Cipp curing quality control method based on deep learning

Abstract

The application discloses a kind of CIPP solidification quality control methods based on deep learning, to solve the problem that multiple-source monitoring data is difficult to carry out time and space alignment fusion in CIPP solidification process, cannot real-time output along the solidification state of pipeline mileage and circumference and accurately determine solidification end point and position under-solidification and over-solidification, the application is by collecting temperature, pressure, flow, traction speed, ultraviolet light intensity and other multi-source data and correlating time stamp, mileage coordinate and circumferential angle, denoising completion, time alignment resampling and cylindrical coordinate gridding;Grid observation input is constrained by residual error of heat conduction equation and residual error of solidification reaction kinetics equation cylindrical coordinate Fourier neural operator model obtains initial temperature field, solidification degree and solidification rate;Then adopt sensor reliability weighted observation assimilation formula residual error Transformer to correct, and based on solidification standard condition and rate stability condition determination solidification end point and output abnormal area early warning, realize the technical effect of solidification state real-time output, end point determination reliability improvement and local anomaly positioning early warning.

Description

Technical Field

[0001] This invention relates to the field of pipeline lining curing construction, and in particular to a deep learning-based method for quality control of CIPP curing. Background Technology

[0002] CIPP (Cured-In-Place Pipe) is one of the mainstream technologies for trenchless repair of municipal drainage pipelines. It typically involves pulling or flipping a resin-impregnated lining material into the pipeline to be repaired, and then curing it using hot water, steam, or ultraviolet light. The curing quality directly affects the mechanical properties, corrosion resistance, and service life of the lining. Therefore, during construction, it is usually necessary to monitor parameters such as temperature, pressure, medium flow rate, pulling speed, and ultraviolet light intensity, and to control the curing process quality in accordance with process specifications. In existing technologies, on the one hand, thermocouple temperature measurement, pressure gauges, flow meters, and other sensors are commonly used on-site to record key parameters, and manual judgment is made based on experience or process curves. On the other hand, some technologies attempt to establish curing heat conduction and reaction kinetic models, or use data-driven methods to estimate the curing state, in order to improve the ability to determine the endpoint and identify anomalies.

[0003] However, existing technologies still have the following shortcomings:

[0004] 1. The data from multiple monitoring sources vary greatly in terms of sampling frequency, clock drift, and installation location, making it difficult to achieve reliable time and space alignment and fusion, resulting in unstable solidification status assessment results;

[0005] 2. Most solutions rely on a small number of sensor point values ​​or overall average values, making it difficult to output real-time field information such as the degree of curing along the pipeline mileage and circumferential distribution, and thus unable to effectively locate and warn of local under-curing or over-curing.

[0006] 3. The determination of the curing endpoint often relies on fixed thresholds or empirical rules, without fully considering the changes in curing rate and the reliability of the sensor. It is easily affected by noise, missing data and outliers, which may lead to misjudgment or lag.

[0007] Therefore, a CIPP curing quality control method that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a deep learning-based CIPP curing quality control method. Addressing the challenges of existing technologies such as difficulty in achieving temporal and spatial alignment and fusion of multi-source monitoring data, difficulty in obtaining the curing state distribution along pipeline mileage and circumference, susceptibility of curing endpoint determination to noise and missing data, and inability to locate and warn of local under-curing and over-curing, the following technical solution is proposed: Multi-source monitoring data, including timestamps, mileage coordinates, and circumferential angles, is collected and correlated. After denoising, completion, unified time axis alignment, and resampling, it is gridded in cylindrical coordinates. The gridded observations are input into a cylindrical coordinate Fourier neural operator model constrained by the residuals of the heat conduction equation and the curing reaction kinetic equation to obtain the initial temperature field, curing degree, and curing rate. Correction is then performed using the grid field as the query, sensor point observations as the key and value, and a residual Transformer weighted by sensor reliability. The curing endpoint determination is output based on the curing compliance condition and rate stability condition, and under-curing and over-curing location warnings are achieved based on the grid connected regions. This invention offers the technical advantages of real-time output of the spatiotemporal distribution of curing state, improved reliability of endpoint determination, and location warnings for local anomalies.

[0009] This invention provides a deep learning-based CIPP solidification quality control method, comprising:

[0010] S1. Collect multi-source monitoring data during the CIPP curing process, recording timestamps and corresponding mileage coordinates and circumferential angles to obtain raw multi-source monitoring data; S2. Denoise and impute missing values ​​in the raw multi-source monitoring data, perform time alignment and resampling based on timestamps, so that the monitoring data from each source correspond to the same time axis and retain the corresponding mileage coordinates and circumferential angles to obtain aligned multi-source monitoring data; S3. Discretize the mileage coordinates and circumferential angles to construct a cylindrical coordinate grid composed of mileage and circumferential dimensions, converting the aligned multi-source monitoring data into gridded observations indexed by time, mileage, and circumferential dimensions; S4. Input the gridded observations and cylindrical coordinate grid into the curing quality control model, outputting the initial temperature field, initial curing degree, and initial curing rate; S5. Determine the sensor reliability weights for each sensor point based on the aligned multi-source monitoring data, and adjust the initial temperature field, initial curing degree, and initial curing rate accordingly. Using the rate as the query input, the observation data of each sensor point in the aligned multi-source monitoring data is used as the key input and value input. The contribution of cross-attention is weighted using sensor reliability weights. The input residual Transformer network is used to correct the initial temperature field, initial curing degree, and initial curing rate to obtain the corrected temperature field, corrected curing degree, and corrected curing rate. The distribution of corrected curing degree and / or corrected curing rate on the cylindrical coordinate grid is output. S6. Based on the corrected curing degree, corrected curing rate, and cylindrical coordinate grid, each grid cell within the preset judgment range of the cylindrical coordinate grid is judged to meet the curing compliance condition and rate stability condition. When all grid cells within the preset judgment range meet the curing compliance condition and rate stability condition, the curing endpoint judgment result is output. S7. Based on the corrected curing degree and cylindrical coordinate grid, the under-cured area and over-cured area are determined, and the positioning warning result is output.

[0011] Optionally, S1 includes:

[0012] During the CIPP curing process, measurements are taken using at least two types of sensors, including those for acquiring temperature data, pressure data, flow rate data, traction speed data, and ultraviolet light intensity data, and the measured values ​​are output at the preset refresh cycle.

[0013] Assign a timestamp to each measurement;

[0014] The mileage coordinates corresponding to the measured value are determined based on the traction length measuring device or odometer;

[0015] The circumferential angle corresponding to the measured value is determined based on the sensor's installation location;

[0016] The measured values ​​are associated with and stored with the corresponding timestamps, mileage coordinates, and circumferential angles to obtain the original multi-source monitoring data.

[0017] Optionally, S2 includes:

[0018] Outlier removal and denoising are performed on the original multi-source monitoring data to obtain the denoised original multi-source monitoring data.

[0019] The missing values ​​of the denoised original multi-source monitoring data are filled in to obtain the filled original multi-source monitoring data.

[0020] A unified timeline is determined based on the timestamps of the completed original multi-source monitoring data, and the completed original multi-source monitoring data is time-aligned according to the unified timeline.

[0021] The time-aligned monitoring data from each source are resampled according to the unified time axis, so that each source monitoring data has a corresponding monitoring value at each sampling moment on the unified time axis and retains the corresponding mileage coordinates and circumferential angle, thus obtaining aligned multi-source monitoring data.

[0022] Optionally, S3 includes:

[0023] The mileage coordinates are divided into multiple mileage intervals according to the preset mileage segment intervals, and the circumferential angles are divided into multiple circumferential angle intervals according to the preset circumferential angle intervals. The circumferential angle intervals are connected end to end to satisfy the circumferential period boundary.

[0024] A combination of any mileage interval and any circumferential angle interval is defined as a grid cell, and the cylindrical coordinate grid is constructed from all grid cells.

[0025] For each sampling moment of the aligned multi-source monitoring data on a unified time axis, the grid cell to which each aligned multi-source monitoring data belongs is determined based on the mileage coordinates and circumferential angle corresponding to each data point. The same type of multi-source monitoring data falling into the same grid cell are aggregated to obtain the monitoring value of that type for that grid cell.

[0026] When any grid cell lacks a monitoring value of any type, interpolation is performed to complete the value based on the monitoring value of that type from the adjacent grid cells.

[0027] The various types of monitoring values ​​of each grid cell at each sampling time are organized according to the time dimension, mileage dimension, and circumferential dimension to obtain the gridded observation.

[0028] Optionally, S4 includes:

[0029] The gridded observations are organized into input tensors according to the time dimension, mileage dimension, and circumferential dimension, and the input tensors are input into the solidified quality control model along with the cylindrical coordinate grid.

[0030] The input tensor is feature extracted by the cylindrical coordinate Fourier neural operator model in the solidified quality control model. In the circumferential dimension, the input tensor is subjected to Fourier transform and calculated based on frequency domain convolution that satisfies the periodic boundary condition. In the mileage dimension, the input tensor is segmented according to the preset mileage segment length, and Fourier transform and segmented frequency domain convolution based on non-periodic processing are performed on each segment and then concatenated.

[0031] During the training process of the curing quality control model, the residual terms of the heat conduction equation and the residual terms of the curing reaction kinetics equation corresponding to the output results of the curing quality control model are calculated based on the cylindrical coordinate grid, and the residual terms of the heat conduction equation and the residual terms of the curing reaction kinetics equation are used as physical constraints to train the curing quality control model.

[0032] The trained curing quality control model outputs the initial temperature field and initial degree of curing, and calculates the initial curing rate based on the difference in the initial degree of curing at adjacent sampling times on the same time axis.

[0033] Optionally, S5 includes:

[0034] Within a preset time window on a unified time axis, for each sensor point, the missing rate of that sensor point is calculated based on aligned multi-source monitoring data, and the noise intensity of that sensor point is also calculated. When the sensor point contains temperature data, based on the initial temperature field, the initial temperature field values ​​of the sensor point at its corresponding mileage coordinates and circumferential angles are determined, and the consistency between the temperature data of the sensor point and the initial temperature field values ​​is calculated. The missing rate and the noise intensity are normalized and weighted to obtain a basic weight. When the consistency is calculated, the consistency is normalized and incorporated into the basic weight to obtain the sensor reliability weight. The initial temperature field, the initial curing degree, and the... The initial curing rate at each grid cell of the cylindrical coordinate grid is encoded as a query vector. The observation data of each sensor point in the aligned multi-source monitoring data within the preset time window are encoded as a key vector and a numerical vector. In the residual Transformer network, cross-attention weights are calculated based on the query vector and the key vector. The cross-attention weights are weighted using the sensor reliability weights. The numerical vectors are then weighted and summed using the weighted cross-attention weights to obtain the residual correction amount. The residual correction amount is then superimposed on the initial temperature field, the initial degree of curing, and the initial curing rate to obtain the corrected temperature field, the corrected degree of curing, and the corrected curing rate.

[0035] Optionally, S6 includes:

[0036] The preset judgment range is determined based on a cylindrical coordinate grid, wherein the preset judgment range is a set of grid cells in the cylindrical coordinate grid corresponding to a preset mileage range and a preset circumferential angle range;

[0037] At each sampling moment on the unified time axis, based on the degree of correction of curing, it is determined whether each grid cell within the preset judgment range meets the curing compliance condition, and based on the degree of correction of curing rate, it is determined whether each grid cell within the preset judgment range meets the rate stability condition.

[0038] When all grid cells within the preset determination range simultaneously meet the curing compliance condition and the rate stability condition for a continuous sampling time of no less than the preset duration, the first sampling time that meets the condition is determined as the curing endpoint time, and the preset mileage range is determined as the mileage range corresponding to the curing endpoint time, thus obtaining the curing endpoint determination result.

[0039] Optionally, the S7 includes:

[0040] At each sampling moment on the unified time axis, based on the corrected curing degree and the cylindrical coordinate grid, the grid cells with a corrected curing degree less than the preset curing threshold are identified as under-cured grid cells, and adjacent under-cured grid cells are merged into an under-cured region.

[0041] Mesh cells with a corrected curing degree greater than the preset over-curing threshold are identified as over-cured mesh cells, and adjacent over-cured mesh cells are merged into an over-cured region.

[0042] Calculate the mileage range and circumferential angle range corresponding to each under-cured area and the mileage range and circumferential angle range corresponding to each over-cured area respectively.

[0043] When the area of ​​any under-cured region or any over-cured region is not less than a preset area threshold and remains so for a period of time not less than a preset duration, a positioning warning result is output. The positioning warning result includes the time when the warning was triggered, the region type, and the corresponding mileage range and circumferential angle range.

[0044] Optionally, the residual terms of the heat conduction equation include at least one boundary condition residual, which includes residuals related to the heat transfer boundary conditions of the pipe wall and / or residuals related to the heat transfer boundary conditions of the medium inside the pipe, and when the multi-source monitoring data includes ultraviolet light intensity data, the residual terms of the heat conduction equation further include a volume heat source term determined by the ultraviolet light intensity.

[0045] Optionally, after outputting the curing endpoint determination result and / or the positioning warning result, the method further includes: generating a construction parameter adjustment instruction based on the corrected temperature field, the corrected degree of curing, and / or the corrected curing rate, wherein the construction parameter adjustment instruction is used to adjust at least one of the ultraviolet light source power, medium temperature, medium flow rate, pressure, and traction speed.

[0046] The beneficial effects of this invention are:

[0047] 1. Achieve spatiotemporal alignment and fusion of multi-source monitoring data: By denoising and completing data such as temperature, pressure, flow rate, traction speed, and ultraviolet light intensity, unifying time axis alignment and resampling, and performing grid aggregation and interpolation completion in cylindrical coordinate grid according to mileage and circumference, monitoring data from different sources and with different sampling characteristics can be stably fused in the same spatiotemporal coordinate system, improving the robustness and consistency of subsequent solidification state estimation;

[0048] 2. Real-time output of solidification state along pipeline mileage and circumference with physical consistency: The cylindrical coordinate Fourier neural operator constrained by the residuals of the heat conduction equation and the solidification reaction kinetic equation is used to obtain the gridded distribution of temperature field, degree of solidification and solidification rate. The predicted field is corrected by the observation assimilation residual Transformer, so that the output results simultaneously meet the constraints of field observation and physical law, and improve the accuracy and generalization ability of solidification state distribution estimation.

[0049] 3. Improve the reliability of curing endpoint determination and realize local anomaly location and early warning: By simultaneously introducing curing compliance conditions and rate stability conditions for endpoint determination, the misjudgment caused by relying on a single threshold is avoided; and based on the connected regions of the corrected curing degree on the cylindrical coordinate grid, under-cured and over-cured regions are identified, and the corresponding mileage range and circumferential angle range are output, which can accurately locate local anomalies and provide continuous early warning. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a deep learning-based CIPP solidification quality control method proposed in this invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0053] refer to Figure 1A deep learning-based CIPP solidification quality control method includes:

[0054] S1. Collect multi-source monitoring data during the CIPP curing process, recording timestamps and corresponding mileage coordinates and circumferential angles to obtain raw multi-source monitoring data; S2. Denoise and impute missing values ​​in the raw multi-source monitoring data, perform time alignment and resampling based on timestamps, so that the monitoring data from each source correspond to the same time axis and retain the corresponding mileage coordinates and circumferential angles to obtain aligned multi-source monitoring data; S3. Discretize the mileage coordinates and circumferential angles to construct a cylindrical coordinate grid composed of mileage and circumferential dimensions, converting the aligned multi-source monitoring data into gridded observations indexed by time, mileage, and circumferential dimensions; S4. Input the gridded observations and cylindrical coordinate grid into the curing quality control model, outputting the initial temperature field, initial curing degree, and initial curing rate; S5. Determine the sensor reliability weights for each sensor point based on the aligned multi-source monitoring data, and adjust the initial temperature field, initial curing degree, and initial curing rate accordingly. Using the rate as the query input, the observation data of each sensor point in the aligned multi-source monitoring data is used as the key input and value input. The contribution of cross-attention is weighted using sensor reliability weights. The input residual Transformer network is used to correct the initial temperature field, initial curing degree, and initial curing rate to obtain the corrected temperature field, corrected curing degree, and corrected curing rate. The distribution of corrected curing degree and / or corrected curing rate on the cylindrical coordinate grid is output. S6. Based on the corrected curing degree, corrected curing rate, and cylindrical coordinate grid, each grid cell within the preset judgment range of the cylindrical coordinate grid is judged to meet the curing compliance condition and rate stability condition. When all grid cells within the preset judgment range meet the curing compliance condition and rate stability condition, the curing endpoint judgment result is output. S7. Based on the corrected curing degree and cylindrical coordinate grid, the under-cured area and over-cured area are determined, and the positioning warning result is output.

[0055] In this specific embodiment, S1 includes:

[0056] At the CIPP curing construction site, temperature sensors, pressure sensors, flow sensors, traction speed sensors, and ultraviolet light intensity sensors are laid along the pipeline to be repaired, and all sensors are connected to the same data acquisition terminal to form a multi-source monitoring system.

[0057] The data acquisition terminal sets the preset refresh cycle to... ,in This indicates the time interval between two consecutive output measurements, and each sensor is synchronously sampled once and the corresponding measurement value is output within each refresh cycle;

[0058] The temperature sensor outputs the temperature measurement value of the inner surface of the lining or the interface between the lining and the medium; the pressure sensor outputs the pressure measurement value of the medium in the pipe; the flow sensor outputs the volumetric flow rate measurement value of the medium in the pipe; the traction speed sensor obtains the traction speed measurement value by counting the encoder pulses of the traction length measuring device and converting them with the pulse equivalent of the device; and the ultraviolet light intensity sensor outputs the irradiance measurement value of the ultraviolet light source on the surface of the lining.

[0059] The data acquisition terminal performs range conversion and zero-point calibration on the raw electrical signals of each channel, and writes the sensor point number and sensor type number for each measurement value to ensure that the multi-source monitoring data is logically distinguishable and traceable;

[0060] The data acquisition terminal has a built-in unified clock and completes clock synchronization through a time synchronization module, so that all measured values ​​are generated with the same time base to generate timestamps;

[0061] In the A timestamp is generated at the next sampling time. ,in Indicates the sampling sequence number. This indicates the sampling time corresponding to the sampling sequence number;

[0062] Data acquisition terminal will Each sensor measurement value collected at that sampling time is individually bound to the data to avoid time drift caused by sampling delays of different sensors becoming unrecognizable at the data level.

[0063] The mileage coordinates corresponding to the measured values ​​are determined by using a traction length measuring device, and the position where the pipeline inlet lining enters is defined as the zero point of the mileage coordinates.

[0064] The traction length measuring device outputs the cumulative traction length along the traction direction, and the data acquisition terminal will then output the cumulative traction length. The cumulative traction length at each sampling time is directly recorded as the mileage coordinate. ,in Indicates the timestamp The corresponding coordinates of the reference point at the front end of the lining or the curing equipment in the axial direction of the pipe, and With the same timestamp Temperature, pressure, flow, traction speed, and ultraviolet light intensity measurements are stored synchronously and in association, giving each piece of multi-source monitoring data a unique axial spatial positioning information;

[0065] The circumferential angle corresponding to the measured value is determined by the sensor installation position, and the reference direction of the circumferential angle is defined as the reference direction at the top of the pipe section, and the direction of increasing circumferential angle is defined as the clockwise direction when viewed along the positive direction of the pipe axis.

[0066] For each sensor point Record its fixed circumferential angle after installation. ,in Indicates the sensor point number, This represents the circumferential angle between the sensor point and the reference direction, and It remains unchanged throughout the entire construction process and serves as the circumferential positioning label for all measurements at that sensor point;

[0067] For any sensor point Any sensor type In the The measured value output at the next sampling time is denoted as ,in Indicates the sensor type number, This indicates the time stamp of the sensor point. The measured values ​​below;

[0068] The data acquisition terminal organizes the raw multi-source monitoring data into records:

[0069] ;

[0070] in This represents a single original multi-source monitoring data record, and will By timestamp The data is incrementally written to local storage and synchronously uploaded to the backend database to obtain the original multi-source monitoring dataset containing measurement values, timestamps, mileage coordinates, and circumferential angles.

[0071] In this specific embodiment, S2 includes:

[0072] For the original multi-source monitoring dataset Denoising, missing value completion, time alignment, and resampling are performed, among which... Indicates sensor point Sensor type In sampling sequence number The original record corresponding to the timestamp, and the original record contains at least the measurement value. timestamp Mileage coordinates With circumferential angle ,in Indicates the timestamp The measured value, Indicates the timestamp Corresponding mileage coordinates Indicates sensor point A fixed circumferential angle;

[0073] First, press the sensor point With sensor type All records are sorted in ascending order by timestamp to form a time series. Then, outlier removal and denoising are performed on each time series. Outlier removal uses the Hampel sliding window detection model with a window length set to [value missing]. There are 1 sampling point and the judgment threshold coefficient is set to 1. ,in This represents the number of sampling points within the window used to calculate the local robustness statistic. This represents the threshold coefficient used when comparing the degree of deviation with the degree of local dispersion.

[0074] The sampling points that fall into the anomaly detection are replaced with the median value within this window to complete the outlier removal;

[0075] After outlier removal, the Savitzky-Golay denoising model is executed on each time series, and the window size is set to... And let the order of the polynomial be set as ,in This indicates that the number of consecutive sampling points participating in the least squares fitting is odd. This indicates the order of the fitted polynomial;

[0076] During noise reduction, each time step is compared with the time steps before and after it. Each sampling point was used The least squares method of order polynomial fitting is used, and the fitted value is used to replace the original value to suppress high-frequency noise;

[0077] Subsequently, missing value completion was performed on the denoised monitoring data of each source, with missing value determination based on a preset refresh cycle. Based on, where This represents the time interval between two adjacent expected samples. When the timestamp difference between two adjacent records is greater than... The system determines if there are missing sampling points in the middle and generates a set of missing timestamps.

[0078] For missing time periods not exceeding Missing points are filled using linear interpolation, and for missing time periods exceeding [a certain length], [the missing points are filled using linear interpolation]. The missing points are filled using zero-order hold-for-all, and the filled values ​​are clipped to the corresponding sensor range. This represents the longest consecutive missing duration when using linear interpolation for completion;

[0079] After the data completion is complete, a unified timeline is constructed based on the timestamps of all the original multi-source monitoring data after the completion. ,in Represents the first on the unified timeline There are 1 sampling time points and the interval between adjacent time points is fixed. ;

[0080] For each sensor point With sensor type The time series, its measured time series are time-aligned along a uniform time axis and in each Resampling is performed, and the resampling uses a linear interpolation model to calculate the aligned measurements using the following formula. :

[0081] ;

[0082] in Indicates sensor point Sensor type At a unified timeline Resampled measurements, and These represent the data of this type from the sensor point that satisfy the following conditions: Two adjacent original timestamps, Indicates the original timestamp sequence number. and These represent the measurement values ​​corresponding to the original timestamps;

[0083] when If the timestamp is less than the minimum timestamp of the time series or greater than the maximum timestamp of the time series, then The measured value corresponding to the nearest boundary timestamp is taken to ensure that a monitoring value exists at every sampling moment on a unified time axis;

[0084] The mileage coordinates are also based on the time series corresponding to the traction length measuring device, arranged along a unified time axis. Perform isomorphic outlier removal, denoising, missing value completion, time alignment, and resampling to obtain the mileage coordinates at each sampling time. and circumferential angle As a sensor point The fixed properties remain unchanged throughout the entire time period;

[0085] At each unified timeline moment Will and Associative storage allows monitoring data from various sources to be mapped to the same time axis while retaining the corresponding mileage coordinates and circumferential angles, resulting in aligned multi-source monitoring data.

[0086] In this specific embodiment, S3 includes:

[0087] Using aligned multi-source monitoring data as input, the mileage coordinates and circumferential angles are discretized and a cylindrical coordinate grid is constructed. Then, the aligned multi-source monitoring data is converted into gridded observations indexed by time, mileage, and circumferential dimensions.

[0088] Define the mileage coordinates corresponding to the pipeline inlet as The end mileage coordinates of this construction section are defined as follows: And from the aligned mileage coordinate sequence The maximum value is determined, where Represents the first on the unified timeline Each sampling time, Indicates timestamp as Mileage coordinates at that time;

[0089] Set the preset mileage segment interval to ,in This represents the distance between adjacent mileage grid boundaries, yielding the number of grid cells in the mileage dimension. And satisfy ,in Indicates the number of grid cells in the mileage dimension. Indicates rounding up;

[0090] Define the circumferential angle reference direction as And corresponding to the reference direction at the top of the pipe section, and defining the circumferential angle range as To satisfy the circumferential periodic boundary;

[0091] Set the preset circumferential angle interval to ,in This represents the angle between adjacent circumferential angular grid boundaries, yielding the number of circumferential dimension grids. ,in Indicates the number of grid cells in the circumferential dimension;

[0092] Any mileage range With any circumferential angle interval A combination of these elements is defined as a grid cell, where... This represents the grid index for the mileage dimension and its value range is... Indicates the circumferential dimension grid index and its value range is 1. ;

[0093] For each sampling moment of the unified time axis For each measurement value in the aligned multi-source monitoring data Perform grid positioning, where Indicates sensor point Sensor type At any moment Resampled measurements, Indicates sensor point For a fixed circumferential angle, the grid positioning determines its corresponding grid cell index according to the following rules. :

[0094] ;

[0095] in This indicates rounding down. This indicates that the modulo operation is used to achieve the circumferential angle at... The beginning and end of a cycle are connected in a mapping;

[0096] After completing the positioning, for each moment Each grid cell and each sensor type Resample all measurements that fall into the same grid cell and are of the same sensor type. Aggregate the data and take the arithmetic mean as the monitoring value of this type for the grid cell. During aggregation, record the number of measurements involved in the aggregation for subsequent quality traceability.

[0097] At a certain moment A certain grid cell Missing a certain sensor type When monitoring values, the grid cell is recorded as a missing grid cell of this type, and interpolation is performed based on adjacent grid cells. The adjacent grid cells adopt an 8-neighborhood set that is "adjacent in both the mileage dimension and the circumferential dimension", and are further interpolated in the circumferential dimension according to... Rules for wrapping connections;

[0098] Interpolation completion searches layer by layer according to the neighborhood level. First, it collects grid cells with the same type of monitoring value in the first-level neighborhood and assigns weights to them. The completed value is obtained by performing a weighted average, where the weights are... Let be the reciprocal of the neighborhood step size, and let the neighborhood step size be defined as... This represents the difference in the mileage dimension index between adjacent grid cells and the missing grid cell. This represents the difference in the circumferential dimension index between adjacent grid cells and missing grid cells, calculated using the shortest wraparound distance;

[0099] If no monitoring value of this type exists in the first-level neighborhood, extend to the second-level neighborhood and calculate the complete value according to the same rules. If no monitoring value of this type exists in the second-level neighborhood either, take the monitoring value of this type for that grid cell as the same time. The arithmetic mean of the monitoring values ​​of this type for all non-missing grid cells is used to ensure tensor integrity.

[0100] Each sampling time All grid cells All sensor types The monitored values ​​are organized into a gridded observation tensor according to the time, mileage, and circumferential dimensions. ,in Indicates gridded observation, Indicates at time Grid cells Sensor type The monitored values ​​are used to obtain gridded observations indexed by time, mileage, and circumferential dimensions.

[0101] In this specific embodiment, S4 includes:

[0102] Gridded observation tensor The initial temperature field, initial degree of curing, and initial curing rate are output along with the cylindrical coordinate mesh in the curing quality control model.

[0103] in, Indicates the unified timeline Each sampling time and mileage dimension grid index is Circumferential dimension grid index is Sensor type index is The monitoring value, For sampling sequence number, For mileage dimension grid cell indexing, For circumferential dimension grid cell index, The sensor type index corresponds to temperature, pressure, flow rate, traction speed, and ultraviolet light intensity in this embodiment, respectively.

[0104] The solidified quality control model consists of a cylindrical coordinate Fourier neural operator model and an output header. The cylindrical coordinate Fourier neural operator model uses the multi-channel monitoring values ​​of each grid cell at the same time as input features, and adds normalized coordinate codes of mileage coordinates and circumferential angles as additional channel concatenation inputs to explicitly introduce geometric position information. The normalized coordinate codes employ a mileage dimension... Scalar channel and in the circumferential dimension and Two scalar channels, of which Take the starting point of the mileage coordinates and take Mileage segment intervals, This is the coordinate of the end point of this construction section. The circumferential dimension index is The circumferential angle of the center of the grid cell;

[0105] The network width of the cylindrical coordinate Fourier neural operator model is set to And the number of layers is set to ,in Indicates the number of feature channels. Indicates the number of layers in the spectral convolution layer;

[0106] Each spectral convolutional layer contains a circumferential spectral convolutional branch, a segmented spectral convolutional branch in the equation dimension, and a point convolutional branch, which are then added and fused together. The circumferential spectral convolutional branch has a length of [missing information - likely a number]. Perform real-number fast Fourier transform on the feature sequence and extract low-frequency mode numbers. Then, it is multiplied by the learnable complex weight tensor and then subjected to an inverse transform to satisfy the circumferential periodic boundary condition. The mileage-dimensional piecewise spectral convolution branch will then perform a mileage-dimensional convolution with a length of [missing information]. The characteristic sequence is arranged according to a fixed segment length Divide the segment into segments and fill the last segment with zeros if it is not long enough. ,in Indicates the number of grid cells contained in a mileage segment;

[0107] Perform a real-number fast Fourier transform on each mileage segment and extract the low-frequency mode numbers. Then, the segmented output is multiplied by the learnable complex weight tensor corresponding to the segment and then an inverse transformation is performed to obtain the segmented output. All segmented outputs are then concatenated in the original mileage order to achieve non-periodic segmented frequency domain convolution in the mileage dimension.

[0108] Point convolution branches use A linear transformation is used to mix the channels and add them to the outputs of the two spectral convolution branches. Then, the output of this layer is obtained by passing the GELU nonlinear activation function.

[0109] The output header consists of two parallel, two-layer fully connected networks that share the output features of the cylindrical coordinate Fourier neural operator model. One of the networks outputs the initial temperature field. Furthermore, a linear output is used, with another network outputting the initial curing degree. Furthermore, a Sigmoid function is applied at the output to limit the curing degree. The interval, in which This indicates the initial temperature field at time [time]. With grid cells The value of , Indicates the initial degree of cure at time [time]. With grid cells The possible values ​​of ;

[0110] During the training phase of the solidified quality control model, a joint loss function of "observation fitting loss + physical constraint loss" is used to train the network parameters. The observation fitting loss will... Bilinear interpolation is performed at the mileage coordinates and circumferential angles corresponding to the temperature sensor to obtain the predicted temperature at the sensor location. The mean square error is calculated with the temperature measurement values ​​in the aligned multi-source monitoring data. The physical constraint loss consists of the residual terms of the heat conduction equation and the residual terms of the curing reaction kinetic equation, and both are calculated on the cylindrical coordinate grid.

[0111] The residual terms of the heat conduction equation employ a thin-layer equivalent heat conduction model in cylindrical coordinates. The second derivative in the mileage direction is calculated using a second-order central difference on a discrete grid, and the second derivative in the circumferential direction is calculated using a second-order central difference satisfying the circumferential periodic boundary. The time derivative is calculated using a first-order backward difference. The gridded ultraviolet intensity corresponding to the ultraviolet light intensity channel is used as a volume heat source term in the residual calculation. This volume heat source term converts the ultraviolet light intensity into unit volume heat generation power with a linear proportional relationship, and the conversion factor is set to [value missing] in this embodiment. ,in It represents the coefficient for converting surface irradiance to volume heat source and is determined by the absorption rate of the lining material to ultraviolet light and the thickness of the lining.

[0112] The residual terms of the heat conduction equation further include boundary condition residuals, which include residuals related to the heat transfer boundary conditions of the pipe wall and residuals related to the heat transfer boundary conditions of the medium inside the pipe. The heat transfer boundary conditions of the pipe wall adopt a convective heat transfer model and the convective heat transfer coefficient is set to... The heat transfer boundary conditions of the medium inside the pipe adopt a convective heat transfer model and the convective heat transfer coefficient is set to 0. ,in This represents the convective heat transfer coefficient between the pipe wall and the external environment. This represents the convective heat transfer coefficient between the inner surface of the lining and the medium inside the pipe;

[0113] The residual terms of the curing reaction kinetic equation are calculated using an Arrhenius-type autocatalytic curing kinetic model. The time derivative of the degree of curing is calculated on a discrete grid using first-order backward difference, and the residual is calculated with the curing rate given by the kinetic model. In this embodiment, the parameters of the kinetic model are set as pre-exponential factors. ,activation energy Gas constant Reaction order and ,in This represents the pre-exponential factor of the reaction rate. Indicates the activation energy of the reaction. Represents the gas constant. and Indicates the reaction order in the curing kinetics model;

[0114] The solidified quality control model was trained using the Adam optimizer with the learning rate set to [value missing]. The number of training rounds was set to 200, the batch size was set to 8, and the weights of the observation fitting loss and the physical constraint loss were both set to 1 to ensure that the observation consistency and physical consistency converged simultaneously.

[0115] During the model inference phase, the trained and solidified quality control model is used for each time step. Perform forward computation to obtain and The initial curing rate was calculated based on the difference in initial curing degree at adjacent sampling times on the same time axis. The formula for its calculation is:

[0116] ;

[0117] in Indicates the initial curing rate at time [time]. With grid cells The value of , and These represent the initial degree of curing at adjacent sampling times. This represents the time interval between adjacent sampling moments on a unified time axis, and is set to 1 second. Moment Set to 0 to complete the definition of the initial curing rate.

[0118] In this specific embodiment, S5 includes:

[0119] Based on aligned multi-source monitoring data and initial temperature field Initial curing degree and initial curing rate Perform observation assimilation correction to obtain the corrected temperature field Correcting the degree of curing and correct curing rate ;

[0120] in, and Representing the unified timeline Each sampling time and mileage dimension grid index is Circumferential dimension grid index is The initial temperature field value, the initial degree of cure value, and the initial curing rate value are determined. Indicates the sampling sequence number. This represents a grid index based on mileage dimensions. Indicates the circumferential dimension grid index;

[0121] Aligning multi-source monitoring data It means that among them Indicates the sensor point number, Indicates the sensor type number, Indicates the first The timestamp of each sampling moment Indicates sensor point Sensor type In timestamp The corresponding resampled measurement values, and for each sensor point With a fixed circumferential angle And at every moment Associated mileage coordinates ;

[0122] Use preset time window length Each sampling point and the current time The right endpoint of the window is used to define the time window as... ,in This represents the number of sampling points used in the time window for calculating reliability and coded observations;

[0123] During the missing value completion process, for each Synchronously store completion flags to indicate whether the measurement value was obtained from field measurement or generated by a completion algorithm for each sensor point. Calculate the missing rate within the time window. ,in This indicates the number of sampling points marked as completed within the time window for that sensor point, compared to the total number of sampling points in the time window. The ratio, and when the sensor point It has multiple sensor types At that time, the arithmetic mean of the missing rates corresponding to each type is taken as the missing rate of that sensor point. ;

[0124] For each sensor point Calculate noise intensity within the time window ,in This represents the standard deviation of the first-order difference sequence of the resampled measurements at that sensor point within the time window, and when the sensor point... It has multiple sensor types At that time, the arithmetic mean of the standard deviations corresponding to each type is taken as the noise intensity of that sensor point. ;

[0125] When sensor point It contains temperature data and its temperature type number is denoted as At that time, based on the initial temperature field At the sensor point The predicted temperature value is calculated at the mileage coordinates and circumferential angle position and compared with the value. A consistency index is obtained by comparing points one by one within the time window. ,in This represents the score obtained by normalizing and inversely mapping the average absolute error between the sensor's temperature measurement and the initial temperature field prediction within the time window. The range of values ​​is and The larger the value, the higher the consistency.

[0126] For missing rate Noise intensity Consistency Indicators The min-max normalization is performed over a time window across the entire set of sensor points to obtain... and ,in This represents the normalized missing rate. This represents the normalized noise intensity. This represents the normalized consistency index, and when sensor points... When temperature data is not included The value is set to 0.5 to ensure that this item makes a neutral contribution to the reliability weight;

[0127] Then, for each sensor point Calculate sensor reliability weights And the observation assimilation weights used within this time window are fixed, along with the sensor reliability weights. Calculate using the following formula:

[0128] ;

[0129] in, Indicates sensor point The sensor reliability weight and its value range is This represents the weighting coefficient for the missing rate term. This represents the weighting coefficient for the noise intensity term. This represents the weighting coefficient of the consistency term. Indicates sensor point Normalized missing rate, Indicates sensor point Normalized noise intensity, Indicates sensor point Normalized consistency index;

[0130] In obtaining Then, the initial temperature field Initial curing degree With initial curing rate The values ​​at each grid cell in the cylindrical coordinate grid are encoded as query vectors, and the observation data of each sensor point within the time window in the aligned multi-source monitoring data are encoded as key vectors and numerical vectors, where the query vector is based on each grid cell. At the present moment triples Normalized encoding of the center coordinates of the grid cell The components are assembled and mapped to the model dimension via a linear layer. ,in The coordinates of the center mileage of the grid cell are represented by The normalized scalar, Indicates the circumferential angle of the center of the grid cell. Indicates the mileage coordinates at the end of the construction section;

[0131] Key vectors and numerical vectors for each sensor point Time window observation matrix constitute, and Represents the relative sequence number within the time window and satisfies This represents the total number of sensor types and is taken in this embodiment. ;

[0132] Will Input a time-series encoder to extract sensor points The observation representation vector, wherein the time series encoder consists of two one-dimensional convolutional layers and the number of input channels of the first convolutional layer is . The first convolutional layer has 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The second convolutional layer has 64 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. A GELU activation function is used between the two convolutional layers, and global average pooling is performed on the temporal dimension after the second convolutional layer to obtain a 64-dimensional vector. This 64-dimensional vector is then encoded with the sensor point positions. The concatenation is then mapped to a key vector and a numerical vector via two independent linear layers, respectively. Indicates sensor point At the present moment Mileage coordinates according to Normalized scalar;

[0133] The query vector, key vector, and value vector are input into a residual Transformer network for cross-attention assimilation. The residual Transformer network employs multiple layers. Multiple heads Hidden Dimensions in Feedforward Layer The calculation is performed within each layer in the following order: "Cross-attention sublayer—Residual connection and layer normalization—Feedforward sublayer—Residual connection and layer normalization," where the cross-attention sublayer uses the grid cell query vector as... And using the sensor point key vector as And using the sensor point numerical vector as ;

[0134] When calculating cross-attention, each sensor point The corresponding attention score is multiplied by the sensor reliability weight for that sensor point. Then, Softmax normalization is performed, which allows sensor points with higher reliability to make a greater contribution to the assimilation and suppresses the influence of sensor points with lower reliability on the assimilation.

[0135] The output of the residual Transformer network is for each mesh cell at the current time. The residual correction latent vector is mapped to temperature residual correction, curing degree residual correction, and curing rate residual correction through a three-channel output head. The three-channel output head consists of a linear layer. GELU activation function and linear layer The three scalars that constitute and output correspond to respectively and ,in This represents the residual correction amount for the temperature field. This indicates the amount of residual correction for degree of cure. This indicates the residual correction amount for the curing rate;

[0136] Will and Superimposed on each and get and and will Cut off to The interval satisfies the physical value range of the degree of cure, and the corrected temperature field, corrected degree of cure, and corrected curing rate are obtained, and the distribution of the corrected degree of cure and the corrected curing rate on the cylindrical coordinate grid are output.

[0137] In this specific embodiment, S6 includes:

[0138] Based on the correction degree of curing Correcting the curing rate The solidification endpoint is determined by performing a solidification process on the cylindrical coordinate mesh.

[0139] in Indicates the unified timeline Each sampling time The lower mileage dimension grid index is And the circumferential dimension grid index is The corrected curing degree value, Indicates the unified timeline Each sampling time The lower mileage dimension grid index is And the circumferential dimension grid index is The corrected curing rate is taken as a value and the unit is . Indicates the sampling sequence number. Indicates the grid cell index of the mileage dimension. Indicates the circumferential dimension grid cell index;

[0140] A preset judgment range is determined based on a cylindrical coordinate grid, and the preset judgment range corresponds to a preset mileage range. With preset circumferential angle range ,in Indicates the starting point of the mileage coordinates and takes This represents the mileage coordinates at the end of the construction section and is determined by the mesh construction in step S3. Indicates the circumferential angular periodic boundary;

[0141] Therefore, the preset judgment range covers all grid cells and corresponds to all indexes in the mileage dimension. All indexes in the circumferential dimension ,in Indicates the number of grid cells in the mileage dimension. Indicates the number of grid cells in the circumferential dimension;

[0142] Define the curing compliance conditions as Not less than the preset curing threshold ,in This indicates the solidification compliance threshold and is dimensionless.

[0143] Define the rate stability condition as The absolute value is not greater than the preset rate stabilization threshold. ,in Represents the rate stability threshold and its unit is . ;

[0144] At each sampling moment on the unified time axis For each grid cell within the preset judgment range Simultaneously, it is determined whether the curing compliance condition and the rate stability condition are met, and the joint judgment rule is as follows:

[0145] ;

[0146] in This represents the logical AND operator. This represents the absolute value operation;

[0147] If and only if all grid cells within the preset judgment range are sampled at the same time... When all conditions are met according to the above joint determination rules, the sampling time is recorded as the "satisfaction time", and a continuous satisfaction counter is maintained. ,in Indicates the time up to the sampling point The number of consecutive satisfying times, and when To meet the time and season requirements ,when Not meeting the time season ;

[0148] Set the preset duration to ,in This represents the shortest duration for continuously meeting the requirements, and since the adjacent sampling interval of the unified time axis in step S2 is... ,in This indicates the sampling time interval, therefore it will be continuous for no less than The determination is equivalent to continuous not less than The determination of each sampling time, among which This indicates the number of sampling points corresponding to the duration;

[0149] When it first appears When the conditions are met, the first sampling time is determined as the solidification endpoint time. ,in Indicates the curing endpoint time. This indicates a backward regression from the current sampling time. Sampling time after each sampling interval;

[0150] Simultaneously, the preset mileage range Determined as the curing endpoint time The corresponding mileage range is then output, along with the curing endpoint determination result, which includes the curing endpoint time. Preset mileage range With preset circumferential angle range .

[0151] In this specific embodiment, S7 includes:

[0152] At each sampling moment on the unified time axis Based on the correction degree of curing The process involves using a cylindrical coordinate mesh to determine under-cured and over-cured regions and outputting positioning warning results.

[0153] in Indicates time The lower mileage dimension grid index is And the circumferential dimension grid index is The corrected curing degree value, Indicates the sampling sequence number. Indicates the grid cell index of the mileage dimension. Indicates the circumferential dimension grid cell index;

[0154] The cylindrical coordinate grid is determined by step S3, and the mileage segment interval is... Circumferential angle interval is The starting point of the mileage coordinates is The range of circumferential angles is And satisfy the circumferential periodic boundary;

[0155] Set the preset curing compliance threshold to And set the preset over-curing threshold to ,in This represents the dimensionless threshold used to determine under-cured properties. This represents the degree of cure threshold used to determine over-curing and is dimensionless;

[0156] At each sampling time For all grid cells Performing cell-by-cell classification will satisfy The mesh elements are marked as undercured mesh elements, which will satisfy... The grid cells are marked as over-cured grid cells, and the remaining grid cells are marked as normal grid cells;

[0157] Within the same sampling time, connected component merging is performed on the under-cured and over-cured mesh cell sets respectively. Connectivity is determined by the four-adjacency rule, and the cells are connected end-to-end along the periodic boundary in the circumferential dimension. The four-adjacency rule is defined as any mesh cell... and Adjacent grid cells and circumferential indexed by right Modulus extraction enables wraparound connections. Indicates the number of grid cells in the circumferential dimension;

[0158] Region merging employs a breadth-first search to traverse adjacent grid cells with the same type of label and assigns a region number to each connected component, thus achieving the desired result at time [time value missing]. Obtain the sets of under-cured regions and the sets of over-cured regions;

[0159] For each region, calculate its corresponding mileage range and circumferential angle range, where the mileage range is determined by the minimum mileage index of all grid cells contained in that region. With maximum mileage index Determine and output as The circumferential angle range is determined and output by the minimum covering arc on the circumference of the entire set of circumferential indices contained in the region. and and All fell Inside, among which Indicates the starting angle of the minimum coverage arc. This represents the minimum coverage arc termination angle. The minimum coverage arc is obtained by sorting the center angles of all grid cells corresponding to all circumferential indices in the region in ascending order and finding the largest adjacent angle gap, and then taking its supplementary arc. This ensures that when crossing the 0 circumferential boundary, it still outputs a unique and shortest circumferential angle range.

[0160] The area of ​​each region is calculated and used for early warning triggering. The area is determined by the number of grid cells contained in the region. With unit area Multiplying them together, we get: Indicates the number of grid cells in the region. This represents the surface area of ​​the lining corresponding to a single grid cell, and Calculate using the following formula:

[0161] ;

[0162] in Indicates the interval between mileage segments. This represents the inner radius of the pipe to be repaired, which is calculated from the pipe diameter measured before construction and is taken as the inner radius in this embodiment. Indicates the circumferential angular interval;

[0163] Set the preset area threshold to ,in This represents the minimum area required to trigger an alert.

[0164] Set the preset duration to And it is sampled at a uniform time axis interval The threshold for the number of continuous sampling points is obtained through conversion. ,in Indicates the minimum duration required to trigger an alert. Indicates the time interval between adjacent sampling times. Indicates the minimum number of continuous sampling points;

[0165] To achieve persistence determination, the same type of region association is performed for each region between adjacent sampling times and the persistence count is accumulated. The region association is based on the condition that "the intersection of the region grid cell sets is not empty" as the determination condition for the same continuous region. When the association is true, the persistence count of the continuous region is incremented by one. When the association is false, the persistence count is set to zero and a new persistence count is established for the newly appearing region.

[0166] When any under-cured region or any over-cured region satisfies the condition that its area is not less than 1 at any sampling time. And its continuous count is not less than At that time, a location warning result is output, which includes the time when the warning was triggered. Regional types Mileage range Circumferential angle range ,in This indicates the sampling time at which the area and duration conditions are first met. Indicates the region type and specifies either under-cured or over-cured. This indicates the mileage range corresponding to this area. This indicates the range of circumferential angles corresponding to this region.

[0167] In this specific embodiment, after outputting the curing endpoint determination result and the positioning warning result, the construction control module generates construction parameter adjustment instructions and sends them to the ultraviolet light source power supply, the medium heating unit, the medium circulation pump and regulating valve, the pressure regulating valve and the traction drive to achieve closed-loop adjustment.

[0168] The construction control module is deployed within an industrial computer and samples moments along a unified timeline. As the instruction refresh time, among which Indicates the first The timestamp of each sampling moment Indicates the sampling sequence number;

[0169] When step S7 is at time... When outputting location warning results, the set of grid cells in the cylindrical coordinate grid of the area that triggered the warning is defined as the adjustment area. and in the adjustment area Internal correction of curing degree The arithmetic mean was used to obtain the regional average degree of curing. ,in Indicates time The lower mileage dimension grid index is And the circumferential dimension grid index is The corrected curing degree value, This represents a grid index based on mileage dimensions. Indicates the circumferential dimension grid index. Indicates the area to be adjusted At any moment Average degree of cure;

[0170] When step S6 is at time... When outputting the curing endpoint determination result, the set of grid cells corresponding to the preset determination range in step S6 is defined as the adjustment region. Calculate in the same way The curing endpoint time is recorded as follows: To trigger the endpoint closing control, where Indicates the curing endpoint time;

[0171] The construction control module sets the curing degree reference value to And calculate the curing deviation. ,in Indicates time curing deviation and Indicating a trend of under-consolidation This indicates a trend towards solidification;

[0172] Under conditions of under-curing or over-curing, the construction control module is based on the curing deviation. Incremental commands are generated for the ultraviolet light source power setting, medium temperature setting, medium flow rate setting, pressure setting, and traction speed setting. The increment is calculated as follows:

[0173] ;

[0174] in This indicates the increment of the ultraviolet light source power setting value, expressed as a percentage point. Indicates the increment of the medium temperature setpoint and the unit is Indicates the increment of the medium flow rate setpoint, and the unit is... Indicates the pressure setpoint increment and the unit is Indicates the increment of the traction speed setpoint, and the unit is . This represents the power proportionality coefficient of the ultraviolet light source. Indicates the proportionality coefficient of the medium temperature. This represents the proportionality coefficient of the medium flow rate. This represents the pressure proportionality coefficient. This represents the traction speed proportionality coefficient. Indicates curing deviation;

[0175] The construction control module reads the current set value. , and Each of these values ​​is added to its corresponding increment to obtain a new setpoint. Then, a limit is applied to this new setpoint to meet the equipment's safety boundaries. Specifically, the ultraviolet light source power setpoint is limited to... Percentage point, medium temperature setpoint limit to Medium flow rate setpoint limit to Pressure setpoint limit to Traction speed setpoint limit to ;

[0176] When the endpoint closing control is triggered and the endpoint time is fixed. When output is complete, the construction control module, without changing the amplitude limiting rules, directly sets the traction speed setting to 0 and reduces the ultraviolet light source power setting to 0 in a ramp manner, decreasing by 20 percentage points with each refresh until it reaches 0. Simultaneously, the medium temperature setting decreases with each refresh. The slope is lowered until And maintain the pressure setting value at no less than 0.05 MPa to ensure the liner fits and cools and forms.

[0177] Construction parameter adjustment instructions are output in the form of instruction frames and include the instruction generation timestamp. The system sends the target device identifier, parameter name, and new settings to the corresponding actuator via industrial Ethernet to adjust at least one of the following: UV light source power, medium temperature, medium flow rate, pressure, and traction speed.

[0178] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0179] This invention addresses the technical problems in CIPP curing quality control, namely, the "dispersed and noisy multi-source monitoring data, difficulty in spatiotemporal alignment and fusion, difficulty in obtaining the curing state field along the mileage and circumference in real time and reliably determining the curing endpoint, and difficulty in locating local under-curing and over-curing." It employs a joint modeling approach of "physically constrained neural operators and data-driven residual Transformers": First, the neural operator maps multi-source monitoring information into field quantities such as temperature field, curing degree, and curing rate on a cylindrical coordinate grid. During the training phase, residuals from the heat conduction equation and the curing reaction kinetic equation are introduced to ensure the output satisfies the physical consistency of the curing process, thereby improving generalization ability under varying operating conditions. Subsequently, cross-attention assimilation is performed using the grid field as the query and sensor point observations as the key and value. Sensor reliability weights are introduced to suppress the influence of anomalies and low-reliability observations on correction, achieving real-time residual correction of the initial field. This allows for stable output of the curing state along the mileage and circumference even under noisy, missing, and asynchronous sampling conditions, providing a reliable basis for endpoint determination and anomaly warning.

[0180] In terms of algorithm structure, this invention makes specific improvements to address the characteristics of "coexistence of cylindrical pipeline geometry and circumferential periodicity with mileage aperiodicity": In the neural operator, spectral convolution satisfying periodic boundary conditions is used for the circumferential dimension, and piecewise Fourier transform and aperiodic piecewise frequency domain convolution are used for the mileage dimension, followed by concatenation, to better match the characteristics of the solidification thermal field being continuously closed along the circumference and segmentally changing along the mileage; at the same time, the determination of the solidification endpoint is endogenized into a joint output of solidification degree and solidification rate, and the stopping is determined by superimposing the solidification compliance condition and the rate stability condition, avoiding misjudgment and lag caused by relying solely on a single threshold; furthermore, based on the regional connectivity of the corrected solidification degree on the cylindrical coordinate grid, the positioning and early warning of under-solidification and over-solidification are realized, thereby more effectively achieving the technical effects of real-time quality control along the mileage and circumference, reliable endpoint determination, and accurate local anomaly positioning.

Claims

1. A CIPP solidification quality control method based on deep learning, characterized in that, include: S1. Collect multi-source monitoring data during the CIPP curing process, record the timestamps and corresponding mileage coordinates and circumferential angles to obtain the raw multi-source monitoring data; S2. Denoise and impute missing values ​​in the original multi-source monitoring data, perform time alignment and resampling based on timestamps to map the monitoring data of each source to the same time axis and retain the corresponding mileage coordinates and circumferential angles, thus obtaining aligned multi-source monitoring data; S3. Discretize the mileage coordinates and circumferential angles to construct a cylindrical coordinate grid composed of mileage and circumferential dimensions, converting the aligned multi-source monitoring data into gridded observations indexed by time, mileage, and circumferential dimensions; S4. Input the gridded observations and cylindrical coordinate grid into the solidification quality control model, and output the initial temperature field, initial solidification degree, and initial solidification rate; S5. Based on the aligned multi-source monitoring data, determine the sensor reliability weights for each sensor point. Use the initial temperature field, initial curing degree, and initial curing rate as query inputs, and the observation data of each sensor point in the aligned multi-source monitoring data as key and value inputs. Use the sensor reliability weights to weight the contribution of cross-attention, and input the residual Transformer network to correct the initial temperature field, initial curing degree, and initial curing rate to obtain the corrected temperature field, corrected curing degree, and corrected curing rate. Output the distribution of corrected curing degree and / or corrected curing rate on the cylindrical coordinate grid. S6. Based on the corrected curing degree, corrected curing rate, and cylindrical coordinate grid, determine whether each grid cell within the preset judgment range of the cylindrical coordinate grid meets the curing compliance condition and the rate stability condition. When all grid cells within the preset judgment range meet the curing compliance condition and the rate stability condition, output the curing endpoint judgment result. S7. Based on the correction degree of curing and the cylindrical coordinate grid, determine the under-cured area and the over-cured area, and output the positioning warning result.

2. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S1 includes: During the CIPP curing process, measurements are taken using at least two types of sensors, including those for acquiring temperature data, pressure data, flow rate data, traction speed data, and ultraviolet light intensity data, and the measured values ​​are output at the preset refresh cycle. Assign a timestamp to each measurement; The mileage coordinates corresponding to the measured value are determined based on the traction length measuring device or odometer; The circumferential angle corresponding to the measured value is determined based on the sensor's installation location; The measured values ​​are associated with and stored with the corresponding timestamps, mileage coordinates, and circumferential angles to obtain the original multi-source monitoring data.

3. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S2 include: Outlier removal and denoising are performed on the original multi-source monitoring data to obtain the denoised original multi-source monitoring data. The missing values ​​of the denoised original multi-source monitoring data are filled in to obtain the filled original multi-source monitoring data. A unified timeline is determined based on the timestamps of the completed original multi-source monitoring data, and the completed original multi-source monitoring data is time-aligned according to the unified timeline. The time-aligned monitoring data from each source are resampled according to the unified time axis, so that each source monitoring data has a corresponding monitoring value at each sampling moment on the unified time axis and retains the corresponding mileage coordinates and circumferential angle, thus obtaining aligned multi-source monitoring data.

4. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S3 include: The mileage coordinates are divided into multiple mileage intervals according to the preset mileage segment intervals, and the circumferential angles are divided into multiple circumferential angle intervals according to the preset circumferential angle intervals. The circumferential angle intervals are connected end to end to satisfy the circumferential period boundary. A combination of any mileage interval and any circumferential angle interval is defined as a grid cell, and the cylindrical coordinate grid is constructed from all grid cells. For each sampling moment of the aligned multi-source monitoring data on a unified time axis, the grid cell to which each aligned multi-source monitoring data belongs is determined based on the mileage coordinates and circumferential angle corresponding to each data point. The same type of multi-source monitoring data falling into the same grid cell are aggregated to obtain the monitoring value of that type for that grid cell. When any grid cell lacks a monitoring value of any type, interpolation is performed to complete the value based on the monitoring value of that type from the adjacent grid cells. The various types of monitoring values ​​of each grid cell at each sampling time are organized according to the time dimension, mileage dimension, and circumferential dimension to obtain the gridded observation.

5. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S4 includes: The gridded observations are organized into input tensors according to the time dimension, mileage dimension, and circumferential dimension, and the input tensors are input into the solidified quality control model along with the cylindrical coordinate grid. The input tensor is feature extracted by the cylindrical coordinate Fourier neural operator model in the solidified quality control model. In the circumferential dimension, the input tensor is subjected to Fourier transform and calculated based on frequency domain convolution that satisfies the periodic boundary condition. In the mileage dimension, the input tensor is segmented according to the preset mileage segment length, and Fourier transform and segmented frequency domain convolution based on non-periodic processing are performed on each segment and then concatenated. During the training process of the curing quality control model, the residual terms of the heat conduction equation and the residual terms of the curing reaction kinetics equation corresponding to the output results of the curing quality control model are calculated based on the cylindrical coordinate grid, and the residual terms of the heat conduction equation and the residual terms of the curing reaction kinetics equation are used as physical constraints to train the curing quality control model. The trained curing quality control model outputs the initial temperature field and initial degree of curing, and calculates the initial curing rate based on the difference in the initial degree of curing at adjacent sampling times on the same time axis.

6. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S5 include: Within a preset time window of a unified time axis, for each sensor point, the missing rate of the sensor point is calculated based on the aligned multi-source monitoring data, and the noise intensity of the sensor point is also calculated. When the sensor point contains temperature data, the initial temperature field value of the sensor point at its corresponding mileage coordinates and circumferential angle is determined based on the initial temperature field, and the consistency between the temperature data of the sensor point and the initial temperature field value is calculated. The missing rate and the noise intensity are normalized and weighted to obtain the basic weights; When the consistency is calculated, the consistency is normalized and incorporated into the basic weight to obtain the sensor reliability weight; The values ​​of the initial temperature field, the initial degree of curing, and the initial curing rate at each grid cell of the cylindrical coordinate grid are encoded as query vectors. The observation data of each sensor point in the aligned multi-source monitoring data within the preset time window are encoded as key vectors and numerical vectors. In the residual Transformer network, cross-attention weights are calculated based on the query vectors and the key vectors. The cross-attention weights are weighted using the sensor reliability weights. The numerical vectors are then weighted and summed using the weighted cross-attention weights to obtain the residual correction amount. The residual correction amount is then superimposed on the initial temperature field, the initial degree of curing, and the initial curing rate to obtain the corrected temperature field, the corrected degree of curing, and the corrected curing rate.

7. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S6 include: The preset judgment range is determined based on a cylindrical coordinate grid, wherein the preset judgment range is a set of grid cells in the cylindrical coordinate grid corresponding to a preset mileage range and a preset circumferential angle range; At each sampling moment on the unified time axis, based on the degree of correction of curing, it is determined whether each grid cell within the preset judgment range meets the curing compliance condition, and based on the degree of correction of curing rate, it is determined whether each grid cell within the preset judgment range meets the rate stability condition. When all grid cells within the preset determination range simultaneously meet the curing compliance condition and the rate stability condition for a continuous sampling time of no less than the preset duration, the first sampling time that meets the condition is determined as the curing endpoint time, and the preset mileage range is determined as the mileage range corresponding to the curing endpoint time, thus obtaining the curing endpoint determination result.

8. The CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, S7 includes: At each sampling moment on the unified time axis, based on the corrected curing degree and the cylindrical coordinate grid, the grid cells with a corrected curing degree less than the preset curing threshold are identified as under-cured grid cells, and adjacent under-cured grid cells are merged into an under-cured region. Mesh cells with a corrected curing degree greater than the preset over-curing threshold are identified as over-cured mesh cells, and adjacent over-cured mesh cells are merged into an over-cured region. Calculate the mileage range and circumferential angle range corresponding to each under-cured area and the mileage range and circumferential angle range corresponding to each over-cured area respectively. When the area of ​​any under-cured region or any over-cured region is not less than a preset area threshold and remains so for a period of time not less than a preset duration, a positioning warning result is output. The positioning warning result includes the time when the warning was triggered, the region type, and the corresponding mileage range and circumferential angle range.

9. A deep learning-based CIPP solidification quality control method according to claim 5, characterized in that, The residual terms of the heat conduction equation include at least one boundary condition residual, which includes residuals related to the heat transfer boundary conditions of the pipe wall and / or residuals related to the heat transfer boundary conditions of the medium inside the pipe. When the multi-source monitoring data includes ultraviolet light intensity data, the residual terms of the heat conduction equation further include a volume heat source term determined by the ultraviolet light intensity.

10. A CIPP solidification quality control method based on deep learning according to claim 1, characterized in that, After outputting the curing endpoint determination result and / or the positioning warning result, the method further includes: generating a construction parameter adjustment instruction based on the corrected temperature field, the corrected degree of curing and / or the corrected curing rate, wherein the construction parameter adjustment instruction is used to adjust at least one of the ultraviolet light source power, medium temperature, medium flow rate, pressure and traction speed.

Images