Intelligent grouting big data-based real-time warning method for abnormal grouting construction

Abstract

The application discloses a grouting abnormal construction real-time early warning method based on intelligent grouting big data. S1, a normal construction reconstruction model is established, which is used for reconstructing reconstruction construction data of a current construction time period under a normal construction mode according to construction data of a historical time period; S2, a normal construction prediction model is established, which is used for predicting prediction construction data of a next time node according to the construction data of the historical time period; S3, actual construction data generated by grouting construction is acquired in real time, which is input into the two models, the reconstruction construction data of the current time period under the normal construction mode is calculated, and the prediction construction data of the next time node is predicted, which is used for error calculation; S4, early warning is performed according to the values of mode reconstruction error and time sequence evolution error. The application has the advantages of constructing an interpretable normal benchmark, overcoming label dependency, and realizing prediction and reconstruction double-error fusion discrimination for online early warning, and giving consideration to the advantages of early identification of sudden and gradual abnormalities.

Description

Technical Field

[0001] This invention relates to the technical field of artificial intelligence applied to geotechnical engineering, specifically to a method for early warning of abnormal grouting construction, particularly a real-time early warning method for abnormal grouting construction based on intelligent grouting big data. Background Technology

[0002] Curtain grouting is a crucial process in the seepage prevention system of water conservancy, hydropower, and underground engineering, directly affecting the seepage prevention effect and operational safety of projects such as dam foundations, underground powerhouses, and tunnels. With the expansion of project scale and the increasing complexity of geological conditions, abnormal conditions during grouting construction occur frequently and are gradually becoming important factors affecting project quality and efficiency. Abnormal fluctuations in grouting pressure can cause the grout diffusion path to deviate from the design; sudden increases or decreases in flow rate may correspond to leakage, cross-contamination, or blockage; fluctuations in the water-cement ratio can alter the rheological and setting characteristics of the grout, resulting in insufficient filling or poor consolidation. Due to the coupled effects of factors such as geological fissures, karst channels, and permeable interlayers, the three-dimensional pressure-flow-water-cement ratio (PQC) curve during construction exhibits strong nonlinearity and uncertainty. Abnormal construction conditions are often concealed in the early stages but evolve rapidly. If they are not identified and intervened in a timely manner, they can easily lead to rework, material waste, project delays, and even long-term leakage risks.

[0003] Traditional anomaly identification mainly relies on human experience and simple threshold interpretations (such as pressure stabilization and changes in grout intake). However, modern curtain grouting projects involve numerous borehole segments, frequent cycles, and high sampling frequencies. Each borehole segment generates a three-dimensional time series of PQC (Pressure Quality Control) data that varies with injection time. This data is massive and highly dynamic, making manual identification extremely labor-intensive, inefficient, and susceptible to subjective influences, resulting in insufficient consistency and accuracy. Furthermore, construction conditions are influenced by multiple factors, including geological variations, borehole depth stress, grouting techniques, mix adjustments, and the cumulative effect of time, leading to poor predictability. Traditional manual identification and fixed thresholds struggle to depict complex evolutionary relationships, especially under special conditions such as leakage, cross-contamination, and blockage, making early warning difficult and often resulting in only post-incident detection, missing the optimal window for intervention. Therefore, there is an urgent need for a real-time early warning method for grouting anomalies based on intelligent grouting big data, tailored to the entire construction process. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] One of the technical problems this application aims to solve is that traditional methods for identifying anomalies in curtain grouting require manual operation, which is labor-intensive, inefficient, and highly subjective, resulting in insufficient consistency and accuracy in judgment.

[0006] Another technical problem that this application aims to solve is that traditional curtain grouting anomaly identification methods are difficult to characterize complex evolutionary relationships, especially in special conditions such as leakage, cross-contamination, and blockage, making it even more difficult to achieve early warning. They often can only be discovered after the fact, missing the best window for handling.

[0007] (II) Technical Solution

[0008] To address the aforementioned technical problems, this invention proposes a real-time early warning method for abnormal grouting construction based on intelligent grouting big data, comprising the following steps: S1, establishing a normal construction reconstruction model, which is used to reconstruct the reconstruction construction data of the current time period under the normal construction mode based on the actual construction data of the current time period; S2, establishing a normal construction prediction model, which is used to calculate the predicted construction data of the next time node of the current time period based on the actual construction data of the current time period; S3, acquiring the actual construction data generated by grouting construction in real time, and inputting the acquired actual construction data into the normal construction reconstruction model and the normal construction prediction model respectively, to calculate the reconstruction construction data of the current time period and the predicted construction data of the next time node of the current time period; S4, taking the difference between the actual construction data and the reconstructed construction data as the mode reconstruction error, and taking the difference between the actual construction data and the predicted construction data as the time evolution error, and issuing an early warning based on the values ​​of the mode reconstruction error and the time evolution error.

[0009] According to a preferred embodiment of the present invention, the construction data is represented by a data sequence that changes over time; the method further includes step S0: constructing a time-series feature extractor, which is used to convert the construction data sequence generated during grouting construction into a time-series feature vector.

[0010] According to a preferred embodiment of the present invention, the temporal feature extractor is composed of an LSTM encoder of an LSTM autoencoder network.

[0011] According to a preferred embodiment of the present invention, the construction reconstruction model is trained by the LSTM decoder of the LSTM autoencoder network.

[0012] According to a preferred embodiment of the present invention, the time-series feature vector is obtained by the time-series feature extractor based on normalized three-dimensional data of the injection pressure P, injection flow rate Q, and water-cement ratio C.

[0013] According to a preferred embodiment of the present invention, step S4 includes:

[0014] S4.1 Construct comprehensive early warning indicators and threshold benchmarks;

[0015] S4.2 Calculate the pattern reconstruction error for the current time period in real time;

[0016] S4.3 Calculate the temporal evolution error of the current time period in real time;

[0017] S4.4 Calculate the fusion error based on the comprehensive early warning index and the threshold benchmark, and issue an early warning based on the threshold benchmark.

[0018] According to a preferred embodiment of the present invention, in step S4.1, the comprehensive error is used as the comprehensive early warning index and calculated according to the following formula: In the formula, It is a comprehensive error. It is the normalized model reconstruction error. It is the normalized time series evolution error. and It is the weighting coefficient.

[0019] According to a preferred embodiment of the present invention, the normalized mode repetition error is calculated according to the following formula:

[0020] ,

[0021] in, To reconstruct the error of the normalized model, This represents the current number of nodes. This represents a certain type of data in 3D construction data. This indicates the weight of a certain type of construction data. For a specific value of a certain type of data at a certain node, Reconstruct values ​​for a certain type of data at a certain node.

[0022] According to a preferred embodiment of the present invention, the normalized time series evolution error is calculated according to the following formula:

[0023] ,

[0024] in, It is the normalized time series evolution error. This represents the current number of nodes. This represents a certain type of data in 3D construction data. This indicates the weight of a certain type of construction data. For a specific value of a certain type of data at a certain node, This refers to the predicted construction data for the next time point in the current time period.

[0025] (III) Beneficial Effects

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. Build interpretable normal benchmarks to overcome label dependencies.

[0028] This invention can still automatically screen normal samples and classify working conditions even under conditions of missing labels, mixed construction conditions, and significant geological differences, and has good interpretability and transferability.

[0029] 2. A dual-error fusion discrimination method using segments and nodes is adopted to take into account both abrupt changes and gradual anomalies.

[0030] This invention enables earlier and more stable anomaly warnings during construction. It can capture instantaneous and sudden anomalies such as leakage and cross-contamination, as well as gradual and cumulative anomalies such as blockage and mix drift, thus achieving a better balance between sensitivity and robustness. Attached Figure Description

[0031] Figure 1 This is a flowchart of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data according to the present invention.

[0032] Figure 2 This is a flowchart illustrating the overall implementation of an embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data according to the present invention.

[0033] Figure 3 This is a schematic diagram of the complete construction data sequence and the prefix sequence of the construction data according to the present invention.

[0034] Figure 4 This is a schematic diagram of the model architecture based on the LSTM autoencoder of the present invention.

[0035] Figure 5 This is a loss curve diagram of the normal construction reconstruction model training process in one embodiment of the present invention.

[0036] Figure 6 This is a visualization of the results of clustering historical construction data under different working conditions in one embodiment of the present invention.

[0037] Figure 7 This is a schematic diagram of the early stage of construction in one embodiment of the present invention, before the model is involved in early warning.

[0038] Figure 8 This is a schematic diagram illustrating the initial participation of the model in early warning in one embodiment of the present invention.

[0039] Figure 9 This is a schematic diagram illustrating the model participating in early warning twice in one embodiment of the present invention.

[0040] Figure 10 This is a schematic diagram illustrating the continuous early warning error judgment of the model in one embodiment of the present invention.

[0041] Figure 11 This is a three-dimensional spatial visualization diagram of the actual construction curve and the reconstruction curve in one embodiment of the present invention. Detailed Implementation

[0042] To address the shortcomings of existing technologies, this invention proposes a real-time early warning method for abnormal grouting construction based on intelligent grouting big data. Figure 1 This is a flowchart of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data according to the present invention. Figure 1 As shown, the method of the present invention includes the following steps:

[0043] S1. Establish a normal construction and reconstruction model.

[0044] The first aspect of this invention establishes a normal construction reconstruction model, which is used to reconstruct the reconstruction construction data for the current time period under normal construction mode based on the actual construction data for the current time period. Since the construction data is sequential data arranged in time, the prefix sequence here refers to the subsequence at the beginning of the construction data arranged in time sequence.

[0045] Specifically, this invention uses the prefix sequence of construction data in a normal construction database (a database that stores construction data under normal construction conditions) as training data to train a reconstruction decoder for normal construction conditions, which serves as a normal construction reconstruction model, enabling it to learn the reconstruction rules of construction data under normal construction conditions.

[0046] Preferably, the prefix sequence used in the training of the normal construction reconstruction model is formed by truncating the complete historical construction data sequence, and is used to represent the short-term construction data continuously collected during actual construction. In the specific data processing, it is preferable to represent the prefix sequence of the construction data as a low-dimensional temporal feature vector. Therefore, the present invention preferably includes a step of constructing a temporal feature extractor before step S1, i.e., step S0.

[0047] The temporal feature extractor is used to extract low-dimensional temporal feature vectors that change over time from the construction data generated during grouting construction, and to serve as intermediate input data for training and computation.

[0048] Therefore, the normal construction reconstruction model of the present invention takes the prefix sequence of normal construction data as input, and obtains the low-dimensional temporal features of the current prefix sequence based on the temporal feature extractor constructed in step S0. The prefix sequence is reconstructed based on the low-dimensional features. The reconstruction decoder is optimized with minimizing the reconstruction error as the training objective, so that it can reconstruct the prefix sequence under the normal construction pattern according to the low-dimensional temporal features.

[0049] The normal construction reconstruction model of this invention can achieve high-precision reconstruction when inputting construction data under normal construction mode. However, when inputting construction data that deviates from the normal construction mode, its reconstruction error will increase significantly. The difference between the actual construction data and the reconstructed construction data is defined as the "mode reconstruction error," which can serve as one of the key indicators for measuring the degree of deviation between the current construction process and the normal construction mode in the online stage.

[0050] S2. Establish a normal construction prediction model.

[0051] The second aspect of this invention establishes a normal construction prediction model, which is used to predict the construction data for the next time node corresponding to the current time period based on the actual construction data of the current time period.

[0052] Specifically, step S2 samples from the normal construction database, constructing training samples of different lengths for each normal construction data sequence under the normal construction mode: the prefix sequence is used as input, and the construction data of its corresponding next time node is used as the prediction target. In offline training, the prefix sequence is first input into the temporal feature extractor constructed in step S0 to obtain a low-dimensional temporal feature vector. Then, the normal construction prediction model is trained based on this low-dimensional temporal feature vector, outputting the predicted construction data value of the next time node in the current time period, thereby obtaining a normal construction prediction model containing the evolution law of normal construction. In the online stage, the model can receive the prefix sequence generated by actual construction in real time and output the predicted value of the construction data of the next time node corresponding to the current time period under the normal construction mode, i.e., the predicted construction data. At the next time node, after the actual construction data is obtained, the predicted construction data can be compared with the actual construction data. The difference is referred to here as the "temporal evolution error". The "temporal evolution error" can be used to characterize whether abnormal evolution trends or abrupt changes occur in the actual construction process.

[0053] S3. Real-time acquisition of construction data, calculation of reconstruction construction data and prediction of construction data.

[0054] After steps S1 and S2 have established the normal construction reconstruction model and the normal construction prediction model respectively, step S3 is the online application step. In step S3, the actual construction data generated by the grouting construction is acquired in real time, and the acquired construction data is input into the normal construction reconstruction model and the normal construction prediction model in real time to calculate the reconstruction construction data for the current construction time period under the normal construction mode, and to calculate the predicted construction data for the next time node of the current time period.

[0055] S4. Calculate the model reconstruction error and temporal evolution error and provide early warning.

[0056] As mentioned above, the present invention uses the difference between the actual construction data and the reconstructed construction data as the "model reconstruction error" and the difference between the actual construction data and the predicted construction data as the "temporal evolution error". Thus, the present invention can provide early warning based on the values ​​of the model reconstruction error and the temporal evolution error.

[0057] Specifically, a comprehensive early warning index is first constructed based on the pattern reconstruction error and temporal evolution error obtained in step S3. This comprehensive early warning index is used to simultaneously capture two types of abnormal signals: one reflects the cumulative deviation between the actual construction process and the normal construction mode, and the other reflects the sudden change or deviation of the construction data at the current time node.

[0058] The thresholds and triggering rules of the comprehensive early warning indicators are determined based on the statistical characteristics of historical normal data, and adapted threshold benchmarks are formed by combining factors such as different hole sections and process stages.

[0059] Next, during the online early warning phase, early warning information is output based on whether the real-time calculation error exceeds the threshold range, and the comprehensive early warning index is judged in real time to determine whether it exceeds the threshold range determined in the offline phase. When the triggering conditions are met, an abnormal early warning is output and feedback is provided on the time node of the abnormality, the current hole section information, and the contribution of the error signal, providing a basis for on-site engineers to handle the situation in a timely manner.

[0060] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, but this does not limit the scope of protection of this application.

[0061] Figure 2 This is a flowchart illustrating the overall implementation of an embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data according to the present invention. Figure 2 As shown, the method of the present invention includes the following steps.

[0062] S1. Establish a normal construction and reconstruction model.

[0063] As mentioned earlier, step S1 uses the prefix sequences in the normal construction database as training data to train a reconstruction decoder oriented towards normal construction conditions. This decoder serves as the normal construction reconstruction model, enabling it to learn the reconstruction patterns of construction data under normal construction conditions. Specifically, this step includes the following sub-steps:

[0064] S1.1 Preprocessing of historical construction data.

[0065] Step S1.1 is used to ensure that historical construction data with multiple holes, multiple durations, and multiple noise levels are unified into a trainable three-dimensional sequence dataset. At the same time, by constructing prefix samples, the input format for model training is made consistent with the input format of online acquisition.

[0066] S1.1.1 Data Acquisition: During historical grouting construction, for each grouting hole section, the intelligent grouting equipment collects and records the grouting pressure P, injection flow rate Q, and water-cement ratio C in real time, and stores them in the database with timestamps as indexes.

[0067] Figure 3 This is a schematic diagram of the complete construction data sequence and the prefix sequence of the construction data according to the present invention. For example... Figure 3 As shown, a complete construction data sequence can have multiple prefix sequences of different lengths. Examples of actual collected raw construction data include grouting pressure P, injection flow rate Q, and water-cement ratio C.

[0068] S1.1.2 Resampling and Data Length Standardization: Considering the varying construction times of different borehole segments, the overall historical construction times of all borehole segments were statistically analyzed. As an example, 80 minutes was selected as the standard length. The original sampled data was resampled at 2-minute intervals, and the construction curve of each borehole segment was represented as a fixed-length three-dimensional sequence, i.e., a 40-row × 3-column three-dimensional time series. For borehole segments exceeding the standard length, truncation was performed, retaining only the first 40 steps; for borehole segments shorter than the standard length, zero-padding was used to fill the tail, and a mask was simultaneously generated to distinguish between the actual sampling points and the padded points, ensuring that subsequent training loss was calculated only for the actual points.

[0069] S1.1.3 Normalization processing: The three-dimensional time series data P, Q, and C are normalized by mean-variance standardization to ensure the stability of subsequent training and the joint optimization of variables with different dimensions.

[0070] S1.1.4 Prefix Sequence Generation: For each complete borehole segment sequence, generate multiple sets of prefix sequences of different lengths to simulate the online scenario of data generation during construction. The prefix length is no less than 5 steps (i.e., 10 minutes), and then gradually increases to the full length; at the same time, retain the mask subset corresponding to each prefix.

[0071] S1.1.5 Dataset Partitioning: The aforementioned steps are used to establish a dataset of construction history data. The dataset is further divided into training set, validation set, and test set (ratio: 7:1:2) based on borehole segments for subsequent model training and threshold calibration.

[0072] As mentioned above, the construction data of this invention is preferably represented by low-dimensional temporal feature vectors. Therefore, this invention further includes a step S0 of constructing a temporal feature extractor before step S1. The temporal feature extractor is used to convert the real data generated during grouting construction into low-dimensional temporal feature vectors, which are used as construction data for training and calculation. The constructed temporal feature extractor allows the model to automatically learn the temporal representation of the pressure-flow-water-cement ratio coupling evolution under unlabeled conditions, providing a separable and measurable feature space for subsequent clustering and screening of normal construction datasets.

[0073] (1) Model structure: Construct an autoencoder network consisting of a long short-term memory network (LSTM) encoder and an LSTM decoder.

[0074] Figure 4 This is a schematic diagram of the LSTM-Autoencoder model of the present invention. Figure 4 As shown, the LSTM autoencoder includes an encoder and a decoder. The intermediate layer uses an LSTM model sensitive to time-varying data to extract temporal features. It extracts temporal dependencies and variable coupling information from the input PQC 3D temporal data, compressing the entire sequence into a low-dimensional feature vector to characterize the overall construction dynamics of the borehole segment. The decoder reconstructs the original 3D temporal sequence based on the low-dimensional temporal features, making the reconstructed sequence as close as possible to the input sequence. In the figure, tanh is the tanh function, an activation function that maps data to [-1,1], and σ is the sigmoid activation function, an activation function that maps data to [0,1].

[0075] (2) Training objective: Using reconstruction error as the loss function, minimize the difference between the input sequence and the reconstructed sequence; when there are padding points, filter out the padding points by masking, and calculate the loss only for the real sampling points. In this embodiment, the pattern reconstruction error is calculated using mean squared error (MSE).

[0076] (3) Network structure and hyperparameter settings: In this embodiment, the LSTM is set to two layers with 64 hidden units and 16 latent feature dimensions; the optimizer is Adam with a learning rate of 1e-3, a batch size of 256, and 1000 training epochs, and an early stopping strategy is configured to avoid overfitting.

[0077] Figure 5 This is a loss curve during training of a normal construction reconstruction model in one embodiment of the present invention. Figure 5 As can be seen, the loss value decreased rapidly, indicating that the model efficiently learned the extraction rules of temporal features and could be used to reconstruct the complete sequence.

[0078] (4) Model solidification: After training converges, the encoder is fixed as a "temporal feature extractor" for subsequent low-dimensional feature extraction, working condition clustering, and front-end feature input for normal construction reconstruction model and normal construction prediction model.

[0079] S1.2, Working condition clustering and construction data type classification.

[0080] Unlabeled historical construction data is transformed into usable resources with groupable working conditions, and highly reliable normal construction samples are selected to provide a data foundation for learning normal patterns and enhance the interpretability of the method.

[0081] S1.2.1 Low-dimensional feature extraction: Using the trained encoder, the complete construction data sequence of all historical construction borehole segments is encoded one by one to obtain the corresponding low-dimensional time-series feature vector set. This feature can comprehensively characterize the rise, stability, decay and fluctuation of pressure, the decay rate and abrupt change behavior of flow rate, the adjustment trajectory and stability of water-cement ratio, and the coupling relationship between the three variables.

[0082] S1.2.2, Working Condition Clustering: HDBSCAN is used to cluster historical construction data based on low-dimensional features. In this embodiment, the minimum cluster size is set to 30 and the minimum number of samples is 5 to automatically identify working condition clusters of different densities and mark outliers as noise or abnormal candidates.

[0083] S1.2.3 Normal Sample Screening: Based on the clustering results, combined with construction records, quality acceptance indicators, final hole standards, and records of leakage, cross-contamination, and blockage, the clustering results are interpreted and judged according to working conditions. Samples that meet the specifications and stably represent normal construction are screened out to form a normal construction database. The remaining samples are used as an abnormal reference set and do not participate in the learning of normal patterns.

[0084] Figure 6 The visualization results of density clustering in the feature space after feature extraction of 3086 actual construction data in this embodiment. Figure 6 It is a visualization display that reduces actual construction data to three dimensions using the UMAP (Uniform Manifold Approximation and Projection) dimensionality reduction algorithm. Figure 6 The clustering results of the construction data show that there are four types of normal construction and one type of abnormal construction. Corresponding to the construction curve analysis of each type of result, normal construction involves different construction types commonly seen in engineering, such as low pressure and high flow, pressure increase and less slurry intake, high pressure and no slurry intake, and high pressure and flow attenuation. Abnormal construction obtained from the clustering is mainly reflected in the abnormal fluctuation of flow.

[0085] S1.3 Training the normal construction reconstruction model.

[0086] By training with only normal construction samples, the normal construction reconstruction model internalizes the normal construction pattern, thereby naturally generating a larger reconstruction deviation when abnormal inputs occur, thus enabling anomalies to be detected.

[0087] S1.3.1 Encoder Fixation: The temporal feature extractor trained previously is used as the front-end feature extraction module, and its parameters are fixed in this stage to maintain the stability and consistency of the feature space.

[0088] S1.3.2 Decoder Training: Using prefix sequences of different lengths from the normal construction database as input, low-dimensional features are first obtained through a temporal feature extractor, and then input into the reconstruction decoder to reconstruct the corresponding prefix sequences. The training objective is to minimize the difference between the reconstructed sequence and the prefix sequence of the actual construction data; when there are padding points, the loss is calculated only for the real sampling points in combination with mask marking.

[0089] S1.3.3, Consolidation of the Normal Construction Reconstruction Model: A reconstruction model that internalizes the normal construction patterns is obtained through training. During the online phase, the model receives real-time prefix sequences (construction data for the current time period) and outputs reconstructed construction data, namely the "reconstruction prefix sequence under normal mode," and quantifies the degree of deviation between the current construction process and the normal construction mode using "mode reconstruction error."

[0090] S2. Establish a normal construction prediction model.

[0091] This step establishes and trains a normal construction prediction model, which can predict construction data at the next time point. By training using only normal construction samples, the model learns the short-term evolution patterns of construction data under normal construction patterns, which is used in the online phase to determine whether the latest data point deviates from the normal evolution direction.

[0092] S2.1 Training Sample Construction: Samples are taken from the normal construction database to construct training samples for each complete construction data sequence of a normal construction mode: prefix sequences of different lengths are taken as input, and the construction data (PQC data) of the next time node corresponding to them is taken as the prediction target. In this embodiment, to ensure the effectiveness of the prediction, the lower limit of the prefix length is set to 5 steps, and 10 prefix sequences of different lengths are randomly selected from each construction data sequence for training.

[0093] S2.2 Training of the normal construction prediction model: Taking the prefix sequence as input, a low-dimensional feature vector is first extracted by a fixed encoder, and then the prediction head network is connected to the back end of the low-dimensional feature vector to output the three-dimensional prediction value of the next time node. In this embodiment, the prediction head network adopts a two-layer fully connected structure with 32 hidden layers and the output layer is 3D PQC data. MSE is used as the error for target optimization during training.

[0094] S2.3, Solidification of the Normal Construction Prediction Model: After the model training is completed, the normal construction prediction model is obtained. In the online stage, the model can receive the prefix sequence of real-time construction data and output the value of the predicted construction data of the next time node in the normal construction mode.

[0095] S3. Real-time acquisition of actual construction data, calculation of reconstruction construction data and prediction of construction data.

[0096] This step involves real-time acquisition of actual construction data and construction data sequence construction. Based on the raw construction data collected by the intelligent grouting equipment during actual construction, this data is processed and used as a prefix sequence for subsequent error calculation. This step also utilizes models to calculate and reconstruct construction data and predict construction data.

[0097] S3.1 Real-time data acquisition: This step receives the data stream consisting of pressure P, flow rate Q, and water-cement ratio C of the current borehole section in real time during the actual construction process, and stores it in the database with timestamp as the index.

[0098] S3.2 Online Data Preprocessing: This step involves online data preprocessing, which is the same as step S1, namely, resampling, missing data handling, outlier handling, and normalization, to ensure that the online data distribution is consistent with the training data.

[0099] S3.3 Prefix Maintenance: As construction progresses, the prefix sequence of the ever-growing real-time construction data needs to be dynamically maintained. For subsequent judgment and early warning, this embodiment of the invention stores the continuously updated actual construction data in real time and caches it as the prefix sequence input to the model.

[0100] S3.4 Calculate reconstructed construction data and predicted construction data: The prefix sequence of the actual construction data collected in real time is extracted into low-dimensional time-series feature vectors by the time-series feature extractor trained by S0. Then, it is input into the normal construction reconstruction model and normal construction prediction model trained by S1 and S2 respectively to obtain the reconstructed construction data corresponding to the current prefix sequence (current time period) under the normal construction mode and the predicted construction data of the next time node of the current time period.

[0101] S4. Calculate the model reconstruction error and temporal evolution error and provide early warning.

[0102] In this embodiment of the present invention, when the prefix sequence length of the actual construction data is ≥5 steps (i.e., 5 construction data are obtained), an early warning judgment can be started, and the judgment result is updated once after each new time node is added.

[0103] This embodiment of the invention performs real-time prediction and early warning according to the following steps:

[0104] S4.1 Construct comprehensive early warning indicators and threshold benchmarks.

[0105] This step integrates the reconstruction error and the temporal evolution error of the integrated model to construct an early warning indicator that can be triggered online, and calibrates the threshold on normal construction data.

[0106] S4.1.1 Error Benchmark Statistics: On the validation set of normal construction data, the distribution characteristics of the model reconstruction error and the temporal evolution error, such as mean, standard deviation and quantile, are statistically analyzed to form an error benchmark.

[0107] S4.1.2 Construction of Comprehensive Early Warning Indicators: The model reconstruction error and the temporal evolution error are normalized and weighted and fused to form a comprehensive early warning indicator. As an example, the comprehensive error is used. As a comprehensive early warning indicator, it is calculated according to this formula: In the formula, It is a comprehensive error. It is the normalized model reconstruction error. It is the normalized time series evolution error. and These are weighting coefficients, which are calibrated during the training phase using a normal construction dataset. In this embodiment... The value is 0.642. The value of 0.358 indicates that in this embodiment, the reconstruction error has a higher weight and the temporal evolution error has a slightly lower weight. That is, the model pays great attention to whether the actual construction curve conforms to the normal construction mode, and the prediction error of the next time node is used as an important reference. This weighting coefficient setting can prevent frequent false alarms caused by single-point anomalies during the online use phase.

[0108] S4.1.3, Constructing Early Warning Thresholds and Triggering Rules: As an example, the 95th percentile of the normal validation set can be used as the early warning threshold. Specifically, a weak early warning action can be set for a single point exceeding the threshold, providing only an informational notification; and a strong early warning action can be triggered when the threshold is exceeded multiple times consecutively (e.g., 3 times), such as a buzzer or light warning. This early warning method can reduce false alarms caused by noise.

[0109] S4.2 Calculate the pattern reconstruction error for the current time period in real time.

[0110] S4.2.1 Feature Extraction: Based on the actual construction data (prefix sequence) acquired in real time, input the time-series feature extractor fixed in step S0 to obtain its low-dimensional feature vector.

[0111] Figure 7 This is a schematic diagram illustrating an embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data, according to the present invention, where the initial construction model has not yet participated in the early warning process. In the diagram, hollow circles represent one actual construction data point at one time node. Model A represents the normal construction reconstruction model, and Model B represents the normal construction prediction model. In this embodiment, construction data is collected every 2 minutes for feature extraction. Figure 7 The graph shows four data points collected 8 minutes before construction began. As can be seen from the graph, since only four data points were collected at this time, neither Model A nor Model B started calculating the early warning.

[0112] S4.2.2 Calculate the reconstruction construction data for the current time period: Input the low-dimensional feature vector into the normal construction reconstruction model trained in step S1, and output the reconstruction construction data, that is, the prefix sequence of the reconstruction construction data under the normal construction mode.

[0113] Figure 8 This is a schematic diagram of the model's initial participation in the early warning in one embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data of the present invention. Figure 8 This corresponds to the first 10 minutes, during which 5 actual construction data sequences were acquired. Both Model A and Model B were triggered, generating reconstructed construction data and predicted construction data, respectively. From Figure 8 It can be seen that the reconstructed construction data and the actual construction data are data from the same time period. The reconstructed construction data reflects the construction data that should be exhibited under the normal construction mode, and is represented by solid dots. The predicted construction data is the predicted construction data for the next time node (i.e., the 12th minute) based on the actual construction data. It represents the construction data that is most likely to be obtained at the next time node if the normal construction mode is followed.

[0114] S4.2.3 Calculate the reconstruction error: Compare the reconstructed construction data with the original construction data, calculate the difference as the reconstruction error, and use it to assess the deviation of the current construction process from the normal construction mode.

[0115] exist Figure 8 In the example shown, the actual construction data obtained at minutes 2, 4, 6, 8, and 10 are compared with the reconstruction construction data at minutes 2, 4, 6, 8, and 10 calculated by the normal construction reconstruction model. The average, standard deviation, or quantile of the difference is calculated as the deviation of the current construction process from the normal construction mode. Figure 8 middle The mean square error is represented by the difference between the actual construction data and the reconstructed construction data at the 10th minute.

[0116] Specific calculation method: ,in, To reconstruct the error of the normalized model, This represents the current number of nodes. This represents a certain type of 3D construction data (P, Q, C). This indicates the weight of a certain type of construction data. For a specific value of a certain type of data at a certain node, The normalized pattern reconstruction error is the final reconstructed value of a certain type of data at a certain node. This represents the combined error of the three types of data: P, Q, and C.

[0117] S4.3 Calculate the temporal evolution error of the current time period in real time.

[0118] S4.3.1 Calculate and cache the predicted construction data for the next time node of the current time period: Based on the actual construction data of the current time period, input it into the time series feature extractor fixed in step S0 to extract low-dimensional time series features, and then input it into the normal construction prediction model obtained in step S2, i.e. Figure 8 Model B outputs the predicted construction data for the next time node in the current time period and caches this value.

[0119] exist Figure 8 In the example shown, the predicted construction data calculated by Model B (normal construction prediction model) is the predicted construction data at the 12th minute, which is represented by a circle with a shaded line in the figure. It cannot be used at the current 10th minute, so it needs to be cached.

[0120] S4.3.2 Calculate the time series evolution error: After advancing one time node in the current time period, that is, after the actual construction data of the next time node is obtained, retrieve the cached predicted construction data, compare it with the actual construction data, and calculate the deviation as the time series evolution error.

[0121] Specific calculation formula: , It is the normalized time series evolution error. This represents the predicted construction data for the next time point in the current time period. This time series evolution error... It can be used to determine whether the construction data collected at the latest time point conforms to the normal construction evolution pattern.

[0122] Figure 9 This is a schematic diagram of the model participating in the early warning twice in one embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data of the present invention. Figure 9 The display shows 12 minutes of construction, with 6 actual construction data sequences acquired. Model A and Model B regenerate the reconstructed construction data and predicted construction data, respectively. It should be noted that although Model A has already reconstructed the construction data for minutes 2, 4, 6, 8, and 10 at minute 10, at minute 12, it will still reconstruct the reconstructed construction data for minutes 2, 4, 6, 8, 10, and 12 again based on the updated prefix sequence, i.e., the actual construction data for minutes 2, 4, 6, 8, 10, and 12. This is because the reconstruction model focuses on whether the current construction mode conforms to normal construction. Each time a new data point is added, the construction mode for the corresponding current time period changes. Therefore, reconstruction must be performed based on the latest prefix sequence. Furthermore, a well-trained normal construction reconstruction model only requires one set of inputs and outputs a corresponding set of numbers, making it highly efficient and capable of performing each reconstruction efficiently.

[0123] like Figure 9As shown, at the third time node, the root mean square error at that time node is calculated based on the recalculated reconstruction construction data. The pattern reconstruction error at the current time point.

[0124] Since the predicted construction data for the current time point (min 12) calculated by Model B from the previous time point (min 10) was cached in the previous step, we can now compare the actual construction data for the current time point with the predicted construction data and calculate the deviation as the time series evolution error. Here, we use... express.

[0125] S4.4 Calculate the fusion error based on the comprehensive early warning indicators and the threshold benchmark, and issue an early warning based on the threshold benchmark.

[0126] Figure 10 This is a schematic diagram of a model continuously judging early warning errors in one embodiment of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data of the present invention. Figure 10 The display shows the first 24 minutes, with 12 actual construction data sequences acquired. It can be seen that starting from the 12th minute, at each time point, Model A and Model B respectively regenerate the reconstructed construction data and predicted construction data, obtaining the pattern reconstruction error and temporal evolution error for each time point. Based on these errors, this invention can calculate the fusion error and provide intelligent early warning.

[0127] S4.4.1 Comprehensive Index Calculation: The normalized model reconstruction error and temporal evolution error obtained from S4.2 and S4.3 are weighted and fused according to step S4.1 to obtain a comprehensive early warning index. As an example, the comprehensive error is used. As a comprehensive early warning indicator, according to The calculated comprehensive early warning index in the current embodiment is 0.185.

[0128] S4.4.2 Trigger Decision: In the current embodiment, the comprehensive early warning index at the 24th minute is 0.285, which does not exceed the abnormal early warning threshold of 0.791 calibrated at the 95th percentile during the offline training phase. In this embodiment, the trigger decision can be made using the "single-point threshold exceedance + continuous threshold exceedance" rule established in step S4.1.3.

[0129] S4.4.3 Output early warning information: When the triggering conditions are met, output abnormal early warning information. The early warning information includes the time point of the abnormality, the hole segment number, the comprehensive index value, and the contribution of the two types of errors, so as to assist on-site personnel in quickly locating the source of the abnormality and taking disposal measures.

[0130] S4.4.4 Normal Construction: When the calculated comprehensive early warning indicators do not trigger an abnormal early warning, normal construction shall be carried out and S4.2 to S4.4 shall be repeated until the construction is completed or an early warning is triggered.

[0131] Figure 11 This is a three-dimensional visualization diagram of the actual construction curve and the reconstruction curve after inverse normalization for the subsequent 32 minutes of construction in one embodiment of the present invention. As can be seen in the figure, in the current embodiment, the reconstruction curve always has the same trend as the actual construction curve and matches well. There is only a slight error in the flow rate at the beginning of construction, and the time evolution error has never caused the comprehensive early warning index to reach the abnormal threshold, indicating that the construction in the current embodiment conforms to normal construction.

[0132] The above describes specific embodiments of the real-time early warning method for abnormal grouting construction based on intelligent grouting big data according to the present invention. As can be seen from the description of the above embodiments, the present invention has the following advantages.

[0133] First, this invention constructs an interpretable normal benchmark, overcoming label dependency.

[0134] This invention first utilizes an LSTM autoencoder for unsupervised representation learning on the full historical construction data, compressing the complex nonlinear construction process into a low-dimensional feature space. Then, density clustering is employed in this feature space to form different construction condition clusters. Combined with construction records and quality inspection results, a normal construction database is selected as the benchmark for the "normal construction mode range." Compared to existing technologies that rely on manual threshold interpretation or a large number of labeled samples, this invention can still achieve automatic screening of normal samples and classification of construction conditions even under conditions of missing labels, mixed construction conditions, and significant geological differences, and possesses good interpretability and transferability.

[0135] Secondly, the present invention adopts a dual-error fusion discrimination of sections and nodes, which takes into account the identification of sudden changes and gradual anomalies during the construction process.

[0136] This invention addresses the real-time accumulation of construction data by training a normal construction reconstruction model on a normal construction database. This model quantifies the cumulative deviation between the prefix sequence of the current construction data and the normal construction pattern. A time-series-based normal construction prediction model is also trained to characterize whether the construction data at the latest time point deviates from the evolution direction of the normal construction pattern. In the online phase, this invention simultaneously calculates pattern reconstruction error and time-series evolution error and constructs a comprehensive early warning index and threshold triggering rules. Compared to existing methods that rely on a single index / single error criterion or are biased towards post-hoc complete sequence detection, this invention can achieve earlier and more stable anomaly warnings during construction. It can capture both instantaneous abrupt anomalies such as leakage and cross-contamination, as well as progressively cumulative anomalies such as blockage and mix ratio drift, thus achieving a better balance between sensitivity and robustness.

[0137] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time early warning of abnormal grouting construction based on intelligent grouting big data, characterized in that, Includes the following steps: S0. Construct a time-series feature extractor, which is used to convert the real data generated during grouting construction into a low-dimensional time-series feature vector as construction data for training and calculation. The real data generated during grouting construction is a three-dimensional construction data sequence composed of grouting pressure P, injection flow rate Q and water-cement ratio C arranged in time series. S1. Preprocess the historical curtain grouting construction data, construct prefix sequences of different lengths, and train a temporal feature extractor to extract low-dimensional temporal features. Based on the low-dimensional time-series features, historical construction data are clustered according to working conditions, and normal construction samples are selected by combining construction records to construct a normal construction database; the time-series feature extractor is fixed, and a normal construction reconstruction model is trained separately based on the normal construction database. This normal construction reconstruction model is used to reconstruct the reconstruction construction data of the current time period under the normal construction mode based on the actual construction data of the current time period. S2. Fix the time series feature extractor and train a normal construction prediction model separately based on the normal construction database. The normal construction prediction model is used to calculate the predicted construction data for the next time node of the current time period based on the actual construction data of the current time period. S3. In the online phase, the actual construction data of the current hole segment is acquired in real time, and the continuously growing prefix sequence of the actual construction data is dynamically maintained. When the prefix sequence reaches the preset starting length, the updated prefix sequence is processed by the time-series feature extractor to extract features, and then input into the normal construction reconstruction model and the normal construction prediction model respectively to obtain the reconstruction construction data of the current time period and the predicted construction data of the next time node. S4. The difference between the actual construction data and the reconstructed construction data is taken as the model reconstruction error, and the difference between the actual construction data at the current time node and the predicted construction data at the current time node obtained from the previous time node is taken as the temporal evolution error. An abnormal construction warning is given based on the model reconstruction error and the temporal evolution error. In particular, when a new time node is added in the online phase, the model reconstruction error, the temporal evolution error and the corresponding warning result are updated.

2. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 1, characterized in that: The preprocessing includes resampling, data length unification, and normalization. For construction data sequences that are not of uniform length, tail padding is performed, and a mask is generated to distinguish between the real sampling points and the padding points, so that the model training loss is calculated only for the real sampling points.

3. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 1, characterized in that: The temporal feature extractor described in step S0 is used to convert the three-dimensional construction data sequence generated during grouting construction into temporal dependency and variable coupling information.

4. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 1, characterized in that: The temporal feature extractor is composed of the LSTM encoder of the LSTM autoencoder network, the normal construction reconstruction model is trained on the normal construction database by the decoder of the LSTM autoencoder network, and the normal construction prediction model is trained by the prediction head network that receives the low-dimensional temporal feature vector as input.

5. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 1, characterized in that: The prefix sequences of different lengths are formed by truncating complete historical construction data sequences and are used to simulate online scenarios where data is gradually generated during construction.

6. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 5, characterized in that: Step S4 includes: S4.1 Construct comprehensive early warning indicators and threshold benchmarks; S4.2 Calculate the pattern reconstruction error for the current time period in real time; S4.3 Calculate the temporal evolution error of the current time period in real time; S4.4 Calculate the fusion error based on the comprehensive early warning index and the threshold benchmark, and issue an early warning based on the threshold benchmark.

7. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 6, characterized in that: In step S4.1, the comprehensive error is used as the comprehensive early warning index and calculated according to the following formula: In the formula, It is a comprehensive error. It is the normalized model reconstruction error. It is the normalized time series evolution error. and It is the weighting coefficient.

8. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 7, characterized in that: The normalized model reconstruction error is calculated using the following formula: ， in, To reconstruct the error of the normalized model, This represents the current number of nodes. This represents a certain type of data in 3D construction data. This indicates the weight of a certain type of construction data. For a specific value of a certain type of data at a certain node, Reconstruct values ​​for a certain type of data at a certain node.

9. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to claim 7, characterized in that: The normalized time series evolution error is calculated using the following formula: ， in, It is the normalized time series evolution error. This represents the current number of nodes. This represents a certain type of data in 3D construction data. The representative indicates the weight of a certain type of construction data. For a specific value of a certain type of data at a certain node, This refers to the predicted construction data for the next time point in the current time period.

10. The method for real-time early warning of abnormal grouting construction based on intelligent grouting big data according to any one of claims 6 to 9, characterized in that: A weak warning is triggered when the comprehensive early warning indicator exceeds the threshold benchmark in a single instance. When the comprehensive early warning index continuously exceeds the threshold benchmark, a strong early warning is triggered, and the engineering analysis of the cause of the abnormal early warning is guided based on the pattern reconstruction error and the temporal evolution error.

Images