Underground pipeline detection method and system based on motor vehicle traction

Abstract

The invention discloses an underground pipeline detection method and system based on motor vehicle traction. The method comprises the following steps: step 1, collecting vehicle state multi-physics field joint signals and preprocessing the signals to generate a standardized sequence; step 2, inputting the potential vector sequence, the vehicle speed and the attitude into a multi-physics field controllable Kupman self-encoding network, and outputting a potential vector sequence; 3, linear propagation is executed to calculate a trigger interval correction value; 4, adjusting the radar frequency and the electromagnetic phase according to the correction value, collecting a new signal, and calculating a residual error; 5, when the residual exceeds the threshold value, updating the propagation and control matrix and adjusting the vehicle speed and the chassis height; step 6, iterating the above process until residual convergence, and obtaining an aligned signal sequence; and 7, inputting an identification operator to generate an underground pipeline three-dimensional center line model, and outputting a detection result. According to the invention, high-precision alignment of multi-physical field signals and accurate detection of three-dimensional modeling of underground pipelines are realized.

Description

Technical Field

[0001] This invention relates to the field of underground space detection technology, and in particular to a method and system for detecting underground pipelines based on a motor vehicle towing system. Background Technology

[0002] Existing underground pipeline detection technologies mainly rely on single sensor devices such as ground-penetrating radar, electromagnetic induction instruments, and seismic wave sensors. They acquire signal data using methods such as manual carts, handheld devices, or vehicle-mounted platforms, and then use offline signal processing algorithms for target identification and centerline positioning. In complex urban geological environments, these technologies are easily affected by ground structures, interfering power sources, and signal aliasing, leading to a decrease in signal-to-noise ratio, blurred reflection characteristics, and difficulty in accurately distinguishing various underground targets. Traditional systems often rely on fixed-frequency sampling and lack the ability to dynamically control the accuracy of data acquisition under motion conditions. This can easily lead to problems such as trigger frequency drift, electromagnetic excitation delay, and vibration response deviation, affecting the accuracy of multi-source data alignment and comprehensive judgment.

[0003] Existing methods lack structured modeling capabilities in multi-physics information fusion, failing to establish physically consistent signal evolution relationships. This results in the underutilization of spatiotemporal coupling relationships between data from different sensors, limiting modeling accuracy and detection stability. Existing modeling methods are mostly based on black-box deep networks or empirical formulas, making it difficult to provide highly interpretable and generalizable prediction mechanisms. Especially in areas where signals are disturbed, the models often exhibit degraded generalization performance. Existing systems lack closed-loop iterative optimization mechanisms, making it impossible to dynamically adjust sampling strategies and model parameters when detection errors are large, hindering the achievement of accuracy self-calibration and state adaptation.

[0004] Therefore, how to provide a method and system for detecting underground pipelines based on motor vehicle towing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for detecting underground pipelines based on motor vehicle towing. This invention integrates multi-physics joint signal modeling, adaptive sampling control, and conservation constraint state identification technology, and describes in detail the entire process of multi-source signal alignment, trigger compensation, and three-dimensional centerline fitting. It has the advantages of high detection accuracy, strong data robustness, and adaptability to complex environments.

[0006] A method for detecting underground pipelines based on a vehicle-to-motorized system according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect vehicle-state multi-physics joint signals and perform preprocessing to generate standardized multi-physics signal sequences; Step 2: Input the standardized multiphysics signal sequence along with vehicle speed information and traction attitude information into the multiphysics controllable Koopman autoencoder network. The multiphysics controllable Koopman autoencoder network consists of a convolutional encoder, a linear propagation matrix, a control matrix, and a decoder. The latent vector sequence is output through the convolutional encoder. Step 3: Perform a linear propagation on the latent vector sequence using the linear propagation matrix and control matrix within the multiphysics controllable Koopman autoencoder network to obtain the predicted latent vector sequence and calculate the trigger interval correction. Step 4: After adjusting the radar pulse trigger frequency and electromagnetic excitation phase according to the trigger interval correction, acquire the next frame of multiphysics field signal and calculate the residual between the actual multiphysics field signal frame and the predicted multiphysics field signal frame. Step 5: When the residual is greater than the preset threshold, execute the adaptive Koopman parameter update process to update the propagation matrix and control matrix, while reducing the vehicle speed and lowering the deformable chassis height to generate a corrected potential vector sequence. Step 6: Repeat steps 2 to 5 until the residual is no greater than the preset threshold, and integrate to obtain the aligned multi-physics signal sequence; Step 7: Input the aligned multiphysics signal sequence into the conserved constraint Koopman identification operator to generate a three-dimensional centerline model of the underground pipeline and output the underground pipeline detection results.

[0008] Optionally, in step one, the vehicle-state multi-physics joint signal consists of vehicle speed data, traction attitude data, ground-penetrating radar electromagnetic scattering signal, low-frequency electromagnetic induction signal, and surface wave vibration acceleration signal. The preprocessing sequentially performs bandpass filtering for noise reduction, time-scaled unified interpolation, coordinate system transformation synchronization, and amplitude normalization operations to output a standardized multi-physics signal sequence. The standardized multi-physics signal sequence is arranged according to a unified time index and consists of a speed value subsequence, an attitude value subsequence, a normalized electromagnetic scattering waveform subsequence, a normalized electromagnetic induction phase subsequence, and a normalized surface wave acceleration subsequence.

[0009] Optionally, the multiphysics controllable Koopman autoencoder network in step two is composed of a three-branch convolutional encoder, a latent vector construction unit, a linear propagation matrix, a control matrix, and a symmetric structure decoder, specifically including:

[0010] The three-branch convolutional encoder includes an electromagnetic scattering coding branch, an electromagnetic induction coding branch, and a surface wave vibration coding branch.

[0011] The electromagnetic scattering coding branch is used to receive the normalized electromagnetic scattering waveform subsequence and extract local frequency domain features; the electromagnetic induction coding branch is used to receive the normalized electromagnetic induction phase subsequence and extract phase change features; and the surface wave vibration coding branch is used to receive the normalized surface wave acceleration subsequence and extract low-frequency vibration timing features.

[0012] The latent vector construction unit is used to concatenate the feature output of the three-branch convolutional encoder with the velocity numerical subsequence and the attitude numerical subsequence according to time steps to generate a latent vector sequence.

[0013] The linear propagation matrix is ​​used to perform a linear propagation operation on the potential vector sequence, the control matrix is ​​used to convert the velocity numerical subsequence and attitude numerical subsequence into control inputs, and the joint propagation matrix performs linear evolution on the potential vector sequence.

[0014] The symmetric structure decoder is used to receive the propagated potential vector sequence and reconstruct the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence, and the predicted surface wave acceleration subsequence according to the original three-branch structure, and output the predicted multiphysics signal frame.

[0015] Optionally, the calculation of the trigger interval correction amount in step three specifically includes:

[0016] The predicted latent vector sequence is input into the symmetric structure decoder to reconstruct the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence, and the predicted surface wave acceleration subsequence, respectively, forming a predicted multiphysics signal frame.

[0017] The difference between the predicted time interval of adjacent main reflection peaks in the predicted electromagnetic scattering waveform subsequence and the actual time interval of adjacent main reflection peaks in the actual electromagnetic scattering waveform subsequence is calculated, and the electromagnetic scattering triggering error is output.

[0018] The predicted occurrence time of the phase change point in the predicted electromagnetic induction phase subsequence is compared with the actual occurrence time of the phase change point in the actual electromagnetic induction phase subsequence by calculating the time offset, and the electromagnetic induction excitation delay is output.

[0019] The difference between the predicted start time of the main amplitude rise segment in the predicted surface wave acceleration subsequence and the actual start time of the main amplitude rise segment in the actual surface wave acceleration subsequence is calculated, and the vibration response offset is output.

[0020] The electromagnetic scattering triggering error, electromagnetic induction excitation delay, and vibration response offset are weighted and fused according to their respective signal energy weights to generate the trigger interval correction.

[0021] Optionally, step four specifically includes:

[0022] The ground-penetrating radar pulse trigger frequency is adjusted according to the electromagnetic scattering correction component in the trigger interval correction, the low-frequency electromagnetic excitation phase is adjusted according to the electromagnetic induction correction component in the trigger interval correction, and the surface wave vibration sampling time reference is adjusted according to the surface wave vibration correction component in the trigger interval correction to complete the trigger compensation for the next frame of multi-physics field acquisition.

[0023] After the compensation is triggered, the next frame of vehicle-state multi-physics joint signal is acquired, and the next frame of vehicle-state multi-physics joint signal is preprocessed to generate the next frame of standardized multi-physics signal sequence.

[0024] The next frame of standardized multiphysics signal sequence is input into the multiphysics controllable Koopman autoencoder network, and processed by the convolutional encoder, linear propagation matrix, control matrix and symmetric structure decoder to generate the corresponding predicted multiphysics signal frame.

[0025] Align the next frame of the normalized multiphysics signal sequence with the predicted multiphysics signal frame point by point according to a unified time index, calculate the Euclidean residual distance between the velocity numerical subsequence, attitude numerical subsequence, normalized electromagnetic scattering waveform subsequence, normalized electromagnetic induction phase subsequence and normalized surface wave acceleration subsequence respectively, and output the whole frame residual of the next frame.

[0026] Optionally, step five specifically includes:

[0027] Construct a potential vector sequence sample set containing the current time and the previous window of length N. Fit the potential vector sequence sample set using the minimum mean square error criterion, calculate the propagation error vector, and update the propagation matrix element by element-wise gradient descent based on the propagation error vector to obtain the updated propagation matrix.

[0028] Extract the velocity and attitude numerical subsequences within the time period corresponding to the potential vector sequence sample set, construct the control input sample set, calculate the control input error using the minimum control residual criterion, update the corresponding elements in the control matrix, and obtain the updated control matrix.

[0029] The updated propagation matrix and the updated control matrix are input into the multiphysics controllable Koopman autoencoder network, linear propagation operation is performed, and the corrected latent vector sequence is output.

[0030] Vehicle control adjustment operations are performed based on the magnitude of the residual in different physical field channels, and a dynamic compensation strategy based on residual component classification is implemented.

[0031] The corrected potential vector sequence and the updated velocity and attitude numerical subsequences are input into the multiphysics controllable Koopman autoencoder network, and the process of steps three to five is continued.

[0032] Optionally, the dynamic compensation strategy based on residual component classification specifically includes:

[0033] If only the electromagnetic scattering residual exceeds the threshold, a speed adjustment command is sent to the vehicle control module to set the traction speed to 80% of the current speed while keeping the chassis height unchanged.

[0034] If only the electromagnetic induction residual exceeds the threshold, a height adjustment command is sent to the chassis control module to set the deformable chassis height to 70% of the current height while maintaining the traction speed.

[0035] If only the surface wave vibration residual exceeds the threshold, a speed adjustment command is sent to the vehicle control module to set the traction speed to 85% of the current speed, and a command is sent to the time base synchronization module to recalibrate the sampling reference time of the surface wave signal.

[0036] If both the electromagnetic scattering residual and the electromagnetic induction residual exceed the threshold, adjustment commands are sent to the vehicle control module and the chassis control module respectively, setting the traction speed to 75% of the current speed and the chassis height to 70% of the current height.

[0037] If both the electromagnetic scattering residual and the surface wave vibration residual exceed the threshold, the traction speed is set to 75% of the current speed, and a command is sent to the time base synchronization module to improve the sampling time accuracy from 1 millisecond to 0.1 millisecond. At the same time, a command is sent to the radar control module to improve the radar trigger compensation frequency from 100 Hz to 200 Hz.

[0038] If both the electromagnetic induction residual and the surface wave vibration residual exceed the threshold, the chassis height will be set to 70% of the current height, and surface wave sampling time reference calibration and electromagnetic excitation phase compensation operations will be performed simultaneously.

[0039] If the electromagnetic scattering residual, electromagnetic induction residual, and surface wave vibration residual all exceed the threshold, the traction speed is set to 70% of the current speed, the chassis height is set to 60% of the current height, and the three types of signal acquisition and compensation operations are performed simultaneously.

[0040] If the electromagnetic scattering residual, electromagnetic induction residual, and surface wave vibration residual do not exceed the threshold, but the residual increase exceeds 10% for three consecutive frames, a fine-tuning command is sent to the vehicle control module to set the traction speed to 95% of the current speed and simultaneously increase the multi-physics field signal sampling frequency by 5%.

[0041] Optionally, the generation of the aligned multiphysics signal sequence in step six includes:

[0042] The modified latent vector sequence generated in the previous round is input again into the multiphysics controllable Koopman autoencoder network to reconstruct the predicted multiphysics signal frame;

[0043] The residuals are calculated based on the predicted multiphysics signal frames and the current actual acquisition frames. According to the distribution characteristics of the three types of physical components in the residuals, the corresponding dynamic compensation strategy based on residual component classification is selected to update and correct the latent vector sequence.

[0044] The difference between the current round residual and the previous round residual is calculated to obtain the residual change rate, which is compared with the preset convergence rate threshold. If the residual change rate is less than the convergence rate threshold for three consecutive rounds, or the absolute value of the current round residual is lower than the preset residual threshold, then convergence is determined.

[0045] When the convergence condition is met, the iteration process is terminated, the current modified latent vector sequence is taken as the final latent state representation, and the final predicted multiphysics signal frame is generated based on it. The time axis alignment operation is performed with the actual acquisition frame to construct the aligned multiphysics signal sequence.

[0046] Optionally, the conservation constraint Koopman identification operator in step seven includes a feature encoding unit, a linear propagation unit, and a centerline fitting unit, specifically including:

[0047] The aligned multiphysics signal sequence is divided into fixed-length time intervals, and the multiphysics spatial feature vectors of each time interval are extracted by the feature encoding unit. The spatial feature vectors are then input into the linear propagation unit to establish a state propagation matrix and superimpose energy conservation terms and phase conservation terms to constrain and correct the state propagation trajectory.

[0048] The constrained and corrected state propagation trajectory is input into the centerline fitting unit. Based on the regression of the node centroid trajectory, the three-dimensional centerline coordinate sequence of the underground pipeline is fitted, and the underground pipeline detection results are output.

[0049] According to an embodiment of the present invention, an underground pipeline detection system based on a motor vehicle towed vehicle includes the following modules:

[0050] Multi-physics signal acquisition module, used to acquire combined multi-physics signals of vehicle status;

[0051] The signal standardization module is used to normalize the time scale and amplitude of preprocessed multi-source signals to generate standardized multi-physics signal sequences.

[0052] The multiphysics controllable Koopman modeling module is used to receive standardized multiphysics signal sequences and vehicle state information and extract potential vector sequences;

[0053] The linear propagation and control evolution module is used to perform linear propagation of the latent vector sequence, generate predicted latent vectors, and calculate the trigger interval correction.

[0054] The trigger compensation control module is used to adjust the radar frequency, electromagnetic excitation phase, and vibration sampling time reference based on the trigger interval correction amount.

[0055] The multiphysics acquisition and residual analysis module is used to acquire a new frame of signal and calculate the multidimensional residual between the predicted frame and the actual frame.

[0056] The Koopman parameter adaptive update module is used to update the propagation matrix and control matrix based on residual feedback, and adjust the traction speed and chassis height.

[0057] The residual iterative convergence control module is used to control the step iteration. When the residual is lower than the threshold, it outputs an aligned multiphysics signal sequence.

[0058] The underground pipeline identification and modeling module is used to input alignment signals, perform conservation constraint Koopman identification, and output three-dimensional centerline detection results.

[0059] The beneficial effects of this invention are:

[0060] (1) Improve the alignment accuracy of multi-physics field signals. This invention constructs a multi-physics field controllable Koopman autoencoder network to achieve unified modeling and linear evolution of electromagnetic scattering, low-frequency electromagnetic induction and surface wave vibration signals. Combined with velocity and attitude control inputs for dynamic propagation, it improves the alignment accuracy of multi-source heterogeneous physical signals on the time axis and avoids the signal mismatch problem caused by triggering error and time base drift in traditional methods.

[0061] (2) Real-time adaptive sampling compensation is achieved. In the acquisition process, the present invention introduces a trigger interval correction mechanism. By calculating the electromagnetic trigger error, phase delay and vibration response offset, the radar pulse frequency, electromagnetic excitation phase and vibration sampling reference are adjusted in a closed loop to form an iterative optimization process oriented towards residual drive, thereby enhancing the real-time stability and sampling robustness of the system in complex geological environments.

[0062] (3) To ensure the spatial structure accuracy of underground pipeline detection results, by introducing energy conservation and phase conservation terms into the Koopman identification operator, the physical consistency of the state propagation trajectory is constrained. Based on the node centroid trajectory, three-dimensional regression fitting is performed to improve the spatial structure accuracy of the final output three-dimensional centerline model of underground pipeline, which is suitable for the fine modeling needs of various types of underground structures. Attached Figure Description

[0063] 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:

[0064] Figure 1 This is an overall flowchart of an underground pipeline detection method based on motor vehicle towing proposed in this invention;

[0065] Figure 2 This is a module connection diagram of an underground pipeline detection system based on a motor vehicle towing mechanism proposed in this invention;

[0066] Figure 3This is a Koopman autoencoder network structure diagram of an underground pipeline detection method based on motor vehicle towing proposed in this invention;

[0067] Figure 4 This is a block diagram of the parameter update and control strategy for an underground pipeline detection method based on motor vehicle towing proposed in this invention. Detailed Implementation

[0068] 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.

[0069] refer to Figure 1-4 A method for detecting underground pipelines based on motor vehicle towing includes the following steps:

[0070] Step 1: Collect vehicle-state multi-physics joint signals and perform preprocessing to generate standardized multi-physics signal sequences; Step 2: Input the standardized multiphysics signal sequence along with vehicle speed information and traction attitude information into the multiphysics controllable Koopman autoencoder network. The multiphysics controllable Koopman autoencoder network consists of a convolutional encoder, a linear propagation matrix, a control matrix, and a decoder. The latent vector sequence is output through the convolutional encoder. Step 3: Perform a linear propagation on the latent vector sequence using the linear propagation matrix and control matrix within the multiphysics controllable Koopman autoencoder network to obtain the predicted latent vector sequence and calculate the trigger interval correction. Step 4: After adjusting the radar pulse trigger frequency and electromagnetic excitation phase according to the trigger interval correction, acquire the next frame of multiphysics field signal and calculate the residual between the actual multiphysics field signal frame and the predicted multiphysics field signal frame. Step 5: When the residual is greater than the preset threshold, execute the adaptive Koopman parameter update process to update the propagation matrix and control matrix, while reducing the vehicle speed and lowering the deformable chassis height to generate a corrected potential vector sequence. Step 6: Repeat steps 2 to 5 until the residual is no greater than the preset threshold, and integrate to obtain the aligned multi-physics signal sequence; Step 7: Input the aligned multiphysics signal sequence into the conserved constraint Koopman identification operator to generate a three-dimensional centerline model of the underground pipeline and output the underground pipeline detection results.

[0071] In this embodiment, the vehicle-state multi-physics joint signal in step one consists of vehicle speed data, traction attitude data, ground-penetrating radar electromagnetic scattering signal, low-frequency electromagnetic induction signal, and surface wave vibration acceleration signal. The preprocessing sequentially performs bandpass filtering denoising, time-scaled unified interpolation, coordinate system transformation synchronization, and amplitude normalization operations to output a standardized multi-physics signal sequence. The standardized multi-physics signal sequence is arranged according to a unified time index and consists of a speed value subsequence, an attitude value subsequence, a normalized electromagnetic scattering waveform subsequence, a normalized electromagnetic induction phase subsequence, and a normalized surface wave acceleration subsequence.

[0072] In this embodiment, the multiphysics controllable Koopman autoencoder network in step two is composed of a three-branch convolutional encoder, a latent vector construction unit, a linear propagation matrix, a control matrix, and a symmetric structure decoder, specifically including:

[0073] The three-branch convolutional encoder includes an electromagnetic scattering coding branch, an electromagnetic induction coding branch, and a surface wave vibration coding branch.

[0074] The electromagnetic scattering coding branch is used to receive the normalized electromagnetic scattering waveform subsequence and extract local frequency domain features; the electromagnetic induction coding branch is used to receive the normalized electromagnetic induction phase subsequence and extract phase change features; and the surface wave vibration coding branch is used to receive the normalized surface wave acceleration subsequence and extract low-frequency vibration timing features.

[0075] The electromagnetic scattering encoding branch includes multiple one-dimensional convolutional layers, batch normalization layers, and activation function layers. It receives normalized electromagnetic scattering waveform subsequences in chronological order, performs fixed-length sliding convolution operations on the electromagnetic reflection waveforms within each time window, extracts reflection intensity envelope features, main frequency amplitude features, and instantaneous frequency change features, and outputs a sequence of electromagnetic scattering radio frequency domain feature vectors.

[0076] The electromagnetic induction coding branch includes a phase demodulation preprocessing unit and a multi-layer one-dimensional convolution module. It receives the normalized electromagnetic induction phase sub-sequences in time steps, first performs local phase unwrapping and phase difference operations on each sampling point to obtain a continuous phase change curve, and then extracts the phase jump amplitude features and phase change rate features through the one-dimensional convolution module to output the electromagnetic induction phase feature vector sequence.

[0077] The surface wave vibration coding branch consists of a sliding window segmenter, a low-pass convolutional unit, and a temporal feature extraction convolutional unit. It continuously receives normalized surface wave acceleration subsequences, segments each acceleration data according to a set window, inputs it into the low-pass convolutional unit to extract the main waveband features, and uses the temporal feature extraction convolutional unit to extract low-frequency dynamic features such as the surface wave amplitude envelope, the main amplitude attenuation rate, and the vibration duration, and outputs a surface wave vibration temporal feature vector sequence.

[0078] The latent vector construction unit is used to concatenate the feature output of the three-branch convolutional encoder with the velocity numerical subsequence and the attitude numerical subsequence according to time steps to generate a latent vector sequence.

[0079] The linear propagation matrix is ​​used to perform a linear propagation operation on the potential vector sequence, the control matrix is ​​used to convert the velocity numerical subsequence and attitude numerical subsequence into control inputs, and the joint propagation matrix performs linear evolution on the potential vector sequence.

[0080] The linear propagation operation models the relationship between the latent vector at the current time step and the latent vector at the previous time step as a linear transformation of a fixed dimension by setting a propagation matrix. Specifically, it performs a matrix multiplication operation on the latent vector at each time step and outputs a linear propagation sequence of latent vectors. The propagation matrix can be updated during training to maintain a linear approximate expression of the latent dynamics evolution process.

[0081] The control matrix is ​​used to concatenate the velocity numerical subsequence and the attitude numerical subsequence by time step to form a state control vector. In each time step, the state control vector is input into the control matrix for linear mapping to generate a control input feature vector that matches the dimension of the potential vector.

[0082] The linear evolution process involves applying the propagation matrix to the current latent vector and adding the control input feature vector under the control matrix. This completes the linear propagation and state control adjustment of the latent vector at the current time step, enabling the latent vector sequence to continuously evolve according to the changes in control variables within a unified latent space, and outputting the predicted latent vector sequence.

[0083] The symmetric structure decoder is used to receive the propagated potential vector sequence and reconstruct the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence, and the predicted surface wave acceleration subsequence according to the original three-branch structure, and output the predicted multiphysics signal frame.

[0084] Specifically, this includes: receiving the frequency domain feature dimension of the propagated potential vector sequence in the electromagnetic scattering decoding branch, gradually restoring the time dimension resolution through multi-layer one-dimensional deconvolution operation, reconstructing the predicted electromagnetic scattering waveform subsequence, and restoring the intensity distribution of the reflected signal from the underground medium;

[0085] The electromagnetic induction decoding branch receives the phase feature dimension in the propagated latent vector sequence, performs one-dimensional deconvolution and inverse phase difference integration operations, and restores the predicted latent features to a continuously changing predicted electromagnetic induction phase subsequence, characterizing the induction response trend of the underground metal pipe.

[0086] The low-frequency dynamic feature dimension of the latent vector sequence after propagation is received by the surface wave vibration decoding branch. One-dimensional deconvolution and reverse envelope construction operations are used to reconstruct the predicted surface wave acceleration subsequence, thereby restoring the vibration energy distribution and foundation continuity of the target area.

[0087] The predicted sub-sequences output by the three branch decoders are spliced ​​together according to a unified time index to form a predicted multiphysics signal frame.

[0088] In this embodiment, the calculation of the trigger interval correction amount in step three specifically includes:

[0089] The predicted latent vector sequence is input into the symmetric structure decoder to reconstruct the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence, and the predicted surface wave acceleration subsequence, respectively, forming a predicted multiphysics signal frame.

[0090] The difference between the predicted time interval of adjacent main reflection peaks in the predicted electromagnetic scattering waveform subsequence and the actual time interval of adjacent main reflection peaks in the actual electromagnetic scattering waveform subsequence is calculated, and the electromagnetic scattering triggering error is output.

[0091] In the electromagnetic scattering waveform subsequence, the main reflection peak usually represents the highly reflected echo signal of the underground structure interface or object boundary, and its time domain location has obvious peak characteristics. The process of calculating the electromagnetic scattering triggering error includes: firstly, extracting the predicted time points of two adjacent main reflection peaks in the predicted electromagnetic scattering waveform subsequence and calculating the predicted time interval; then, extracting the actual time points of the corresponding two adjacent main reflection peaks in the actual electromagnetic scattering waveform subsequence and calculating the actual time interval; and finally, performing a difference operation between the predicted time interval and the actual time interval to obtain the electromagnetic scattering triggering error, which is used to measure the degree of time drift of the radar pulse in the current state.

[0092] The predicted occurrence time of the phase change point in the predicted electromagnetic induction phase subsequence is compared with the actual occurrence time of the phase change point in the actual electromagnetic induction phase subsequence by calculating the time offset, and the electromagnetic induction excitation delay is output.

[0093] In the electromagnetic induction phase subsequence, the phase abrupt change point usually reflects the phase jump phenomenon of the excitation electromagnetic field's response to the underground conductor, and has stable structural timing characteristics. The process of calculating the electromagnetic induction excitation delay includes: first, detecting the predicted occurrence time of the phase abrupt change point in the predicted electromagnetic induction phase subsequence, and then detecting the actual occurrence time of the corresponding abrupt change point in the actual electromagnetic induction phase subsequence; by comparing the predicted time with the actual time one by one and obtaining the time offset, the electromagnetic induction excitation delay is output, which is used to quantify the time domain error of the signal response under the current electromagnetic excitation control.

[0094] The difference between the predicted start time of the main amplitude rise segment in the predicted surface wave acceleration subsequence and the actual start time of the main amplitude rise segment in the actual surface wave acceleration subsequence is calculated, and the vibration response offset is output.

[0095] In the surface wave acceleration subsequence, the rising segment of the main amplitude reflects the first effective response of the surface or underground structure caused by traction vibration. The start time of the rising segment usually corresponds to the starting point of the initial energy coupling. To calculate the vibration response offset, it is necessary to locate the predicted start time of the rising segment of the main amplitude in the predicted surface wave acceleration subsequence and extract the corresponding actual start time in the actual surface wave acceleration subsequence. By calculating the time difference between the two, the vibration response offset is obtained, which is used to estimate the synchronization error between vehicle vibration and underground structure response.

[0096] The electromagnetic scattering triggering error, electromagnetic induction excitation delay, and vibration response offset are weighted and fused according to their respective signal energy weights to generate the trigger interval correction.

[0097] In this embodiment, step four specifically includes:

[0098] The ground-penetrating radar pulse trigger frequency is adjusted according to the electromagnetic scattering correction component in the trigger interval correction, the low-frequency electromagnetic excitation phase is adjusted according to the electromagnetic induction correction component in the trigger interval correction, and the surface wave vibration sampling time reference is adjusted according to the surface wave vibration correction component in the trigger interval correction to complete the trigger compensation for the next frame of multi-physics field acquisition.

[0099] In this embodiment, to address the issue of inconsistent acquisition times between ground-penetrating radar, electromagnetic induction, and surface wave vibration signals, the trigger interval correction amount generated in step three is used to adjust the triggering mechanism of each physical field during the acquisition of the next frame of data, thereby achieving dynamic correction of synchronous sampling accuracy. The correction amount includes:

[0100] Electromagnetic scattering correction component ;

[0101] Electromagnetic induction correction component ;

[0102] Surface wave vibration correction component ;

[0103] The specific adjustment process is as follows:

[0104] I. Adjustment of the pulse trigger frequency of electromagnetic scattering signals:

[0105] The original pulse trigger frequency of the ground-penetrating radar system was: The pulse time interval is When a difference is detected between the predicted interval and the actual interval of the main reflection peak When the error is used as the trigger frequency adjustment amount, compensation is performed using the following formula:

[0106] ;

[0107] frequency It is written in real time to the radar pulse controller to drive the transmission of the next frame of radar waves.

[0108] For example, if the current radar trigger frequency is 500Hz (i.e. Since the main reflection peak has a prediction delay of 0.3ms, we have:

[0109] ;

[0110] II. Excitation Phase Adjustment of Electromagnetic Induction Signal:

[0111] For low-frequency electromagnetic excitation signals, modulation is typically performed in the form of a sine wave or sawtooth wave, and the initial phase of the excitation signal is set to be... When the predicted phase abrupt change point differs from the actual observation in time... At that time, in order to avoid misalignment of the sensing response, Converted into excitation phase adjustment amount And modify the start phase in the excitation control command:

[0112] ;

[0113] in, The period of the electromagnetic excitation signal.

[0114] For example, if the excitation period is 20ms and the detection delay for the phase transition point is 5ms, then:

[0115] ;

[0116] Phase adjustment is used to modulate the starting position of the waveform to ensure that the excitation signal and the target structure's response time are consistent.

[0117] III. Adjustment of sampling time reference for surface wave vibration signals:

[0118] Surface wave vibration signal acquisition typically uses a fixed sampling period, with the starting sampling timestamp being... If there is a deviation between the predicted and actual starting time of the main amplitude rise segment... Then this offset is used as the sampling reference delay term: ;

[0119] That is, update the sampling start time of the vibration sampling module to maintain phase alignment of the target signal;

[0120] For example, if the current starting sampling time is ms, If the prediction is too early (ms), then:

[0121] ;

[0122] The sampling controller will start the sampling process in advance to capture the complete rising waveform.

[0123] IV. Joint Execution of Triggered Compensation:

[0124] After completing the above three adjustments, update parameters are sent to the ground-penetrating radar control unit, the electromagnetic excitation modulator, and the vibration signal acquisition and control module, respectively. , , The system controls the data acquisition module to perform the synchronously corrected multi-physics data acquisition task, generates a new standardized multi-physics signal frame, and enters the next round of residual evaluation process.

[0125] After the compensation is triggered, the next frame of vehicle-state multi-physics joint signal is acquired, and the next frame of vehicle-state multi-physics joint signal is preprocessed to generate the next frame of standardized multi-physics signal sequence.

[0126] The next frame of standardized multiphysics signal sequence is input into the multiphysics controllable Koopman autoencoder network, and processed by the convolutional encoder, linear propagation matrix, control matrix and symmetric structure decoder to generate the corresponding predicted multiphysics signal frame.

[0127] Align the next frame of the normalized multiphysics signal sequence with the predicted multiphysics signal frame point by point according to a unified time index, calculate the Euclidean residual distance between the velocity numerical subsequence, attitude numerical subsequence, normalized electromagnetic scattering waveform subsequence, normalized electromagnetic induction phase subsequence and normalized surface wave acceleration subsequence respectively, and output the whole frame residual of the next frame.

[0128] In this embodiment, step five specifically includes:

[0129] Construct a potential vector sequence sample set containing the current time and the previous window of length N. Fit the potential vector sequence sample set using the minimum mean square error criterion, calculate the propagation error vector, and update the propagation matrix element by element-wise gradient descent based on the propagation error vector to obtain the updated propagation matrix.

[0130] Extract the velocity and attitude numerical subsequences within the time period corresponding to the potential vector sequence sample set, construct the control input sample set, calculate the control input error using the minimum control residual criterion, update the corresponding elements in the control matrix, and obtain the updated control matrix.

[0131] Specifically, the minimum control residual criterion refers to minimizing the control input error when updating the control matrix. It involves calculating the residual tensor between the latent vector sequence and the control input, and then adjusting the elements of the control matrix using the least squares optimization method to minimize the sum of squared control errors between the predicted latent state and the actual propagation result. The minimum control residual criterion ensures that the control matrix has optimal matching in the feedback adjustment process, improving the controllability and prediction accuracy of Koopman linear propagation.

[0132] The updated propagation matrix and the updated control matrix are input into the multiphysics controllable Koopman autoencoder network, linear propagation operation is performed, and the corrected latent vector sequence is output.

[0133] Vehicle control adjustment operations are performed based on the magnitude of the residual in different physical field channels, and a dynamic compensation strategy based on residual component classification is implemented.

[0134] The corrected potential vector sequence and the updated velocity and attitude numerical subsequences are input into the multiphysics controllable Koopman autoencoder network, and the process of steps three to five is continued.

[0135] In this embodiment, the dynamic compensation strategy based on residual component classification specifically includes:

[0136] If only the electromagnetic scattering residual exceeds the threshold, a speed adjustment command is sent to the vehicle control module to set the traction speed to 80% of the current speed while keeping the chassis height unchanged.

[0137] If only the electromagnetic induction residual exceeds the threshold, a height adjustment command is sent to the chassis control module to set the deformable chassis height to 70% of the current height while maintaining the traction speed.

[0138] If only the surface wave vibration residual exceeds the threshold, a speed adjustment command is sent to the vehicle control module to set the traction speed to 85% of the current speed, and a command is sent to the time base synchronization module to recalibrate the sampling reference time of the surface wave signal. The time base synchronization module is used to uniformly manage the sampling time reference of various multi-physics field signals. This module receives the system's global time pulse signal and performs time base calibration on the samplers of subsystems such as ground-penetrating radar, electromagnetic induction, and surface wave vibration, ensuring that different types of signals are collected in alignment under the same time index, supporting the accuracy requirements of residual comparison, synchronous modeling, and control compensation.

[0139] If both the electromagnetic scattering residual and the electromagnetic induction residual exceed the threshold, adjustment commands are sent to the vehicle control module and the chassis control module respectively, setting the traction speed to 75% of the current speed and the chassis height to 70% of the current height.

[0140] If both the electromagnetic scattering residual and the surface wave vibration residual exceed the threshold, the traction speed is set to 75% of the current speed, and a command is sent to the time base synchronization module to improve the sampling time accuracy from 1 millisecond to 0.1 milliseconds. At the same time, a command is sent to the radar control module to improve the radar trigger compensation frequency from 100 Hz to 200 Hz. The radar control module is used to control the transmission and reception process of the ground-penetrating radar. Based on the pulse trigger frequency adjustment command from the host computer, it dynamically configures the radar pulse transmission frequency, period, and threshold parameters, and coordinates the signal acquisition channel and storage buffer in real time to ensure that the ground-penetrating radar signal is acquired with high precision according to the corrected time interval.

[0141] If both the electromagnetic induction residual and the surface wave vibration residual exceed the threshold, the chassis height will be set to 70% of the current height, and surface wave sampling time reference calibration and electromagnetic excitation phase compensation operations will be performed simultaneously.

[0142] Specifically, the surface wave sampling time reference calibration operation refers to adjusting the trigger clock start point of the surface wave signal sampling module according to the magnitude of the vibration response offset. If the vibration response offset exceeds ±2 milliseconds, the sampling start time is moved forward or backward by the corresponding offset value to achieve waveform alignment. For example, when the offset is +3ms, the surface wave sampling trigger reference is delayed by 3ms to ensure that the main amplitude rise segment is aligned with the time axis of historical sampling data.

[0143] Electromagnetic excitation phase compensation operation refers to adjusting the output initial phase angle of the low-frequency electromagnetic excitation source according to the electromagnetic induction excitation delay. If the delay is Δθ (in milliseconds), the phase drift of the corresponding excitation waveform is calculated according to the corresponding time offset, φ=2πfΔθ, and the electromagnetic excitation initial phase angle is corrected to the original angle minus φ, thereby compensating for the phase misalignment caused by the excitation response delay.

[0144] If the electromagnetic scattering residual, electromagnetic induction residual, and surface wave vibration residual all exceed the threshold, the traction speed is set to 70% of the current speed, the chassis height is set to 60% of the current height, and the three types of signal acquisition and compensation operations are performed simultaneously.

[0145] Specifically, the three types of signal acquisition and compensation operations target three main sub-sequences in the standardized multiphysics signal sequence: the electromagnetic scattering waveform sub-sequence, the electromagnetic induction phase sub-sequence, and the surface wave acceleration sub-sequence. The following acquisition and compensation strategies are executed respectively:

[0146] Electromagnetic scattering acquisition compensation operation: Adjust the transmission time reference of the ground-penetrating radar pulse according to the electromagnetic scattering trigger error. Specifically, if the current trigger error is +Δt, the transmission time of the next ground-penetrating radar pulse will be delayed by Δt to ensure that the reflection peak time window is aligned with the prediction and to avoid early or late triggering that causes echo waveform drift.

[0147] Electromagnetic induction acquisition and compensation operation: Adjust the starting phase of the low-frequency electromagnetic source excitation according to the electromagnetic induction excitation delay. Specifically, convert the delay into the phase drift value at the corresponding excitation frequency, modify the initial phase angle of the excitation source, and achieve consistency between the time-domain excitation waveform and the predicted position of the phase change point.

[0148] Surface wave vibration acquisition compensation operation: Adjust the sampling clock reference of the surface wave sampling module according to the vibration response offset; specifically: when the offset is positive (i.e. the predicted main amplitude starts earlier than the actual amplitude), the sampling start time is delayed by the corresponding milliseconds, and vice versa, so as to achieve time window alignment of the main vibration response segment.

[0149] If the electromagnetic scattering residual, electromagnetic induction residual, and surface wave vibration residual do not exceed the threshold, but the residual increase exceeds 10% for three consecutive frames, a fine-tuning command is sent to the vehicle control module to set the traction speed to 95% of the current speed and simultaneously increase the sampling frequency of the multi-physics field signal by 5% to suppress the residual growth trend.

[0150] In this embodiment, the generation of the aligned multiphysics signal sequence in step six includes:

[0151] The modified latent vector sequence generated in the previous round is input again into the multiphysics controllable Koopman autoencoder network to reconstruct the predicted multiphysics signal frame;

[0152] The residuals are calculated based on the predicted multiphysics signal frames and the current actual acquisition frames. According to the distribution characteristics of the three types of physical components in the residuals, the corresponding dynamic compensation strategy based on residual component classification is selected to update and correct the latent vector sequence.

[0153] The difference between the current round residual and the previous round residual is calculated to obtain the residual change rate, which is compared with the preset convergence rate threshold. If the residual change rate is less than the convergence rate threshold for three consecutive rounds, or the absolute value of the current round residual is lower than the preset residual threshold, then convergence is determined.

[0154] When the convergence condition is met, the iteration process is terminated, the current modified latent vector sequence is taken as the final latent state representation, and the final predicted multiphysics signal frame is generated based on it. The time axis alignment operation is performed with the actual acquisition frame to construct the aligned multiphysics signal sequence.

[0155] In this embodiment, the conservation constraint Koopman identification operator in step seven includes a feature encoding unit, a linear propagation unit, and a centerline fitting unit, specifically including:

[0156] The aligned multiphysics signal sequence is divided into fixed-length time intervals, and the multiphysics spatial feature vectors of each time interval are extracted by the feature encoding unit. The spatial feature vectors are then input into the linear propagation unit to establish a state propagation matrix and superimpose energy conservation terms and phase conservation terms to constrain and correct the state propagation trajectory.

[0157] Among them, the energy conservation term is defined as the square of the normalized signal energy difference between adjacent multiphysics field signal frames, which constrains the stability of the overall energy during propagation; the phase conservation term is defined as the change in the main phase of the electromagnetic induction signal between adjacent frames, which constrains the continuity and interpretability of phase evolution.

[0158] The two conservation terms mentioned above are incorporated into the objective function in a weighted form, and together with the prediction residual term, they constitute a joint optimization objective. After solving for the optimal state propagation matrix, a linear transformation is performed on the potential state sequence. Under the premise of maintaining the minimum propagation error, the physical rationality of the state trajectory in the energy and phase dimensions is ensured, and the constraint correction of the propagation trajectory is achieved.

[0159] To achieve physical consistency constraints on the state propagation trajectory, the following objective function is constructed. :

[0160] ;

[0161] in, Indicates time The potential state vector at time t; Here is the state propagation matrix. For the control matrix, The control input consists of velocity and attitude information; Indicates the first The normalized energy value of the multi-physics field signal is calculated as follows: ,in The amplitude of the signal in each channel; Indicates the first The dominant phase of the frame electromagnetic induction signal; The weighting coefficient controls the degree of influence of the conservation term on the total loss function.

[0162] The three terms in the objective function correspond to the optimization objectives of state propagation accuracy, energy conservation, and phase continuity, respectively. These are achieved by minimizing... The optimal propagation matrix can be obtained. With control matrix This allows for the conservation constraints and corrections on the state propagation trajectory.

[0163] The constrained and corrected state propagation trajectory is input into the centerline fitting unit. Based on the regression of the node centroid trajectory, the three-dimensional centerline coordinate sequence of the underground pipeline is fitted, and the underground pipeline detection results are output.

[0164] In this embodiment, after completing the energy conservation and phase conservation constraint correction of the state propagation trajectory, the constrained state propagation trajectory is input into the centerline fitting unit. The centroid position of the multi-physics field signal coupling response region at each time moment is extracted as a representative node. By performing a three-dimensional regression fitting operation on the node centroid sequence, a continuous spatial curve is constructed to obtain the three-dimensional centerline coordinate sequence of the underground pipeline, thereby achieving high-precision restoration of the underground pipeline's direction and spatial position.

[0165] According to an embodiment of the present invention, an underground pipeline detection system based on a motor vehicle towed vehicle includes the following modules:

[0166] Multi-physics signal acquisition module, used to acquire combined multi-physics signals of vehicle status;

[0167] The signal standardization module is used to normalize the time scale and amplitude of preprocessed multi-source signals to generate standardized multi-physics signal sequences.

[0168] The multiphysics controllable Koopman modeling module is used to receive standardized multiphysics signal sequences and vehicle state information and extract potential vector sequences;

[0169] The linear propagation and control evolution module is used to perform linear propagation of the latent vector sequence, generate predicted latent vectors, and calculate the trigger interval correction.

[0170] The trigger compensation control module is used to adjust the radar frequency, electromagnetic excitation phase, and vibration sampling time reference based on the trigger interval correction amount.

[0171] The multiphysics acquisition and residual analysis module is used to acquire a new frame of signal and calculate the multidimensional residual between the predicted frame and the actual frame.

[0172] The Koopman parameter adaptive update module is used to update the propagation matrix and control matrix based on residual feedback, and adjust the traction speed and chassis height.

[0173] The residual iterative convergence control module is used to control the step iteration. When the residual is lower than the threshold, it outputs an aligned multiphysics signal sequence.

[0174] The underground pipeline identification and modeling module is used to input alignment signals, perform conservation constraint Koopman identification, and output three-dimensional centerline detection results.

[0175] Example 1:

[0176] To verify the feasibility of this invention in practice, it was applied to a typical urban road section. Due to the presence of various old or unknown underground pipelines beneath the road, ground-penetrating radar and low-frequency electromagnetic detection operations are often affected by complex road surface structures, vehicle motion disturbances, and timing inconsistencies between sensors. This results in misalignment of multiphysics detection signals, large errors in reconstructing the centerline, and the inability to effectively fuse the three heterogeneous signals: electromagnetic scattering, induced phase, and surface wave vibration. The vehicle-mounted multiphysics alignment sensing and underground pipeline identification method based on a controllable Koopman network provided by this invention has been fully validated in the aforementioned scenario.

[0177] First, an integrated multi-physics joint sensing device, including ground-penetrating radar, electromagnetic excitation coils, and ground micro-vibration sensors, was installed on the road section. As the vehicle moved forward at a traction speed of about 30 km / h, three types of synchronous signals were collected and sent to the signal standardization module. Through unified time reference and normalization processing, a standardized multi-physics signal sequence was generated. In the Koopman autoencoder network, the frequency domain local features (electromagnetic scattering), phase change features (electromagnetic induction), and low-frequency trend features (surface wave acceleration) were extracted by the three-branch convolutional encoders, and the potential vector sequence was output.

[0178] Subsequently, during the propagation phase, the potential vector is jointly evolved using the linear propagation matrix and the control matrix. The trigger interval correction is calculated by the main peak time difference, the phase change point offset, and the amplitude rise time error. The acquisition module dynamically adjusts the radar pulse frequency (increased by 10% to 20%), the electromagnetic excitation phase (fine-tuned within ±15°), and the vibration sampling time reference (accuracy improved to 2ms) based on this correction, thereby achieving trigger compensation before the acquisition of the next frame signal.

[0179] In the residual calculation stage, the residual tensor is obtained by comparing the predicted signal frame with the actual signal frame. If the residual is higher than the threshold of 0.12, the adaptive update module is triggered to re-evaluate the propagation matrix and control matrix, and a control command is sent to reduce the vehicle speed to 85% of the original value and adjust the chassis height to 75% of the original setting, thereby enhancing the detection stability of soft soil layers and heterogeneous buried areas. All updates are based on the state classification decision made according to the residual distribution of the most recent five frames, ensuring that each frame acquisition meets the alignment standard.

[0180] After 10 rounds of iterative optimization, the system automatically detects that the residuals tend to converge, outputs an aligned multiphysics signal sequence, and inputs it into a conservation constraint Koopman identification unit. In this unit, a state propagation matrix is ​​constructed and energy conservation and phase conservation terms are superimposed to achieve physical consistency correction of the propagation trajectory on the power boundary and phase trend. Finally, the three-dimensional centerline coordinates are fitted by centroid trajectory regression.

[0181] This implementation scheme completed the detection of 1420 meters of pipeline in the actual test section. Compared with the direct reconstruction method of misaligned signals, this method reduces the centerline fitting error by approximately 63% and has a significant advantage in identifying intersecting heterogeneous pipeline structures with burial depths between 1.2 and 2.8 meters. The table lists comparative data from three representative test sections, demonstrating the significant improvement of this invention in multiphysics signal alignment and centerline detection accuracy.

[0182] Table 1: Comparison of Underground Pipeline Identification Results Before and After Multiphysics Alignment

[0183]

[0184] Table 1 above lists the comparison results of the three-dimensional centerline identification of underground pipelines in three typical test sections before and after implementing the method of this invention. From the error data, without multi-physics alignment, the centerline reconstruction errors of each section are relatively high: 38.2 cm for sections A, 42.7 cm for sections B, and 51.3 cm for sections C. This indicates that there is a fitting deviation for different types of underground pipelines under traditional processing methods. After applying the multi-physics controllable Koopman alignment modeling and conservation constraint identification method proposed in this invention, the errors of the three sections were reduced to 14.5 cm, 17.2 cm, and 19.0 cm, respectively, with an error reduction of over 59%, reaching a maximum of 68.1%. This demonstrates that this invention improves the alignment accuracy of multi-source heterogeneous signals and the overall reliability of underground pipeline reconstruction.

[0185] Specific analysis of various pipeline types revealed that the reconstruction accuracy of power pipelines and gas pipelines was significantly improved in areas with greater burial depth or complex structures. This indicates that the proposed method has strong adaptability to deeply buried heterogeneous media, with the average residual reduction consistently exceeding 60%. This verifies the combined effect of the proposed trigger compensation mechanism, residual-driven propagation update strategy, and conservation constraint correction operator, achieving high-precision underground pipeline detection in complex geological environments. This result demonstrates the practical application value of this invention in improving the alignment quality of multiphysics signals and the accuracy of pipeline 3D modeling.

[0186] 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.

Claims

1. A method for detecting underground pipelines based on motor vehicle traction, characterized in that, The method comprises the following steps: Step one, collecting vehicle state multi-physical field joint signals, and performing preprocessing to generate standardized multi-physical field signal sequences; Step two, inputting the standardized multi-physical field signal sequences together with vehicle speed information and traction attitude information into a multi-physical field controllable Kramers self-encoding network, the multi-physical field controllable Kramers self-encoding network being composed of a convolutional encoder, a linear propagation matrix, a control matrix and a decoder, and outputting a latent vector sequence through the convolutional encoder; Step three, performing linear propagation on the latent vector sequence by using the linear propagation matrix and the control matrix inside the multi-physical field controllable Kramers self-encoding network to obtain a predicted latent vector sequence and calculate a trigger interval correction amount; Step four, adjusting the radar pulse trigger frequency and the electromagnetic excitation phase according to the trigger interval correction amount, collecting the next frame of multi-physical field signals, and calculating the residual error between the actual multi-physical field signal frame and the predicted multi-physical field signal frame; Step five, when the residual error is greater than a preset threshold, performing an adaptive Kramers parameter updating process to update the propagation matrix and the control matrix, while reducing the vehicle speed and lowering the deformable chassis height, and generating a corrected latent vector sequence; Step six, repeating steps two to five until the residual error is not greater than the preset threshold, and integrating to obtain aligned multi-physical field signal sequences; Step seven, inputting the aligned multi-physical field signal sequences into a conservation constraint Kramers identification operator to generate an underground pipeline three-dimensional center line model and output an underground pipeline detection result.

2. A method for detecting underground pipelines based on motor vehicle traction according to claim 1, characterized in that, The vehicle state multi-physical field joint signals in step one are composed of vehicle speed data, traction attitude data, geological radar electromagnetic scattering signals, low-frequency electromagnetic induction signals and surface wave vibration acceleration signals, the preprocessing sequentially performs band-pass filtering denoising, time scale unification interpolation, coordinate system conversion synchronization and amplitude normalization operations, and outputs standardized multi-physical field signal sequences, the standardized multi-physical field signal sequences are arranged according to a unified time index, and are composed of a speed value subsequence, an attitude value subsequence, a normalized electromagnetic scattering waveform subsequence, a normalized electromagnetic induction phase subsequence and a normalized surface wave acceleration subsequence.

3. A method for detecting underground pipelines based on motor vehicle traction according to claim 2, characterized in that, The multi-physical field controllable Kramers self-encoding network in step two is sequentially composed of a three-branch convolutional encoder, a latent vector construction unit, a linear propagation matrix, a control matrix and a symmetric structure decoder, and specifically comprises: The three-branch convolutional encoder comprises an electromagnetic scattering encoding branch, an electromagnetic induction encoding branch and a surface wave vibration encoding branch; The electromagnetic scattering encoding branch is used for receiving the normalized electromagnetic scattering waveform subsequence and extracting frequency domain local features, the electromagnetic induction encoding branch is used for receiving the normalized electromagnetic induction phase subsequence and extracting phase change features, and the surface wave vibration encoding branch is used for receiving the normalized surface wave acceleration subsequence and extracting low-frequency vibration time sequence features; The latent vector construction unit is used for concatenating the features output by the three-branch convolutional encoder with the speed value subsequence and the attitude value subsequence according to time steps to generate a latent vector sequence; The linear propagation matrix is used for performing a linear propagation operation on the latent vector sequence, the control matrix is used for converting the speed value subsequence and the attitude value subsequence into control inputs, and the joint propagation matrix is used for linear evolution on the latent vector sequence. The symmetric structure decoder is used for receiving the propagated potential vector sequence and respectively reconstructing the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence and the predicted surface wave acceleration subsequence according to the original three-branch structure, and outputting the predicted multi-physical field signal frame.

4. A method for detecting underground pipelines based on motor vehicle traction according to claim 3, characterized in that, The step three specifically comprises: The predicted potential vector sequence is input into the symmetric structure decoder to respectively reconstruct the predicted electromagnetic scattering waveform subsequence, the predicted electromagnetic induction phase subsequence and the predicted surface wave acceleration subsequence, and form the predicted multi-physical field signal frame; The predicted time interval of adjacent main reflection peaks in the predicted electromagnetic scattering waveform subsequence is calculated by difference with the actual time interval of adjacent main reflection peaks in the actual electromagnetic scattering waveform subsequence, and an electromagnetic scattering trigger error is output; The predicted occurrence time of the phase mutation point in the predicted electromagnetic induction phase subsequence is calculated by time offset with the actual occurrence time of the phase mutation point in the actual electromagnetic induction phase subsequence, and an electromagnetic induction excitation delay is output; The predicted starting time of the main amplitude rising section in the predicted surface wave acceleration subsequence is calculated by difference with the actual starting time of the main amplitude rising section in the actual surface wave acceleration subsequence, and a vibration response offset is output; The electromagnetic scattering trigger error, the electromagnetic induction excitation delay and the vibration response offset are fused by weighting according to the respective signal energy weight, and a trigger interval correction quantity is generated.

5. A method for detecting underground pipelines based on motor vehicle traction according to claim 4, characterized in that, The step four specifically comprises: According to the electromagnetic scattering correction component in the trigger interval correction quantity, the geological radar pulse trigger frequency is adjusted, according to the electromagnetic induction correction component in the trigger interval correction quantity, the low-frequency electromagnetic excitation phase is adjusted, and according to the surface wave vibration correction component in the trigger interval correction quantity, the surface wave vibration sampling time reference is adjusted, and the trigger compensation of the next frame of multi-physical field acquisition is completed; After the trigger compensation is completed, the next frame of vehicle state multi-physical field joint signal is acquired, and the next frame of vehicle state multi-physical field joint signal is preprocessed to generate the next frame of standardized multi-physical field signal sequence; The next frame of standardized multi-physical field signal sequence is input into the multi-physical field controllable Kupmann self-encoding network, and processed by the convolutional encoder, the linear propagation matrix, the control matrix and the symmetric structure decoder to generate the corresponding predicted multi-physical field signal frame; The next frame of standardized multi-physical field signal sequence and the predicted multi-physical field signal frame are aligned point by point according to the unified time index, the Euclidean residual distance between the speed value subsequence, the attitude value subsequence, the normalized electromagnetic scattering waveform subsequence, the normalized electromagnetic induction phase subsequence and the normalized surface wave acceleration subsequence is calculated respectively, and the next frame of whole frame residual is output.

6. A method of underground pipeline detection based on motor vehicle traction according to claim 5, characterized in that, The step five specifically comprises: A potential vector sequence sample set containing the current time and a window length of N is constructed, the potential vector sequence sample set is fitted by using the least mean square error criterion, a propagation error vector is calculated, the propagation matrix is updated by element-by-element gradient descent according to the propagation error vector, and an updated propagation matrix is obtained; The speed value subsequence and the attitude value subsequence in the corresponding time period of the potential vector sequence sample set are extracted, a control input sample set is constructed, a control input error is calculated by using the least control residual criterion, corresponding elements in the control matrix are updated, and an updated control matrix is obtained. The updated propagation matrix and the updated control matrix are input into the multi-physical-field controllable Kubo-Martin-Schwinger auto-encoding network to perform linear propagation operation, and an updated latent vector sequence is output; According to the numerical size of the residual error in different physical field channels, a vehicle control adjustment operation is performed, and a dynamic compensation strategy based on residual component classification is executed; The updated latent vector sequence, the updated speed numerical subsequence, and the updated attitude numerical subsequence are input into the multi-physical-field controllable Kubo-Martin-Schwinger auto-encoding network, and the processes of steps three to five are continuously executed.

7. A method of underground pipeline detection based on motor vehicle traction according to claim 6, characterized in that, The dynamic compensation strategy based on residual component classification specifically includes: If only the electromagnetic scattering residual error exceeds the threshold, a speed adjustment instruction is sent to the motor vehicle control module, the traction speed is set to 80% of the current speed, and the chassis height is kept unchanged; If only the electromagnetic induction residual error exceeds the threshold, a height adjustment instruction is sent to the chassis control module, the deformable chassis height is set to 70% of the current height, and the traction speed is kept unchanged; If only the surface wave vibration residual error exceeds the threshold, a speed adjustment instruction is sent to the motor vehicle control module, the traction speed is set to 85% of the current speed, and an instruction is sent to the time base synchronization module to recalibrate the sampling reference time of the surface wave signal; If the electromagnetic scattering residual error and the electromagnetic induction residual error exceed the threshold at the same time, adjustment instructions are sent to the motor vehicle control module and the chassis control module respectively, the traction speed is set to 75% of the current speed, and the chassis height is set to 70% of the current height; If the electromagnetic scattering residual error and the surface wave vibration residual error exceed the threshold at the same time, the traction speed is set to 75% of the current speed, an instruction is sent to the time base synchronization module to improve the sampling time precision, the sampling time precision is improved from 1 millisecond to 0.1 millisecond, an instruction is sent to the radar control module to improve the radar trigger compensation frequency, and the geologic radar pulse trigger frequency is improved from 100 Hz to 200 Hz; If the electromagnetic induction residual error and the surface wave vibration residual error exceed the threshold at the same time, the chassis height is set to 70% of the current height, and the surface wave sampling time reference calibration and electromagnetic excitation phase compensation operations are performed synchronously; If the electromagnetic scattering residual error, the electromagnetic induction residual error, and the surface wave vibration residual error all exceed the threshold, the traction speed is set to 70% of the current speed, the chassis height is set to 60% of the current height, and the three types of signal acquisition compensation operations are performed synchronously; If the electromagnetic scattering residual error, the electromagnetic induction residual error, and the surface wave vibration residual error all do not exceed the threshold but the continuous three-frame residual error increases by more than 10%, a fine adjustment instruction is sent to the motor vehicle control module to set the traction speed to 95% of the current speed, and the multi-physical-field signal sampling frequency is increased by 5% synchronously.

8. A method of underground pipeline detection based on motor vehicle traction according to claim 7, characterized in that, The generation of the aligned multi-physical-field signal sequence in step six includes: The updated latent vector sequence generated in the last round is input into the multi-physical-field controllable Kubo-Martin-Schwinger auto-encoding network again to reconstruct a predicted multi-physical-field signal frame; Based on the predicted multi-physical-field signal frame and the current actual acquisition frame, the residual error is calculated, the corresponding dynamic compensation strategy based on residual component classification is selected according to the distribution characteristics of the three types of physical components in the residual error, and the updated latent vector sequence is updated; The current round residual is subtracted from the last round residual to obtain a residual change rate, which is compared with a preset convergence rate threshold, and if the residual change rates of three consecutive rounds are less than the convergence rate threshold, or the absolute value of the current round residual is less than a preset residual threshold, it is determined that the convergence is achieved; When the convergence condition is met, the iteration process is terminated, the current modified latent vector sequence is taken as the final latent state representation, and a time axis alignment operation is performed based on the final predicted multi-physical field signal frame and the actually collected frame to construct an aligned multi-physical field signal sequence.

9. A method of underground pipeline detection based on motor vehicle traction according to claim 8, characterized in that, The step seven conservation constraint Kramers identification operator includes a feature encoding unit, a linear propagation unit and a center line fitting unit, and specifically includes: The aligned multi-physical field signal sequence is divided into fixed length time periods, and the multi-physical field spatial feature vector of each time period is extracted by inputting the feature encoding unit; the spatial feature vector is input into the linear propagation unit, the state propagation matrix is established and the energy conservation term and the phase conservation term are superimposed, and the state propagation trajectory is corrected and constrained; The state propagation trajectory after constraint correction is input into the center line fitting unit, and the three-dimensional center line coordinate sequence of the underground pipeline is regressed and fitted based on the node barycenter trajectory, and the underground pipeline detection result is output.

10. A motor vehicle towed underground pipeline detection system, applied to the motor vehicle towed underground pipeline detection method of any one of claims 1 to 9, characterized in that, It includes the following modules: A multi-physical field signal acquisition module for acquiring vehicle state multi-physical field joint signals; A signal standardization module for uniformly time-stamping and amplitude normalizing the preprocessed multi-source signals to generate a standardized multi-physical field signal sequence; A multi-physical field controllable Kramers modeling module for receiving the standardized multi-physical field signal sequence and vehicle state information and extracting a latent vector sequence; A linear propagation and control evolution module for performing linear propagation of the latent vector sequence, generating a predicted latent vector and calculating a trigger interval correction amount; A trigger compensation control module for adjusting the radar frequency, electromagnetic excitation phase and vibration sampling time reference according to the trigger interval correction amount; A multi-physical field acquisition and residual analysis module for performing new frame signal acquisition and calculating multi-dimensional residuals between predicted frames and actual frames; A Kramers parameter adaptive updating module for updating the propagation matrix and control matrix according to the residual feedback and adjusting the traction speed and chassis height; A residual iterative convergence control module for controlling the step iteration, and when the residual is lower than the threshold, outputting the aligned multi-physical field signal sequence; An underground pipeline identification and modeling module for inputting the aligned signal, performing conservation constraint Kramers identification and outputting a three-dimensional center line detection result.

Images