. For each candidate root cause node, a do-operation is performed to calculate the average causal effect of the quality indicator deviation after intervention, and the nodes are sorted according to effect size, outputting the root cause process identifier and material defect type. Kernel density estimation is used in the intervention analysis to replace parameter distribution assumptions, responding to causal effect identification with small sample data and ensuring reliable diagnosis even in the early stages of limited construction data accumulation. This module advances quality anomaly detection from correlation analysis to causal inference, providing support for accurate elimination of potential hazards.

[0010] The adaptive control module, based on diagnosed root cause process and material defects, generates construction parameter adjustment instructions using a strategy gradient algorithm and sends them to the on-site control terminal. The module defines a construction state space, consisting of the current quality indicator deviation, environmental sensor readings, and material compliance markers; and an action space, including welding current adjustment, welding speed adjustment, and preheating temperature adjustment. Using a reduction in quality indicator deviation as a reward signal, the strategy network is trained using Monte Carlo strategy gradient update rules. In online application, the module outputs adjustment instructions that conform to process constraints based on the current construction state, sending them to the welding power supply control terminal and preheating equipment via an IoT protocol. Thus, quality deviation information is transformed into specific, executable parameter correction actions in real time, forming a fully automated closed loop from detection and root cause analysis to control, significantly reducing quality response delay and ensuring the consistency and reliability of gas pipeline construction.

[0011] Through the collaborative work of the above modules, this invention integrates multi-source heterogeneous data such as text, time series, and environment during the gas pipeline construction process into a dynamic heterogeneous graph framework. It improves prediction accuracy by using decoupled representation and time series comparative learning, and then uses causal inference to pinpoint the deep-seated root causes of quality defects. Finally, it drives the real-time optimization of construction parameters through policy gradient learning, realizing proactive perception, intelligent diagnosis, and adaptive control of construction quality. This effectively reduces reliance on human experience and post-event inspection, and significantly improves the construction quality management level of gas pipelines in complex environments.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Based on a dual-channel decoupled representation module, a dedicated variational autoencoder is constructed for each node in the dynamic heterogeneous graph. After outputting the latent variable distribution, the encoder samples and separates domain-invariant features that follow a standard Gaussian prior and domain-specific features constrained by modal conditions from the distribution. By constructing a contrastive loss function using domain-invariant features of the same construction process node at consecutive time steps as positive sample pairs and domain-invariant features of different construction process nodes as negative sample pairs, the semantic consistency of domain-invariant features in the temporal dimension is strengthened. The domain-specific features are then mapped to quality influence factors by a fully connected layer. The domain-invariant features constrained by the contrastive loss are fused with the quality influence factors and input into a temporal recursive decoder to generate predicted values ​​of quality indicators for future time moments, thereby calculating the deviation between the predicted values ​​and the actual collected values. This scheme removes interference information related to specific modalities and data acquisition conditions from the representation, enabling temporal contrastive prediction to accurately capture the evolution trend of process quality without being affected by sensor characteristics, recording formats, and environmental background noise, thus solving the problem of heterogeneous representation fusion of multimodal construction data. Based on the causal tracing module, a predefined directed acyclic graph template for construction quality causality is loaded. The deviation of quality indicators is used as an intervention variable. An additive noise model is used to update the conditional distribution of each node in the causal graph. A do-operation is performed on each candidate root cause node, and the average causal effect of the quality indicator deviation after intervention is calculated. The root cause process identifier and material defect type are output based on the magnitude of the average causal effect. This scheme replaces statistical correlation screening with intervention analysis under the framework of a structural causal model. It quantifies the causal contribution of each candidate factor in the pre-assumed causal relationship direction between construction process nodes, material nodes, and quality indicator nodes, enabling the tracing results of quality deviations to trace back from apparent parameter anomalies to fundamental defects at the construction operation and material batch levels. In the on-site adjustment phase, the strategy network uses a reduction in quality indicator deviation as a reward signal. Based on the root cause diagnosis results and the current construction status, it generates construction parameter adjustment instructions in the action space, including adjustments to welding current, welding speed, and preheating temperature. These instructions are then sent to the execution equipment via an IoT protocol, thus forming an adaptive decision-making link from quality deviation identification and causal localization to closed-loop optimization of process parameters. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a schematic diagram of the structure of the gas pipeline construction quality information management system; Figure 2 This is a flowchart of the construction quality cause-and-effect tracing module. Figure 3 This is a flowchart of the closed-loop control process of the adaptive control module. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] See Figure 1 This invention provides an information management system for the construction quality of gas pipelines, comprising: a data perception module, a heterogeneous graph construction module, a dual-channel decoupling representation module, a comparison and prediction module, a causal tracing module, and an adaptive control module. The data perception module collects multimodal heterogeneous data during the gas pipeline construction process, including construction operation log text, pipeline welding timing parameters, environmental sensor readings, and material quality indicators. The heterogeneous graph construction module constructs a dynamic heterogeneous graph based on construction process nodes, geographical location nodes, material batch nodes, and multimodal heterogeneous data, mapping different modal data to node attributes. The dual-channel decoupling representation module performs decoupling representation learning on the dynamic heterogeneous graph, decomposing node attributes into domain-invariant features and domain-specific features using a variational autoencoder. The comparison and prediction module predicts the deviation of quality indicators for each construction process node based on domain-invariant and domain-specific features through time-series comparison learning. The causal tracing module maps the quality indicator deviation to a causal directed acyclic graph, using intervention analysis to identify the root causes of deviations in processes and material defects. The adaptive control module is used to generate construction parameter adjustment instructions based on root cause processes and material defects through a strategy gradient algorithm, and then send them to the on-site control terminal.

[0017] Example 1 In practical implementation, the data perception module collects multimodal heterogeneous data during the gas pipeline construction process. This data includes construction operation log text, pipeline welding timing parameters, environmental sensor readings, and material quality indicators. For the construction operation log text, the data perception module uses natural language processing tools to perform word segmentation and part-of-speech tagging. Operation description tuples are extracted from the segmented and tagged text sequence. These tuples include actions, objects, and construction locations. Word segmentation uses a maximum matching algorithm based on a domain dictionary, and part-of-speech tagging is implemented using a hidden Markov model. Actions correspond to verbs or verb phrases representing construction operations in the text; objects correspond to the pipe material, welding material, or equipment represented by the direct object of the verb; and construction locations correspond to location nouns or codes indicating pipeline sections. The extracted operation description tuples are stored as triples, each with a timestamp and a construction team identifier.

[0018] For pipeline welding timing parameters, the data sensing module collects welding current and welding voltage timing data from the welding power source and wire feeding mechanism. First, the data sensing module segments the welding current and welding voltage timing data according to the welding passes specified in the welding procedure. Within each segment, the data sensing module calculates current stability and voltage fluctuation rate. Current stability is defined as the absolute value of the ratio of the standard deviation to the mean of the welding current timing data within that segment; voltage fluctuation rate is defined as the absolute value of the ratio of the range to the mean of the welding voltage timing data within that segment. Simultaneously, the data sensing module uses a sliding window to further segment the welding current and welding voltage sequences within each segment to extract statistical features. The length of the sliding window adaptively changes according to the welding speed; specifically, the sliding window length L is determined by the following formula: ; Where L represents the sliding window length in the number of sampling points, and v represents the average welding speed within the current segment in millimeters per second. This represents the preset time window constant, selected based on the molten pool stabilization time set in the pipeline welding process qualification, with a value of 2 seconds. Within each sliding window, the data sensing module extracts skewness, kurtosis, and envelope spectral entropy as statistical features. Skewness is the ratio of the third central moment of the data sequence within the window to the cube of the standard deviation; kurtosis is the ratio of the fourth central moment of the data sequence within the window to the fourth power of the standard deviation; the calculation process for envelope spectral entropy is as follows: first, perform a Hilbert transform on the data sequence within the window to obtain the envelope signal; then, perform a Fourier transform on the envelope signal to obtain the envelope spectrum; divide the envelope spectrum into K equal-width frequency bands, where K is set to 16; calculate the energy in each frequency band and divide by the total energy to obtain the normalized energy. Where k is the frequency band number, and the value of k ranges from 1 to K, the envelope spectral entropy is then calculated using the information entropy formula. The data perception module concatenates the skewness, kurtosis, and envelope spectral entropy extracted within each sliding window into a three-dimensional statistical feature vector, and performs min-max normalization on the three-dimensional statistical feature vector. The normalized three-dimensional statistical feature vector serves as a structured representation of the pipeline welding timing parameters.

[0019] For environmental sensor readings, the data sensing module acquires readings from soil temperature and humidity sensors and surface meteorological sensors buried around the trench. The module uses the Grubbs criterion to remove outliers from the environmental sensor readings, considering readings outside the confidence interval as outliers and replacing them with linear interpolation of the preceding and following normal readings to construct an environmental time series. From the environmental time series, the module selects a continuous soil moisture content series and uses moving linear regression to calculate the soil moisture content change trend, represented by the regression slope. From the environmental time series, it selects ground temperature series at different depths, calculates the difference between adjacent depth ground temperature series, divides by the depth difference, and obtains the ground temperature gradient characteristic, represented by the temperature change per meter.

[0020] For material quality indicators, the data sensing module obtains these indicators from the material inspection report. These indicators include the pipe's tensile strength, yield strength, elongation, and the percentage content of the chemical composition of the welding materials. The data sensing module then matches these material quality indicators with a pre-defined batch standard library. This library stores the upper and lower limits of standard technical indicators for each grade of pipe and welding material. The matching process involves comparing each parameter value in the material quality indicators with the allowable range of the corresponding grade's standard parameters in the batch standard library. If all parameters fall within the allowable range, a material compliance mark of "qualified" is generated; if at least one parameter exceeds the allowable range, a material compliance mark of "unqualified" is generated, and the exceeding parameter and its value are recorded.

[0021] Example 2 In practical implementation, the heterogeneous graph construction module acquires the logical sequence of construction procedures specified in the construction plan and the multimodal heterogeneous data output by the data perception module. This multimodal heterogeneous data includes construction operation log text, pipeline welding timing parameters, environmental sensor readings, and material quality indicators. The heterogeneous graph construction module constructs directed edges for each construction procedure node according to the logical sequence in the construction plan, forming a procedure skeleton graph. Each construction procedure node uses a procedure identifier as a unique index, which is composed of the project number, section number, and procedure sequence number. The directed edges point from the preceding procedure node to the subsequent procedure node, and the type of the directed edges is marked as "sequential execution." The heterogeneous graph construction module attaches a procedure attribute vector to each construction procedure node. This vector is composed of the procedure type code, planned start time, planned end time, and construction unit code.

[0022] The heterogeneous graph construction module adds spatial association edges to each construction process node and its corresponding geographical location node. The geographical location node uses geohashing encoding as a unique index and carries encoded coordinate segment information. This coordinate segment information is obtained by dividing the construction area into equal-length segments according to the pipeline route, with each segment being 10 meters long and corresponding to one geographical location node. The encoded coordinate segment information in the geographical location node is obtained by processing the geodetic coordinates of the segment's center point using a geohashing encoding algorithm. Specifically, the geohashing encoding algorithm approximates the longitude and latitude values ​​of the geodetic coordinates using interval binary search, alternately generating binary bit strings, and then converting these binary bit strings into 32-bit strings with a length of 8 bits. Spatial association edges point from the construction process node to the corresponding geographical location node, and the type of the spatial association edge is marked as "located at".

[0023] The heterogeneous graph construction module connects material batch nodes with construction process nodes that use the batch of materials, forming material-process edges. Material batch nodes use the material batch number as a unique index, which consists of the manufacturer code, production date, and furnace batch number. Material-process edges point from material batch nodes to construction process nodes, and their type is labeled "consumption." The heterogeneous graph construction module adds a material attribute vector to each material batch node. This vector is composed of material compliance markers, parameter values ​​from material quality indicators, and material grade codes.

[0024] The heterogeneous graph construction module embeds the time decay weights of corresponding edges into the dynamic features within a time window based on the timestamps of the multimodal heterogeneous data. The time window length is set to 24 hours. For each directed edge in the process skeleton graph, the heterogeneous graph construction module obtains the actual start time of the subsequent construction process node connected by the directed edge as a reference time point and calculates the time interval between the reference time point and the current system time. The time decay weights are calculated as follows: ; in, This represents the time decay weight of the directed edge e. The value ranges from 0 to 1; This represents the time interval between the reference time point of directed edge e and the current system time, in hours; This represents the preset decay time constant, determined based on the statistical value of the average duration of the construction process, and is set to 48 hours. The time decay weights of spatially associated edges and material-process edges are calculated in the same way. The reference time point for spatially associated edges is the actual start time of the associated construction process node, and the reference time point for material-process edges is the time when the batch of materials in the associated construction process node is requisitioned.

[0025] When the heterogeneous graph construction module maps different modalities of data to node attributes, for text modal data, a pre-trained lightweight word vector model is used to encode the operation description tuples extracted by the data awareness module into dense vectors. The lightweight word vector model adopts the FastText model architecture, with a word vector output dimension of 100. The encoding process involves concatenating the three text fields (action, object, and construction location) from the operation description tuple into a text sequence, then obtaining the word vector for each word in the text sequence using the FastText model, and finally performing average pooling on all word vectors to obtain the dense vector representation of the operation description tuple. The pre-trained lightweight word vector model was obtained through unsupervised pre-training on a text corpus in the construction and installation domain, and the model parameters remain unchanged after system deployment.

[0026] For numerical time-series modal data, the heterogeneous graph construction module extracts the energy proportion of each frequency band through multi-level wavelet packet decomposition, forming a frequency domain attribute vector. The multi-level wavelet packet decomposition uses the Daubechies4 wavelet basis, with a decomposition level of 3. For the structured representation sequence of pipeline welding time-series parameters, the heterogeneous graph construction module performs 3-level wavelet packet decomposition to obtain the coefficients of 8 terminal nodes. The energy of each terminal node is calculated and divided by the sum of the energies of all terminal nodes to obtain an 8-dimensional frequency band energy proportion vector, which serves as the frequency domain attribute vector.

[0027] The heterogeneous graph construction module transforms the spatiotemporal coordinates of construction process nodes into low-dimensional embeddings based on geohashing encoding. The spatiotemporal coordinates include the spatial and temporal coordinates of the construction process node. The spatial coordinates are taken from the geodetic coordinates of the corresponding geographic location node, and the temporal coordinates are taken from the planned start time of the construction process node, converted to a Unix timestamp. The heterogeneous graph construction module performs geohashing encoding on the spatial coordinates to obtain a spatial hash string, and performs time interval encoding on the temporal coordinates to obtain a temporal hash string. The spatial hash string and the temporal hash string are concatenated and mapped to a 16-dimensional low-dimensional embedding vector through a fully connected layer. The activation function of the fully connected layer is the ReLU function. Finally, the complete node attributes of a construction process node are composed of a process attribute vector, a frequency domain attribute vector, a dense vector of operation description tuples, a material attribute vector, and a low-dimensional embedding vector.

[0028] Example 3 In practical implementation, the dual-channel decoupled characterization module receives a dynamic heterogeneous graph output by the heterogeneous graph construction module. This dynamic heterogeneous graph includes construction process nodes, geographic location nodes, and material batch nodes, as well as directed edges between these nodes. The dual-channel decoupled characterization module constructs a dedicated variational autoencoder for each type of node in the dynamic heterogeneous graph. The construction process nodes correspond to the first variational autoencoder, the geographic location nodes to the second variational autoencoder, and the material batch nodes to the third variational autoencoder. Each dedicated variational autoencoder has the same network architecture but independent network parameters.

[0029] Taking the first variational autoencoder as an example, it consists of an encoder and a decoder. The encoder's input is the complete node attribute vector of the construction process node. This complete node attribute vector is composed of a process attribute vector, a frequency domain attribute vector, a dense vector of operation description tuples, a material attribute vector, and a low-dimensional embedding vector. The encoder adopts a three-layer fully connected network structure. The input dimension of the first fully connected layer is equal to the dimension of the complete node attribute vector of the construction process node, and the output dimension is 256. The activation function is LeakyReLU. The input dimension of the second fully connected layer is 256, and the output dimension is 128. The activation function is LeakyReLU. The input dimension of the third fully connected layer is 128, and the output dimension is 64. The activation function is linear activation. The encoder outputs two 64-dimensional parameter vectors: one parameter vector represents the mean of the latent variable distribution, and the other parameter vector represents the log-variance of the latent variable distribution. When sampling from the latent variable distribution, the first variational autoencoder uses a reparameterization technique to sample a 64-dimensional random noise vector from the standard Gaussian distribution. The random noise vector is multiplied by the result of taking the exponentiation of each element of the log-variance parameter vector, and then added to the mean parameter vector to obtain the sampled latent variables.

[0030] The decoder structure of the first variational autoencoder is symmetrical to that of the encoder. The decoder takes 64-dimensional latent variables as input and gradually restores them to the original input dimensions through three fully connected layers. The output dimensions of the three fully connected layers are 128-dimensional, 256-dimensional, and the complete node attribute vector dimension of the construction process node, respectively. The activation function of the first two layers is the LeakyReLU function, and the activation function of the last layer is linear activation.

[0031] The second variational autoencoder takes the attribute vector of the geographic location node as its encoder input. This attribute vector is constructed by concatenating a 16-dimensional low-dimensional embedding vector mapped by geohashing with the feature vector of environmental sensor readings. The encoder structure is the same as that of the first variational autoencoder, but its parameters are independent. The third variational autoencoder takes the attribute vector of the material batch node as its encoder input. This attribute vector is constructed by concatenating the material attribute vector and the standardized vectors of the parameter values ​​in the material quality index. The encoder structure is the same as that of the first variational autoencoder, but its parameters are independent.

[0032] After sampling from the latent variable distribution, the dual-channel decoupled representation module separates the sampled latent variables into domain-invariant features that follow a standard Gaussian prior and domain-specific features constrained by modal conditions. The separation method involves dividing the 64-dimensional latent variables into two equal parts along the dimensional direction, with the first 32 dimensions serving as domain-invariant features. The latter 32 dimensions serve as domain-specific features. Modal constraints on domain-specific features are implemented by inputting the domain-specific features into a modality classifier. The modality classifier consists of a two-layer fully connected network with a 32-dimensional input, a 16-dimensional hidden layer, and a 3-dimensional output, corresponding to three node types. The output of the modality classifier is used to assist the modality classification loss term in the loss function.

[0033] In the dual-channel decoupled representation module, the training loss function for all variational autoencoders consists of three parts: the first part is the reconstruction loss of the variational autoencoder, using the mean squared error loss function to measure the difference between the decoder output and the original input; the second part is the KL divergence loss between the latent variable distribution and the standard Gaussian prior distribution, constraining domain-invariant features to follow the standard Gaussian prior; the third part is the cross-entropy loss of the modality classifier, reinforcing the correlation between domain-specific features and node modalities. The weights of the three loss parts are set to 1.0, 0.5, and 0.3, respectively. The variational autoencoder is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 training epochs. Training is terminated early when the validation loss no longer decreases after 20 consecutive epochs.

[0034] After training the dual-channel decoupled representation module, the contrast prediction module uses extracted domain-invariant and domain-specific features to predict the deviation of quality indicators. The contrast prediction module selects domain-invariant features of the same construction process node at consecutive time steps as positive sample pairs and domain-invariant features of different construction process nodes as negative sample pairs to construct a contrastive loss function. A consecutive time step refers to two adjacent construction data acquisition cycles, with a time interval of 4 hours. The domain-invariant feature of the same construction process node at time step t is represented as... The domain-invariant feature at time step t+1 is represented as A positive sample pair consists of two samples. Negative sample pairs are randomly selected from the domain-invariant features of different construction process nodes at any time step, with the number of negative sample pairs selected each time set to twice the number of positive sample pairs. The contrastive loss function is calculated using cosine similarity. The cosine similarity between positive sample pairs is calculated and the negative logarithm is taken, and the cosine similarity between negative sample pairs is also calculated and the negative logarithm is taken. The contrastive loss function tends to increase the similarity of positive sample pairs and decrease the similarity of negative sample pairs.

[0035] The contrast prediction module maps domain-specific features to quality influence factors through fully connected layers. The mapping network consists of a single fully connected layer with 32 input dimensions and 8 output dimensions. The activation function is the Sigmoid function, and the output 8-dimensional vector is the quality influence factor. Each dimension of the quality influence factor takes a value between 0 and 1, representing the degree of influence of the domain-specific feature on the corresponding quality indicator dimension.

[0036] The contrast prediction module fuses the domain-invariant features under the contrast loss constraint with the quality influence factor. The fusion method involves concatenating the domain-invariant features and the quality influence factor along their dimensional axes, resulting in a 40-dimensional fused feature vector. This fused feature vector is input to a temporal recursive decoder, which employs a gated recurrent unit (GRU) network structure. The GRU network contains one hidden layer with 64 neurons. At each time step, the GRU network receives the fused feature vector from the current time step and outputs the predicted value of the domain-invariant features for the next time step. The output of the GRU network at the last time step is mapped to the predicted value of the quality index through a fully connected mapping layer. The fully connected mapping layer has a 64-dimensional input and an output dimension equal to the dimension of the quality index. The dimension of the quality index is determined based on the number of actual monitored quality indicators, and is typically 6-dimensional, corresponding to weld reinforcement height, weld width, undercut depth, number of pores, unfused length, and post-weld pipe deformation. The contrast prediction module calculates the deviation between the predicted quality index value and the actually collected quality index. The deviation is calculated using Euclidean distance, with the following formula: ; in, This represents the deviation of the quality indicator, where M represents the number of dimensions of the quality indicator (M takes a value of 6), and m represents the dimension index of the quality indicator (m ranges from 1 to M). This represents the actual quality index value collected in the m-th dimension. This represents the predicted value of the quality indicator in the m-th dimension. Quality indicator deviation. It reflects the degree of deviation between the current construction quality and the expected construction quality at each construction process node. The larger the deviation value, the more serious the deviation of the construction quality from the expectation.

[0037] Example 4 In specific implementation, please refer to Figure 2 The causal tracing module loads a predefined directed acyclic graph (DAG) template for construction quality causality. The generation of this template is completed offline before system deployment. First, the generation process retrieves historical construction quality accident reports, extracting structured information including the names of the work processes involved, material batches, quality index deviation types, and expert-annotated causal relationships. The generation process converts work process names into work process nodes, material batches into material nodes, and quality index deviation types into quality index nodes. Using these nodes as variables, the generation process employs a PC algorithm to learn the initial causal skeleton. The PC algorithm starts with a fully connected undirected graph and gradually removes edges through conditional independence tests. The conditional independence test uses a partial correlation coefficient combined with Fisher's z-test, with a significance level set to 0.05. The initial causal skeleton output by the PC algorithm contains undirected edges and some directed edges. The generation process submits an initial causal skeleton to domain experts, who determine the direction of the undirected edges. Edges that clearly violate physical laws or construction procedures are deleted or reversed. After verification and correction by domain experts, a causal directed acyclic graph template for construction quality is finalized. This template contains pre-assumed causal relationships between process nodes, material nodes, and quality indicator nodes. The causal relationship direction points from the parent node to the child node, indicating that the execution status of the process or material properties represented by the parent node directly affects the quality indicator or subsequent processes represented by the child node.

[0038] During the online operation phase of the system, the causal tracing module uses the deviation of the quality index output by the comparison prediction module as input for intervention analysis. The causal tracing module first identifies a set of candidate root cause nodes, which includes all process nodes involved in the current construction task and all material batch nodes marked as "unqualified" for material compliance. For each candidate root cause node, the causal tracing module updates the conditional distribution of each node in the directed acyclic graph of construction quality causality using an additive noise model. The additive noise model assumes that the value of each node is generated by superimposing independent noise onto the values ​​of its parent node set through a nonlinear function. The nonlinear function is fitted using a Gaussian process regression model, with a radial basis function kernel (RBM) selected. The length scale parameter of the RBM is set to 2.0, and the kernel variance parameter is set to 1.0. The distribution of independent noise does not adopt a parametric distribution assumption; the causal tracing module models independent noise using kernel density estimation, employing a Gaussian kernel function, with bandwidth calculated using the Silverman rule. The conditional distribution update process is as follows: For each node in the causal directed acyclic graph of construction quality, data samples of the node and all its parent nodes during the historical normal construction period are collected. These data samples are used to train the Gaussian process regression model in the additive noise model. After training, the actual observed values ​​of the node's parent nodes are input into the Gaussian process regression model to obtain the predicted mean. The residual between the actual observed values ​​of the node and the predicted mean is used as the noise sample. The noise sample is used to perform kernel density estimation to obtain the noise distribution.

[0039] The causal attribution module performs a do-operation on each candidate root cause node to calculate the average causal effect of the deviation of quality indicators after intervention. For each candidate root cause node... ,implement The operation represents the candidate root cause nodes. The value is set to the value corresponding to the abnormal state. The value is the candidate root cause node. Observations under abnormal conditions; and execute simultaneously. The operation represents the candidate root cause nodes. The value is set to the reference value corresponding to the normal state. The value is the candidate root cause node. The historical average value under normal conditions. When performing the do-operation, the construction quality causal directed acyclic graph points to candidate root cause nodes. All directed edges are cut off, candidate root cause nodes The conditional distribution of the first node degenerates into a fixed value, while the conditional distributions of the remaining nodes remain unchanged. Based on this, the causal origination module uses the updated additive noise model to sample and generate values ​​for each node sequentially along the directed edges, starting from the root node of the causal directed acyclic graph. Finally, it reaches the quality index node and produces the post-intervention distribution sample of the quality index deviation Y. The number of generated samples is set to 1000. The causal origination module calculates the values ​​of each node in the root node and the value of the remaining nodes. and The expected value of the deviation Y of the quality index under the two interventions, and the formula for calculating the average causal effect are as follows: ; in, Indicates candidate root cause nodes The average causal effect on the deviation of quality indicators, where S represents the number of samples, with a value of 1000, and s represents the sample number, with a value ranging from 1 to S. Indicates in The sample value of the deviation of the quality index obtained from the s-th sampling under intervention. Indicates in The sample value of the deviation of the quality index obtained from the s-th sampling under intervention.

[0040] After calculating the average causal effect for all candidate root cause nodes, the causal tracing module sorts them in descending order based on the magnitude of the average causal effect. It then selects the top K candidate root cause nodes with the largest average causal effect from the sorted results. The value of K is set to 3 according to the construction response strategy. The causal tracing module outputs the root cause process identifier and material defect type. The root cause process identifier is the process identifier of the candidate root cause node belonging to the process node in the sorted results. The material defect type is the non-conforming parameter item and the out-of-standard information in the material compliance mark corresponding to the candidate root cause node belonging to the material node in the sorted results.

[0041] Example 5 In specific implementation, please refer to Figure 3 The adaptive control module receives the quality index deviation from the comparison and prediction module, environmental sensor readings and material compliance markers from the data perception module, and root cause process identifiers and material defect types from the causal tracing module. The adaptive control module defines a construction state space, which consists of the current quality index deviation, environmental sensor readings, and material compliance markers. The current quality index deviation is a scalar value obtained from the comparison and prediction module, denoted as... , Updated once per construction data acquisition cycle, with a cycle interval of 4 hours. Environmental sensor readings include soil moisture content change trends, ground temperature gradient characteristics, atmospheric temperature readings, and atmospheric humidity readings. These readings form a 4-dimensional continuous vector, with each dimension mapped to the interval [-1, 1] using a min-max normalization method. Material compliance is marked as a discrete variable, taking values ​​of 0 or 1, where 0 corresponds to "qualified" and 1 corresponds to "unqualified." At each construction state construction moment, the adaptive control module adjusts the deviation of the current quality index. The normalized 4D environmental sensor reading vector and the material compliance label value are concatenated into a 6D construction state vector, which serves as the input to the strategy network.

[0042] The adaptive control module defines an action space, which includes adjustments for welding current, welding speed, and preheating temperature. The welding current adjustment is in amperes, with a range of [-15, 15] amperes, determined based on the allowable current fluctuation limits in the welding procedure specification. The welding speed adjustment is in millimeters per second, with a range of [-1.5, 1.5] millimeters per second, determined based on the response capability of the wire feeding mechanism and weld formation requirements. The preheating temperature adjustment is in degrees Celsius, with a range of [-25, 25] degrees Celsius, determined based on the power limitations of the preheating equipment and the safety margin of the pipe material's phase transformation point. The action space is a three-dimensional continuous vector, and the value ranges of each dimension constitute the boundary constraints of the actions.

[0043] The adaptive control module uses a reduction in the deviation of quality indicators as a reward signal. The reward signal is calculated based on the change in the deviation of quality indicators between two adjacent construction states. Specifically, the calculation method is as follows: the deviation of the quality indicators in the previous construction state is denoted as... The deviation of the quality index of the current construction status is recorded as The immediate reward for the current action. Defined as: ; in, Indicates the instant reward value; This indicates the deviation of quality indicators obtained under the previous construction condition, with the unit being the Euclidean distance dimension of the deviation. This indicates the deviation of the quality index recalculated after adjustments are made under the current construction condition, expressed in units of [values]. Same dimensions; This is a tiny constant set to avoid division by zero errors, and its value is... .when and When they are completely equal, The value is 0. When the deviation of the quality indicator decreases, When the value is positive, the policy network receives a positive incentive; when the deviation of the quality metric increases, If the result is negative, the policy network is penalized. The adaptive control module creates an interaction sample consisting of the construction state, the action performed, and the immediate reward obtained in each acquisition cycle, which is used for offline training of the policy network.

[0044] The policy network is trained using the Monte Carlo policy gradient update rule. The network structure consists of an input layer, hidden layers, and an output layer. The input layer receives a 6-dimensional construction state vector. The hidden layers contain two fully connected layers: the first fully connected hidden layer has a 6-dimensional input and a 64-dimensional output, with the tanh activation function; the second fully connected hidden layer has a 64-dimensional input and output, with the tanh activation function. The output layer consists of two branches: the mean branch outputs a 3-dimensional mean vector corresponding to the 3D action, with linear activation; the standard deviation branch outputs a 3-dimensional logarithmic standard deviation vector, which is first linearly converted to a 3-dimensional logarithmic standard deviation vector using an exponential function to ensure the standard deviation is always positive. The policy network models the action distribution as a Gaussian distribution with independent dimensions. During the training phase, the adaptive control module samples a batch of complete construction trajectories from historical interaction data of actual construction. Each construction trajectory contains the state sequence, action sequence, and reward sequence of all B interaction steps from the start to the end of a construction task. Using the Monte Carlo policy gradient update rule, the parameters of the policy network are... The update direction in the nth iteration is determined by the gradient ascent term, and the gradient is calculated as follows: ; in, The objective function for policy optimization is represented as follows. Policy network parameters gradient, It is a parameter vector formed by concatenating the weight matrices and bias vectors of all fully connected layers in the policy network; B represents the number of interaction steps contained in a construction trajectory. The value of B is determined by the duration of each construction task and the collection period, and is a variable positive integer; t represents the time step index in the trajectory, and the value of t ranges from 1 to B. This represents the cumulative discount reward from time step t to the end of the trajectory. The calculation method is as follows Where T is the total number of steps on the trajectory, i.e., B. It is the future step offset. It is a discount factor with a value of 0.92. This value is set according to the delay of the action effect adjustment based on the pipeline construction quality. 0.92 ensures that the rewards in the next 10 steps still retain a significant weight. Indicates time step Instant rewards; Indicates a given construction state Time-based policy network output action The probability density; The probability density is logarithmic. Policy network parameters. The update uses the Adam optimizer, with a learning rate of The learning rate is set to 0.0008, and it is decayed by multiplying by 0.6 every 200 iterations to stabilize later training. During training, the number of rounds is set to 500 construction trajectories, and the length B of each construction trajectory is at least equal to 10. If the length of a trajectory in a certain round is less than 10 steps, the trajectory is discarded from the training set.

[0045] During the online phase, the adaptive control module loads the parameters of the policy network after training convergence. And stop gradient updates. At the end of each construction data acquisition cycle, the adaptive control module adjusts the current construction state vector. Calculate the mean using the policy network. and standard deviation The three-dimensional original action vector is sampled from the corresponding Gaussian distribution. The values ​​of the welding current adjustment dimension are clipped to the [-15, 15] interval, the welding speed adjustment dimension to the [-1.5, 1.5] interval, and the preheating temperature adjustment dimension to the [-25, 25] interval, resulting in adjustment instructions that meet the constraints. The adaptive control module encapsulates the adjustment instructions into a JSON-formatted message body and sends it to the welding power control terminal and the preheating equipment via the MQTT IoT protocol. The welding power control terminal modifies the current setting value for the next weld based on the welding current adjustment value in the adjustment instruction, and modifies the wire feeding mechanism speed setting value based on the welding speed adjustment value; the preheating equipment modifies the heating power output based on the preheating temperature adjustment value. After completing one instruction issuance and execution, the adaptive control module waits for the next acquisition cycle to arrive, forming a closed-loop control.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An information management system for the construction quality of gas pipelines, characterized in that, include: The data sensing module is used to collect multimodal heterogeneous data during the construction of gas pipelines. The multimodal heterogeneous data includes construction operation log text, pipeline welding timing parameters, environmental sensor readings, and material quality indicators. The heterogeneous graph construction module is used to construct a dynamic heterogeneous graph based on construction process nodes, geographical location nodes, material batch nodes, and the multimodal heterogeneous data, mapping different modal data to node attributes; The dual-channel decoupled representation module is used to perform decoupled representation learning on the dynamic heterogeneous graph, and decomposes the node attributes into domain-invariant features and domain-specific features through a variational autoencoder. The comparison prediction module is used to predict the deviation of quality indicators of each construction process node based on the domain-invariant features and domain-specific features through time-series comparison learning, and to map the domain-specific features into quality influence factors through a fully connected layer. The domain-invariant features under the contrastive loss constraint are fused with the quality influence factor and input into the time recursive decoder to generate predicted values ​​of quality indicators for future time moments. Calculate the deviation between the predicted quality index value and the actual collected quality index value; The causal tracing module is used to map the deviation of the quality indicators to a causal directed acyclic graph and use intervention analysis to identify the root causes of the deviation, such as process and material defects. The adaptive control module is used to generate construction parameter adjustment instructions based on the root cause process and material defects through a strategy gradient algorithm, and then send them to the on-site control terminal.

2. The gas pipeline construction quality information management system according to claim 1, characterized in that, The data sensing module is specifically used for: The construction operation log text is segmented and part-of-speech tagged, and operation description tuples are extracted. The operation description tuples include actions, objects, and construction locations. The pipeline welding timing parameters are divided into segments according to the welding passes, and the current stability and voltage fluctuation rate in each segment are calculated. Outlier values ​​were removed from the environmental sensor readings to construct an environmental time series, and the trends in soil moisture content and geothermal gradient characteristics were extracted. The material quality indicators are matched with a preset batch standard library to generate a material compliance mark.

3. The gas pipeline construction quality information management system according to claim 1, characterized in that, The heterogeneous graph construction module is specifically used for: The construction process nodes are constructed with directed edges according to the logical order in the construction plan to form a process skeleton diagram; Add a spatial association edge between each construction process node and its corresponding construction geographic location node, wherein the geographic location node carries coded coordinate segment information; Connect the material batch node to the construction process node that uses the batch of material to form a material-process edge; Based on the timestamps of the multimodal heterogeneous data, the dynamic features within the time window are embedded with the time decay weights of the corresponding edges.

4. The gas pipeline construction quality information management system according to claim 1, characterized in that, The dual-channel decoupling characterization module is specifically used for: A unique variational autoencoder is constructed for each node in the dynamic heterogeneous graph, and the encoder output of the variational autoencoder is a latent variable distribution. By sampling from the latent variable distribution, domain-invariant features that follow a standard Gaussian prior and domain-specific features constrained by modal conditions are separated.

5. The gas pipeline construction quality information management system according to claim 4, characterized in that, The comparison and prediction module is specifically used for: Domain-invariant features of the same construction process node at consecutive time steps are selected as positive sample pairs, and domain-invariant features of different construction process nodes are selected as negative sample pairs to construct a contrastive loss function.

6. The gas pipeline construction quality information management system according to claim 5, characterized in that, The causal attribution module is specifically used for: Load a predefined causal directed acyclic graph template for construction quality, wherein the template contains pre-assumed causal relationship directions between process nodes, material nodes, and quality index nodes; The deviation of the quality index is used as an intervention variable, and the conditional distribution of each node in the causal directed acyclic graph is updated using an additive noise model. Perform the do-operation on each candidate root cause node to calculate the average causal effect of the deviation of the quality index after intervention; Based on the magnitude of the average causal effect, the root cause process identifier and material defect type are output.

7. The gas pipeline construction quality information management system according to claim 6, characterized in that, The adaptive control module is specifically used for: Define a construction state space, which consists of the current deviation of quality indicators, environmental sensor readings, and material compliance markers; Define the action space, which includes the welding current adjustment amount, welding speed adjustment amount, and preheating temperature adjustment amount; The policy network is trained using the reduction of quality index deviation as the reward signal, and the policy network adopts the Monte Carlo policy gradient update rule. During the online phase, adjustment instructions that meet the constraints are output based on the current construction status and sent to the welding power supply control terminal and preheating equipment via the Internet of Things protocol.

8. The gas pipeline construction quality information management system according to claim 2, characterized in that, When processing the pipeline welding timing parameters in the data sensing module, a sliding window is used to divide the welding current and voltage sequences. The window length adapts to the welding speed. Skewness, kurtosis, and envelope spectral entropy are extracted as statistical features within the window. These statistical features are normalized and used as a structured representation of the welding timing parameters.

9. The gas pipeline construction quality information management system according to claim 2, characterized in that, When the heterogeneous graph construction module maps different modal data to node attributes, for the text modality, it uses a pre-trained lightweight word vector model to encode the operation description tuple into a dense vector. For numerical time series modes, the energy proportion of each frequency band is extracted by multi-level wavelet packet decomposition to form a frequency domain attribute vector, and the spatiotemporal coordinates are transformed into a low-dimensional embedding based on geo-hash coding.

10. The gas pipeline construction quality information management system according to claim 1, characterized in that, The causal directed acyclic graph template in the causal tracing module is determined based on structured information from historical construction quality accident reports. It learns the initial causal skeleton through a PC algorithm and is then verified and corrected by domain experts. The intervention analysis employs kernel density estimation to replace parameter distribution assumptions in response to causal effect identification in small sample data.

Images