Intelligent monitoring method for pipe blockage during concrete pumping
By setting up an array of acoustic patches and an ultrasonic generator on the concrete pump pipe, the acoustic signals are collected and processed. A lightweight 1D-CNN model is used to identify the pump pipe status, achieving second-level pipe blockage identification and self-repair. This solves the problems of inaccurate pipe blockage location and delayed response in traditional methods, and improves construction efficiency and safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies struggle to accurately pinpoint blockage locations during concrete pumping, and traditional methods suffer from response lag, high costs, and an inability to achieve automatic location, detection, early warning, and intervention loops for blockages, thus impacting construction continuity.
An array of acoustic patches is used to collect acoustic signature signals. After wavelet transform and normalization, a four-dimensional feature vector F is calculated and input into a lightweight 1D-CNN model for state recognition. An ultrasonic generator is set on the outside of the easily clogged area of the pump pipe for self-repair.
It achieves second-level pipe blockage identification and processing, improves the accuracy of pipe blockage identification, supports multi-point blockage identification, reduces equipment and installation costs, reduces material waste, and is adaptable to monitoring in harsh environments.
Smart Images
Figure CN121828172A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a concrete pumping pipe blockage intelligent monitoring method, belonging to the technical field of concrete construction monitoring. BACKGROUND
[0002] During the concrete pumping process, pipe blockage may occur, therefore, concrete pumping state monitoring is an important link to ensure construction quality and efficiency in construction engineering.
[0003] Traditional detection methods include manual knocking method, pressure monitoring method and ultrasonic guided wave monitoring method, which have the following problems: the manual knocking method relies on workers to hold a hammer to knock the pipe and judge the blockage position through sound, which has strong subjectivity, low efficiency and difficulty in accessing high-risk areas; the pressure monitoring method judges pipe blockage by sudden change of pumping pressure, but has response lag (pressure anomaly usually lags behind the actual pipe blockage event by 5-10 seconds), and cannot accurately locate the blockage position; the ultrasonic guided wave monitoring method uses a piezoelectric sensor array to emit ultrasonic waves and receive signals, but requires complex excitation equipment and is expensive, making it difficult to deploy on a large scale.
[0004] Now there is a method of using voiceprint for pipe blockage detection, but there are still the following problems: lack of multi-dimensional feature extraction and dynamic modeling capability of voiceprint signals, unable to distinguish between normal pumping noise and early pipe blockage characteristics; automatic positioning of pipe blockage position has not been achieved, manual segment-by-segment investigation is still required, which delays the disposal time; a "detection-warning-intervention" closed loop has not been formed, and manual shutdown is required after pipe blockage, which affects the continuity of construction. SUMMARY
[0005] The present application provides a concrete pumping pipe blockage intelligent early warning and self-repairing method, which processes the collected voiceprint signals, obtains the key indicators of the voiceprint signals, and then uses a lightweight model to automatically identify the running state of the pump pipe, effectively improving the identification accuracy.
[0006] To solve the above technical problems, the present application includes the following technical solutions:
[0007] A concrete pumping pipe blockage intelligent monitoring method, comprising the following steps:
[0008] Step one, array acoustic patches are attached to the easy-to-block area of the pump pipe, and during the concrete pumping process, the original voiceprint signals are collected through the acoustic patches;
[0009] Step two, the collected original voiceprint signals are preprocessed, including denoising the original voiceprint signals to obtain pure voiceprints, and then normalizing the pure voiceprints;
[0010] Step three, calculate the four-dimensional feature vector F of the preprocessed pure voiceprint signals, wherein,
[0011] F = (H, CF, proportion, ratio),
[0012] H, CF, proportion, ratio are four key indicators, respectively described by numerical range pump pipe blockage state, H is energy entropy, CF is peak factor, these two are time domain indicators; Proportion refers to the proportion of FFT main frequency energy, ratio refers to the proportion of high frequency energy, these two are frequency domain indicators;
[0013] Step four, input F into edge end 1D-CNN lightweight model, the model outputs the probability P of four states state , and the state with the highest probability is taken as the pump pipe running state; Wherein,
[0014] P state = (P normal, P half block, P full block, P empty hit),
[0015] In the formula, P 正常 , P 半堵 , P 全堵 , P 空打 are the probabilities of pump pipe normal, half block, full block and empty hit respectively.
[0016] Further, in step one, multiple acoustic patches are arranged at the same position to form an acoustic array.
[0017] Further, in step two, the original voiceprint signal is denoised to obtain a pure voiceprint, specifically:
[0018] The sound wave is decomposed into different frequency subbands by wavelet transform;
[0019] Set a volume threshold for each layer, and below the threshold is considered as noise and directly zeroed.
[0020] The signal is reconstructed to obtain a pure voiceprint signal.
[0021] Further, in step two, the pure voiceprint is normalized, specifically:
[0022] The pure voiceprint signal amplitude is compressed to the [0, 1] interval;
[0023] The normalized value is calculated, wherein the normalized value = (current value-minimum value) / (maximum value-minimum value).
[0024] Further,
[0025] In the formula, L is the number of signal decomposition layers or subbands, p i is the normalized energy of the i-th subband, and the normalized energy where E i is the energy of the i-th subband.
[0026] Further,
[0027] wherein the peak P max is the maximum absolute value of the sound signal waveform, the effective value V rms is the root mean square value of the sound signal waveform.
[0028] Further,
[0029] wherein the main frequency energy E main is the sum of the energy of the highest energy frequency component in the amplitude spectrum and its adjacent frequency bands, and the total energy ∑E total is the sum of the energy of all frequency components in the entire frequency band of interest.
[0030] Further,
[0031] wherein E HF is the sum of the energy of the high-frequency subbands obtained by wavelet packet decomposition, E total is the total energy of all subbands.
[0032] Further, the 1D-CNN lightweight model is pre-trained, and the training includes:
[0033] Collecting original voiceprint data including four categories of normal, semi-blocked, fully blocked, and empty shots, and label data corresponding to the original voiceprint data, which are labeled as four categories of normal flow, semi-blocked, fully blocked, and empty shots;
[0034] Pretreating the collected original voiceprint signal, including denoising the original voiceprint signal to obtain a pure voiceprint, and then normalizing the pure voiceprint;
[0035] Calculating a four-dimensional feature vector F of the pretreated voiceprint signal;
[0036] Forming a training data set including a plurality of sets of training data, each set of training data including F and corresponding label data;
[0037] Training the 1D-CNN lightweight model using the training data set.
[0038] Further, an ultrasonic generator is arranged outside the easy-to-block area of the pump pipe;
[0039] If the highest probability is P 半堵 , the voiceprint propagation time difference is used to locate the blockage coordinates, and the corresponding ultrasonic generator is turned on to emit a sound wave higher than 20 KHz to vibrate the concrete in the pipe and change its rheological property.
[0040] Compared with the prior art, the present application has the following advantages and positive effects: by arranging an acoustic patch outside the easy-to-block area of the pump pipe, the original acoustic signal of the pump pipe is collected, a pure acoustic signal is obtained after processing, and a four-dimensional feature vector F of the pure acoustic signal is calculated, which is then input into a pre-trained lightweight model to directly obtain the probabilities P of four types of states of the pump pipe operation state and the determined pump pipe operation state, and an alarm information is issued when the pump pipe operation state is abnormal. The present application greatly improves the blockage identification accuracy and the blockage determination rate, supports simultaneous identification of multiple blockages, and solves the limitation of traditional methods that can only handle single-point blockages. In addition, the acoustic array detection of the acoustic patch can eliminate the interference of adverse weather environment factors and can monitor areas that are inaccessible to personnel in high-risk environments. Using the method provided by the present application, the time from judgment to processing of a single blockage can reach seconds, reducing the waste of concrete materials, and the MEMS sensor array is plug and play, greatly reducing the cost of the device itself and installation compared to traditional monitoring systems. BRIEF DESCRIPTION OF DRAWINGS
[0041] Figure 1 A flowchart of a concrete pumping blockage intelligent monitoring method according to an embodiment of the present application. DETAILED DESCRIPTION
[0042] The concrete pumping blockage intelligent monitoring method provided by the present application will be described in further detail below in combination with the drawings and specific embodiments. The advantages and features of the present application will be clearer in combination with the following description. It should be noted that the drawings are greatly simplified and use non-precise proportions, only to facilitate and clarify the purpose of assisting in the description of the embodiments of the present application.
[0043] As Figure 1 shown, the concrete pumping blockage intelligent early warning and self-repairing method provided by the present embodiment includes the following steps:
[0044] Step one, an ultrasonic generator is arranged outside the easy-to-block area of the pump pipe, and an array acoustic patch is attached. During the concrete pumping process, the original acoustic signal is collected through the acoustic patch.
[0045] The acoustic patch can be a MEMS acoustic patch with a sampling frequency of ≥20 kHz and a communication protocol of LoRa / NB-IoT, forming a 3x3 MEMS acoustic array.
[0046] The ultrasonic generator is installed outside the easy-to-block area of the pump pipe, such as the elbow, and three are installed at the same position, with a signal transmission line and a connection to the municipal power grid.
[0047] Step two, preprocessing the collected original voiceprint signal, including denoising (wavelet threshold denoising) the original voiceprint signal to obtain a pure voiceprint, and then normalizing the pure voiceprint.
[0048] After denoising the original voiceprint signal, a pure voiceprint signal can be obtained, specifically: decomposing the sound wave into different frequency subbands by wavelet transform; setting a volume threshold for each layer, and regarding the noise below the threshold as noise and directly setting it to zero; reconstructing the signal to obtain the pure voiceprint signal.
[0049] The normalization process specifically includes: compressing the pure voiceprint signal amplitude to the [0, 1] interval, thereby avoiding that the weak jamming signal at a long distance is drowned out by the normal noise at a short distance; calculating a normalization value for eliminating the interference of the difference in pumping distance, wherein the normalization value = (current value-minimum value) / (maximum value-minimum value).
[0050] Step three, calculating the four-dimensional feature vector F of the preprocessed voiceprint signal, wherein,
[0051] F=(H, CF, proportion, ratio).
[0052] H, CF, proportion, and ratio are four key indicators, respectively described by a numerical range to describe the pump pipe jamming state. H is energy entropy, and CF is peak factor, both of which are time domain indicators. Proportion refers to the FFT main frequency energy proportion, and ratio refers to the high frequency energy proportion, both of which are frequency domain indicators.
[0053] (1) Energy entropy H.
[0054] Energy entropy H represents the degree of disorder or uncertainty of the voiceprint signal. The higher the entropy, the more chaotic and irregular the signal (usually indicating abnormal mixed noise).
[0055]
[0056] In the formula, L is the number of signal decomposition layers or subbands, p i is the normalized energy of the i-th subband. The normalized energy where E i is the energy of the i-th subband.
[0057] Normal flow: low entropy (flat, energy concentrated in low frequency), numerical range about 0.5-1.5.
[0058] Half jamming: high entropy (harsh, scattered spectrum distribution), numerical range about 2.0-3.5.
[0059] Full jamming: signal may disappear or be extremely single, and the entropy value is low.
[0060] (2) Peak factor CF.
[0061] The peak factor CF represents the impact strength of the signal. High peaks of the burst are usually associated with impacts or intense friction, indicating an increased risk of blockage.
[0062]
[0063] where the peak P max is the maximum absolute value of the sound signal waveform, the effective value V rms (Root Mean Square) is the root mean square value of the sound signal waveform.
[0064] Normal flow: Entropy is low (smooth), peak factor value range is about 2.5-3.5.
[0065] Half blockage: Peak factor > 5.5 (signal spike high).
[0066] Full blockage: Peak factor value range is 3.0-4.5
[0067] (3) FFT main frequency energy proportion.
[0068] The FFT main frequency energy proportion represents the concentration of the dominant tone. Normal pumping sound is usually concentrated in the low frequency band, and abnormal friction sound will cause high frequency component energy concentration. FFT is a fast algorithm for computing discrete Fourier transform (DFT) and its inverse transform, which can convert audio signals from time domain to frequency domain. After FFT processing, it can be represented as a combination of different frequency components in the frequency domain.
[0069]
[0070] where the main frequency energy E main is the sum of the energy of the highest energy frequency component and its adjacent frequency band in the amplitude spectrum. The total energy ∑E total is the sum of the energy of all frequency components in the entire frequency band of interest.
[0071] Normal flow: Energy is concentrated in the 200-500Hz frequency band, and the main frequency energy proportion is high (> 0.7).
[0072] Half blockage: The proportion of low frequency main frequency energy decreases, the proportion of main frequency energy is 0.4-0.6, and the energy is dispersed to high frequency.
[0073] Full blockage: The proportion of main frequency energy is <0.3.
[0074] (4) High frequency band energy proportion.
[0075] The physical meaning of the high frequency band energy proportion is the intensity of high frequency friction sound. When there are stones stuck or friction in the pipeline, high frequency spikes will be generated, and this proportion will rise.
[0076] wherein E HF is the total energy of the high frequency sub-band (e.g. above 15 kHz) obtained by wavelet packet decomposition. total is the total energy of all sub-bands.
[0077] Normal flow: low ratio of high frequency energy (e.g. less than 0.1).
[0078] Half-blockage: a spike at 15 kHz occurs, and the ratio of high frequency energy increases significantly (>0.25).
[0079] Full-blockage: low ratio of high frequency energy (e.g. less than 0.1) accompanied by an abnormal peak factor CF. The value of the peak factor CF in full-blockage is higher than that in normal flow, but lower than that in half-blockage.
[0080] Step four, input F into the edge terminal 1D-CNN lightweight model, the model outputs the probability P of four states state , and the state with the highest probability is taken as the pump pipe running state. If the output pump pipe running state is abnormal, a warning information is issued. Wherein,
[0081] P state = (P 正常 , P 半堵 , P 全堵 , P 空打 ).
[0082] P 正常 , P 半堵 , P 全堵 , P 空打 are the probabilities in the normal, half-blockage, full-blockage, and empty-pumping states, respectively.
[0083] Regardless of the type of abnormality (half-blockage, full-blockage, or empty-pumping), the beamforming positioning is started, and the blockage coordinates (error ≤0.3 meters) are pushed to the mobile terminal. The BIM system can also be linked to mark the risk area.
[0084] If the highest probability is P 半堵 , it is identified as "half-blockage", and the ultrasonic generator or reverse pulse self-repair mechanism is triggered. Specifically, the blockage coordinates are located based on the soundprint propagation time difference, and are pushed to the mobile phone or tablet of the site manager to assist in accurate disposal; the ultrasonic generator can be turned on to emit sound waves higher than 20 KHz to vibrate the concrete in the pipe and change its rheological property.
[0085] In addition, P 半堵 and P 全堵 output by the model can also be setThe probability calculation block pipe risk score (0-100 points) is triggered if the score exceeds the set threshold. For example, if the score is greater than or equal to 85, an early warning is triggered. The edge computing unit can be used to calculate the block pipe risk score, and the edge computing unit integrates an AI chip of the 1D-CNN model (such as Horizon Sun X3 and Huawei Ascend 310), to realize local real-time analysis with a delay of less than or equal to 50 ms.
[0086] It should be noted that the 1D-CNN (1D Convolutional Neural Network) is a one-dimensional convolutional neural network, which is a variant of the convolutional neural network, and is specifically used to process sequence data such as time series and text. The lightweight model used is pre-trained, and the training process is as follows:
[0087] The original voiceprint data including normal, half-block, full-block, and empty shots, and the corresponding labeled data of the original voiceprint data, i.e., normal flow, half-block, full-block, and empty shot, are collected, each type of data is greater than or equal to 200, and the time length is 5 seconds per piece.
[0088] After denoising and normalization, the signal-to-noise ratio is improved, and the interference of the pumping distance difference is eliminated.
[0089] Time domain and feature extraction are performed to obtain a four-dimensional feature vector F, and the F and the corresponding label data constitute a training set.
[0090] The training process is as follows:
[0091] (1) Forward propagation: the voiceprint data is calculated from the input layer to the output layer, and the predicted probability is generated (for example, [0.1, 0.8, 0.05, 0.05] represents the probability of “half-block” being 80%);
[0092] (2) Loss calculation: the cross-entropy loss is used to compare the predicted probability with the true label;
[0093] (3) Back propagation: the error is transmitted from the output layer to the input layer, and the convolution kernel weight is updated by gradient descent;
[0094] (4) Output result: four types of state probability (normal / half-block / full-block / empty shot), and the highest probability is taken as the final state.
[0095] Transfer learning optimizes the generalization ability. The training data are all short-distance pumping voiceprints, but the characteristic shift is caused by the 100-meter or more Shen wave attenuation. The short-distance training base model is used to learn the common voiceprint rules, a small amount of long-distance data is added, the first two convolution kernels are frozen (to retain common features), and only the parameters of the last fully connected layer are adjusted.
[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0097] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.
Claims
1. A method for intelligent monitoring of concrete pump pipe blockage, characterized in that, Includes the following steps: Step 1: Install an array of acoustic patches in the easily clogged areas of the pump pipe. During the concrete pumping process, the original acoustic signals are collected through the acoustic patches. Step 2: Preprocess the collected raw voiceprint signal, including denoising the raw voiceprint signal to obtain a clean voiceprint, and then normalizing the clean voiceprint. Step 3: Calculate the four-dimensional feature vector F of the preprocessed clean voiceprint signal, where, F = (H, CF, percentage, proportion) H, CF, percentage, and proportion are four key indicators, each describing the pump pipe blockage state with a numerical range. H is the energy entropy and CF is the peak factor, both of which are time-domain indicators. Percentage refers to the energy percentage of the FFT main frequency, and proportion refers to the energy percentage of the high-frequency band, both of which are frequency-domain indicators. Step 4: Input F into the edge-end 1D-CNN lightweight model. The model outputs the probabilities P of the four states. state The state with the highest probability is taken as the operating state of the pump pipe; among which, P state =(P normal, P partially blocked, P fully blocked, P empty) In the formula, P 正常 ,P 半堵 ,P 全堵 ,P 空打 The probabilities are given for four states: normal pump pipe, partially blocked, fully blocked, and running dry.
2. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, In step one, multiple acoustic patches are placed at the same location to form an acoustic array.
3. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, Step two involves denoising the original voiceprint signal to obtain a clean voiceprint, specifically as follows: Wavelet transform is used to decompose sound waves into different frequency subbands; Set a volume threshold for each layer; anything below the threshold is considered noise and is immediately reduced to zero. The reconstructed signal yields a clean voiceprint signal.
4. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, Step two involves normalizing the pure voiceprint, specifically as follows: Compress the amplitude of the pure voiceprint signal to the [0,1] interval; Calculate the normalized value, where the normalized value = (current value - minimum value) / (maximum value - minimum value).
5. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, In the formula, L is the number of layers or sub-bands of the signal decomposition, and p i It is the normalized energy of the i-th subband. Where E i It is the energy of the i-th subband.
6. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, Among them, peak value P max It is the maximum absolute value and effective value V of the sound signal waveform. rms It is the root mean square value of the sound signal waveform.
7. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, Among them, the main frequency energy E main It is the sum of the energies of the highest-energy frequency component in the amplitude spectrum and its adjacent frequency bands, the total energy ∑E total It is the sum of the energy of all frequency components within the entire frequency band of interest.
8. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, Among them, E HF E is the sum of the energy of the high-frequency subbands obtained through wavelet packet decomposition. total It is the total energy of all subbands.
9. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, The lightweight 1D-CNN model is pre-trained, and the training includes: The raw voiceprint data includes four categories: normal, partially blocked, fully blocked, and dry-running, as well as the corresponding label data of the raw voiceprint data labeled as normal flow, partially blocked, fully blocked, and dry-running. The collected raw voiceprint signal is preprocessed, including denoising the raw voiceprint signal to obtain a clean voiceprint, and then normalizing the clean voiceprint. Calculate the four-dimensional feature vector F of the preprocessed voiceprint signal; A training dataset is formed, which includes several sets of training data, each set of training data including F and the corresponding label data; The lightweight 1D-CNN model was trained using the training dataset.
10. The intelligent monitoring method for concrete pumping pipe blockage as described in claim 1, characterized in that, An ultrasonic generator is installed on the outside of the easily clogged area of the pump pipe. If the highest probability is P 半堵 Based on the time difference of acoustic signature propagation, the blockage coordinates are located, and the corresponding ultrasonic generator is activated to emit sound waves higher than 20 kHz to vibrate the concrete inside the pipe, thereby changing its rheological properties.