Method and system for monitoring concentration in a deep sea mining hydrotransport pipe

Abstract

The application provides a deep-sea mining hydraulic lifting pipeline concentration monitoring method and system, and belongs to the field of deep-sea mineral resource development and detection. The application adopts indoor model experiment data and a small amount of field measurement data to construct a transfer learning neural network model, fuses a multi-scale convolution network, a multi-head attention mechanism and a transfer learning algorithm to form a full closed-loop system from indoor laboratory pre-training to deep-sea online precise concentration measurement, so that the purpose of significantly improving real-time monitoring accuracy is achieved. The deep-sea mining hydraulic lifting pipeline concentration monitoring system comprises a source domain data construction and model pre-training subsystem for constructing a land source domain data set and generating pre-training model parameters, and a target domain data acquisition and online transfer monitoring subsystem for acquiring deep-sea target domain data and loading pre-training parameters to perform online monitoring.

Description

Technical Field

[0001] This application belongs to the field of deep-sea mineral resource development and testing, and specifically proposes a system and method for real-time intelligent monitoring of concentrations in deep-sea hydraulic lifting pipelines. Background Technology

[0002] Existing polymetallic nodule mining systems used in deep-sea areas generally employ vertical hydraulic lifting pipelines, some thousands of meters long, to transport ore collected from the seabed to mining vessels on the surface. During this transportation process, real-time and accurate monitoring of the ore volume concentration within the pipeline is crucial for ensuring the safety of the lifting pump units, preventing pipeline blockages, and optimizing lifting efficiency.

[0003] Currently, deep-sea mining hydraulic lifting pipelines operate in high-pressure environments at depths of several thousand meters. Furthermore, the slurry transported within these pipelines exhibits significant physical characteristics, including a wide particle size distribution, high concentration, discontinuous solid phase, and opacity. Existing concentration detection methods face severe challenges under these unique conditions. For instance, optical methods require the design of transparent windows that can withstand the high pressure of the deep sea. Due to the extremely high concentration and opacity of the slurry, light cannot penetrate the interior, allowing detection of only a very small surface area and creating a large blind zone. Ultrasonic detection methods, on the other hand, suffer from significant challenges due to the wide particle size range of deep-sea ore. Large-diameter particles exhibit strong scattering and attenuation effects on ultrasonic waves, resulting in multiple scatterings and energy loss during signal propagation, leading to distorted echo signals and extremely low signal-to-noise ratios.

[0004] Therefore, existing technologies still face many unresolved technical challenges in achieving accurate online monitoring of hydraulically lifted slurry concentration under deep-sea conditions. In view of this, this application is hereby submitted. Summary of the Invention

[0005] The concentration monitoring method and system for deep-sea mining hydraulic lifting pipelines described in this application aim to solve the problems existing in the prior art by constructing a transfer learning neural network model using indoor model experimental data and a small amount of field measurement data. This model integrates multi-scale convolutional networks, multi-head attention mechanisms, and transfer learning algorithms to form a closed-loop system from indoor laboratory pre-training to precise online concentration measurement in the deep sea, thereby significantly improving the accuracy of real-time monitoring.

[0006] To achieve the above design objectives, the deep-sea mining hydraulic lifting pipeline concentration monitoring system includes a source domain data construction and model pre-training subsystem for constructing a terrestrial source domain dataset and generating pre-trained model parameters, and a target domain data acquisition and online migration monitoring subsystem for collecting deep-sea target domain data and loading pre-trained parameters to perform online monitoring.

[0007] The source domain data construction and model pre-training subsystem includes a stirrer, a pump connected to the outlet of the stirrer for pumping the prepared slurry into the model bend to construct a circulation loop, a first accelerometer installed on the outer wall of the model bend, and a test flow meter installed inside the model bend; both the first accelerometer and the test flow meter are connected to a first data acquisition instrument via a first wire, and the acquired acceleration and flow velocity data are transmitted to a first computer; the first computer is configured to construct a multi-input fusion network model and perform supervised pre-training using the source domain data, and save the converged model parameters;

[0008] The target domain data acquisition and online migration monitoring subsystem includes a relay compartment containing slurry. A relay compartment valve for controlling the slurry's entry into the actual operating environment lifting pipe is connected to the outlet end of the relay compartment. A second accelerometer is installed on the outer wall of the bend of the actual operating environment lifting pipe. A flow meter is installed in the actual operating environment lifting pipe. Both the second accelerometer and the flow meter are connected to a second data acquisition unit via a second wire to acquire real-time operating data and transmit it to a second computer. The second computer is configured to receive pre-trained model parameters generated by a first computer and, in conjunction with the real-time acquired data, perform multi-core maximum mean difference algorithm transfer learning and concentration calculation to output the final slurry concentration monitoring value.

[0009] Furthermore, both the first and second acceleration sensors are installed on the outer arch side of the bend, and both are high-frequency response piezoelectric acceleration sensors.

[0010] Based on the aforementioned concentration monitoring system for deep-sea mining hydraulic lift pipelines, this application also proposes the following method for monitoring concentration in deep-sea mining hydraulic lift pipelines, which includes the following steps:

[0011] S1. Construct the training dataset;

[0012] Acceleration and velocity data of hydraulic lifting pipelines in multiple indoor model experiments were collected as source domain datasets, and acceleration and velocity data of hydraulic lifting pipelines in multiple deep-sea actual operation environments were collected as target domain datasets. The source domain datasets and target domain datasets were used as training datasets.

[0013] S2. Construct a multi-input fusion neural network model;

[0014] The multi-input fusion neural network model includes an acceleration feature extraction branch and a flow velocity feature extraction branch, which fuse the features of the two and output the concentration prediction value through a fully connected layer;

[0015] S3, Pre-training;

[0016] The multi-input fusion neural network model is pre-trained using source domain datasets in a supervised learning manner;

[0017] S4. Introduce transfer learning for retraining and updating;

[0018] Using source domain datasets and target domain datasets, a multi-kernel maximum mean difference transfer learning method is used to retrain and update the multi-input fusion neural network model;

[0019] S5. Real-time measurement;

[0020] The real-time acceleration and velocity data of the deep-sea hydraulic lifting pipeline to be measured are input into the optimized multi-input fusion neural network model, and the real-time measurement results of the slurry volume concentration are output.

[0021] The real-time acceleration and velocity data of the deep-sea hydraulic lifting pipeline to be measured are input into the multi-input fusion neural network model optimized in step S4, and the flow inference and physical post-processing are performed to output the real-time measurement results of the slurry volume concentration.

[0022] Step S1 includes:

[0023] S11. Construct the source domain dataset;

[0024] In the indoor model experiment, the source domain data construction and model pre-training subsystem was used to collect acceleration data and corresponding fluid velocity data of the mining hydraulic lifting pipeline;

[0025] S12. Construct the target domain dataset;

[0026] In the actual deep-sea mining operation environment, the target domain data acquisition and online migration monitoring subsystem is used to collect acceleration data and corresponding fluid velocity data of the mining hydraulic lifting pipeline, and construct the target domain dataset.

[0027] S13. Preprocess the source and target domain data;

[0028] One-dimensional time series data points of a set time window length are collected and extracted using the first and second accelerometers.

[0029] Step S2 includes:

[0030] S21. Construct a deep neural network;

[0031] The deep neural network contains two input branches. The logic for dividing the two input branches is based on the differences in frequency response characteristics and energy contribution mechanisms under different concentrations in the deep-sea booster pipe. The first input branch processes the acceleration time-series signal, while the second input branch processes the relatively low-frequency, gently changing flow velocity signal.

[0032] S22. Construct the acceleration feature extraction branch;

[0033] The acceleration feature extraction branch sequentially includes a one-dimensional multi-scale convolutional neural network, a multi-head self-attention mechanism, and a bidirectional long short-term memory network.

[0034] S23. Construct the flow velocity feature extraction branch;

[0035] A global statistical pooling layer is used to extract the steady-state and trend features of the velocity sequence and map them into a high-dimensional velocity feature vector. The velocity feature extraction branch uses the global statistical pooling layer to extract the steady-state features of the velocity sequence, and performs a dimensionality upscaling operation on the velocity data, which can effectively establish an energy reference system for the acceleration signal.

[0036] S24, Feature Fusion and Regression Layer;

[0037] The first fully connected layer has 256 nodes and is used to receive the acceleration and velocity signals output from the front end. The second fully connected layer has 64 nodes. Through nonlinear transformation, it performs secondary screening and combination of the 256-dimensional redundant features, extracts the high-order features that affect the particle concentration value in the hydraulic lift pipeline, filters out irrelevant variable information, and finally connects to the output layer with 1 node to output the particle concentration value in the hydraulic lift pipeline.

[0038] Step S22 includes setting the following parameters:

[0039] One-dimensional multi-scale convolutional neural network scale 1st branch: The first convolutional layer has 16 kernels and a kernel size of 3×1. This branch is used to capture high-frequency collision features.

[0040] The second branch of the one-dimensional multi-scale convolutional neural network: the first convolutional layer with 16 kernels and a kernel size of 5×1. This branch is used to capture mid-frequency friction features.

[0041] The third branch of the one-dimensional multi-scale convolutional neural network: the first convolutional layer with 16 kernels and a kernel size of 7×1. This branch focuses on capturing low-frequency macroscopic features.

[0042] Multi-head self-attention mechanism layer: This layer is connected after the multi-scale feature extraction layer, with a kernel count of 128 and a Dropout ratio of 0.5;

[0043] Bidirectional Long Short-Term Memory Network Layer: Contains a two-layer stacked structure for extracting temporal dependencies; the first layer of BiLSTM has 128 hidden nodes; the second layer of BiLSTM has 256 hidden nodes.

[0044] Step S23 includes:

[0045] S231. The velocity data and acceleration data maintain a consistent 2048-window length to ensure the extraction of fluid energy background information within the exact same time period; simultaneously, the velocity sequence of length 2048... Global average pooling is performed to obtain steady-state flow velocity characteristics. ;

[0046] S232, will The features are input to a fully connected layer, which maps them to a high-dimensional flow velocity feature vector. ;

[0047] S233. The temporal acceleration feature vector output by the bidirectional long short-term memory network in step S22. The high-dimensional flow velocity feature vector output in step S232 The features are concatenated to obtain a multi-source fused feature vector. .

[0048] Step S3 includes:

[0049] S31. First, the source domain training set is input into the model, and forward propagation calculation is performed. The acceleration signal branch extracts the frequency domain characteristics of particle collision, particle friction and water flow background noise, and the flow velocity branch calculates the background energy gain coefficient. Then, the regression layer outputs the predicted value of ore concentration of deep-sea mining hydraulic lifting pipeline in the source domain by combining the information of both.

[0050] S32. Construct the regression loss function;

[0051] The mean squared error loss function is used, and the total loss function is constructed by combining the sum of squares of all weight parameters in the multi-input fusion neural network model. This is used to calculate the total error between the predicted value output by the model and the real physical label.

[0052] S33. After calculating the loss value of the current training batch in step S32, backpropagation is performed based on the loss value to calculate the gradient of the parameters of each layer of the network, and the Adam optimizer is used to iteratively update the weights and bias parameters in the model to minimize the prediction error.

[0053] S34. Determine whether the training rounds have reached the preset maximum number of iterations or whether the loss function has converged; if so, stop training and save the model parameters trained in the current source domain.

[0054] Step S4 includes:

[0055] S41. Introduce the multi-core maximum mean difference algorithm;

[0056] The distribution distance between source domain data and target domain data in the model fusion feature layer is measured, and the source domain data and target domain data are mapped to the regenerating kernel Hilbert space;

[0057] S42. Perform transfer layer discrimination. For each network layer in the model, determine whether parameter transfer should be performed. If it is determined that the layer should be transferred, proceed to step S43 below. If it is determined that the layer should not be transferred, proceed to step S44 below.

[0058] S43. Load the source domain parameters saved in step S3 as initial values;

[0059] S44. Reinitialize the parameters of the non-transfer layer;

[0060] S45. Construct the total loss function;

[0061] The transfer learning model is obtained by training with target domain data and minimizing the total loss function through backpropagation algorithm.

[0062] By constructing a total loss function, the neural network can extract domain-invariant physical features;

[0063] Calculate source domain distribution and target domain distribution The formula for calculating the maximum mean difference of multiple kernels in the regenerating kernel Hilbert space is as follows:

[0064] ;

[0065] in, Represents source domain distribution and target domain distribution The square of the maximum mean difference between the multi-core samples; Indicates that it follows the source domain distribution Source domain samples; Indicates that it follows the target domain distribution Target domain samples; Represents the feature mapping function; Indicates the distribution of the source domain Find the expected value; Indicates the distribution of the target domain Find the expected value; In the regenerating nucleus Hilbert space The square of the norm in the vector is used to measure the distance between two feature mean vectors;

[0066] With feature mapping function The feature kernel is defined as the weighted sum of three independent Gaussian kernels to represent the scale range of different frequency bands. The formula for calculating the weighted sum is as follows:

[0067] ;

[0068] in, This is a combined kernel function used to measure the difference in distributions; The squared Euclidean distance between the two samples in the feature space; For the first One kernel function; For the first The bandwidth parameters of each Gaussian kernel correspond to frequency bands of different scales; For the first The weight coefficients of each core, and satisfying ;

[0069] The total loss function is then defined as follows:

[0070] ;

[0071] in, This is the balance coefficient; This represents the total loss function value during model training; : Indicates the loss of source domain data; This represents the multi-core maximum mean difference loss, used to measure the distribution difference between the source and target domains in the feature space.

[0072] Step S5 includes:

[0073] First, a sliding window-based real-time streaming inference is performed; the transfer learning model deployed on the mining vessel monitoring host is configured in online inference mode, and the system continuously receives acceleration data streams and velocity data streams from the target domain data acquisition and online migration monitoring subsystem; according to step S13, the source domain and target domain data are preprocessed, and the data streams are truncated using a sliding window mechanism;

[0074] Secondly, perform post-processing algorithms such as exponentially weighted moving average or Kalman filtering;

[0075] The original concentration prediction value is obtained at the output of the multi-input fusion neural network model, and the original concentration prediction value is smoothed by exponential weighted moving average or Kalman filter algorithm.

[0076] Finally, multi-dimensional visualization and industrial closed-loop control are implemented; the final concentration data and early warning status after calculation are encapsulated into a standard industrial data package and transmitted in real time to the mining vessel's central control system via Modbus TCP or MQTT communication protocol; the mining vessel's central control system makes intelligent judgments based on the calculated concentration change rate.

[0077] In summary, this application has the following advantages and beneficial effects compared with the prior art:

[0078] 1. Existing technologies suffer from poor generalization ability in real-world operations due to the significant differences between the physical characteristics of deep-sea measured data and laboratory data. To address this, this application introduces a multi-core maximum mean difference transfer learning algorithm. This algorithm utilizes a large amount of low-cost laboratory source domain data for pre-training and combines it with a small amount of deep-sea target domain data for distribution alignment and parameter updates. This enables high-precision measurement results even in the absence of a large amount of field-labeled data.

[0079] 2. This application completes the entire process from synchronous data acquisition and preprocessing to feature fusion and result output in a short time. It can reflect the instantaneous fluctuation of slurry concentration in the pipe in a timely manner. The system can respond quickly and effectively avoid serious engineering accidents caused by delayed response when there are signs of concentration overload or pipeline blockage.

[0080] 3. Existing technologies suffer from detection blind spots, the need for pressure-resistant viewing windows, and susceptibility to scattering from large particles. This application employs a non-invasive acceleration monitoring scheme, requiring only the sensor to be installed on the outer wall of the bend, thus avoiding damage to the pipeline structure and the need for high-pressure sealing. Therefore, it effectively avoids the risks of seawater corrosion and high-pressure leakage, and also provides an effective new approach to solving the problem of not being able to perform full-section detection in high-concentration, opaque mineral slurries. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the installation layout for the indoor model test proposed in this application;

[0082] Figure 2 This is a schematic diagram of the installation layout for the actual operating environment proposed in this application;

[0083] Figure 3 This is a flowchart of the concentration monitoring method for deep-sea mining hydraulic lifting pipelines;

[0084] Figure 4 This is a schematic diagram of the multi-input fusion neural network model proposed in this application;

[0085] In the above figures, 101 is a stirrer; 102 is a water pump; 103 is a test flow meter; 104 is a first guide wire; 105 is a first accelerometer; 106 is a model bend; 107 is a first data acquisition device; and 108 is a first computer.

[0086] 201. Relay compartment containing slurry; 202. Relay compartment valve; 203. Lifting force pipe for actual operating environment; 204. Flow meter; 205. Second guide wire; 206. Second accelerometer; 207. Second data acquisition instrument; 208. Second computer; Detailed Implementation

[0087] The technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described solutions are some embodiments of this application, but not all embodiments. Based on the embodiments proposed in this application, those skilled in the art can make improvements or modifications without creative effort, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0088] Example 1: To address the problem of distorted slurry concentration measurement in existing technologies, this application proposes a concentration monitoring system for deep-sea mining hydraulic lifting pipelines. This system includes a source domain data construction and model pre-training subsystem for constructing a terrestrial source domain dataset and generating pre-trained model parameters, and a target domain data acquisition and online migration monitoring subsystem for collecting deep-sea target domain data and loading pre-trained parameters to perform online monitoring.

[0089] like Figure 1 As shown, the source domain data construction and model pre-training subsystem includes a stirrer 101, a water pump 102 connected to the outlet end of the stirrer 101 for pumping the slurry prepared in the stirrer 101 into the model bend 106 to construct a circulation loop, a first acceleration sensor 105 installed on the outer wall of the model bend 106, and a test flow meter 103 installed in the model bend 106.

[0090] The first accelerometer 105 and the test flow meter 103 are both connected to the first data acquisition instrument 107 via the first wire 104. The collected acceleration and flow velocity data are transmitted to the first computer 108 respectively. The first computer 108 is configured to build a multi-input fusion network model and perform supervised pre-training using source domain data, and save the converged model parameters.

[0091] like Figure 2 As shown, the target domain data acquisition and online migration monitoring subsystem includes a relay chamber 201 containing slurry. A relay chamber valve 202 for controlling the slurry's entry into the actual working environment lifting pipe 203 is connected to the outlet end of the relay chamber 201. A second acceleration sensor 206 is installed on the outer wall of the bend of the actual working environment lifting pipe 203, preferably on the outer arch side of the bend in this embodiment. A flow meter 204 is installed in the actual working environment lifting pipe 203.

[0092] The second accelerometer 206 and the flow meter 204 are both connected to the second data acquisition device 207 via the second wire 205 to acquire real-time operating data and transmit it to the second computer 208. The second computer 208 is configured to receive pre-trained model parameters generated by the first computer 108, and combine the real-time acquired data to perform multi-core maximum mean difference algorithm transfer learning and concentration calculation, and output the final slurry concentration monitoring value.

[0093] The following is a further optimized implementation plan:

[0094] Agitator 101 is used to prepare mineral slurries of different concentrations;

[0095] Pump 102 is used to pump the prepared slurry into the model bend 106 to simulate the fluid transport conditions of the hydraulic lifting pipeline in deep-sea mining.

[0096] Test flow meter 103 is used to collect fluid flow velocity signals;

[0097] The first accelerometer 105 is preferably installed on the outer arch side of the model bend 106 to collect the pipe wall acceleration signal. Both the first accelerometer 105 and the second accelerometer 206 are high-frequency response piezoelectric accelerometers, capable of sensitively measuring the pipe wall acceleration signal. During actual monitoring, the centrifugal effect generated at the outer arch side when the slurry flows through the bend causes ore particles with a density much greater than seawater to be thrown towards the outer wall and collide with it. The installation of both the first accelerometer 105 and the second accelerometer 206 on the outer arch side of the bend essentially constructs a physical signal amplifier, making the sensors particularly sensitive to particle collision signals that are difficult to capture in straight pipe sections but significantly enhanced by inertial impact at the bend. When the slurry concentration fluctuates slightly, the time-frequency signal collected by the acceleration on the outer arch side will show significant changes, enabling the system to capture the implicit characteristics of concentration changes caused by changes in ore particles.

[0098] The model bend 106, serving as the physical carrier of the source domain data, can be made of transparent acrylic tubing, possessing a geometrically curved structure similar to that of actual deep-sea hydraulic lifting pipes. When constructing the source domain data, a sufficiently long distance is maintained between the bend in the pipe and the pump 102, allowing the flow of slurry particles within the pipe to be considered stable. This ensures that the collected source domain data is physically consistent with actual operating conditions, preventing model training deviations caused by flow anomalies and guaranteeing the integrity and effectiveness of the source domain data at the physical level.

[0099] The test flow meters 103 and 204 are both non-contact electromagnetic or ultrasonic flow meters used to collect the average flow velocity of the fluid in the pipe in real time. This equipment features a high sampling rate. The collected flow velocity data is processed by the system's built-in global average pooling algorithm, transforming it into a steady-state flow velocity feature vector. This vector serves as an independent input branch of a deep neural network, decoupling the nonlinear gain interference caused by flow velocity fluctuations on particle impact energy. The conversion of the collected flow velocity data into a steady-state feature vector allows the constructed neural network to effectively perceive the current energy background, thereby determining the nonlinear acceleration gain interference caused by flow velocity fluctuations when calculating concentration.

[0100] Both the first data acquisition unit 107 and the second data acquisition unit 207 are multi-channel dynamic signal data acquisition terminals, equipped with high-precision analog-to-digital conversion and anti-aliasing filtering functions. They are responsible for synchronously receiving analog signals from the accelerometer and flow meter and converting them into digital signals. Their built-in signal conditioning module can compensate for signal attenuation caused by long-distance transmission in the deep sea, ensuring accurate data acquisition.

[0101] The first computer 108, as the training host of the source domain model, integrates a high-performance graphics processor and a large-capacity storage unit and can exchange data with the second computer 208. It is responsible for building a neural network model including acceleration feature extraction branch and flow velocity feature extraction branch, and performing fully supervised learning using source domain data with accurate labels. It optimizes the model weights through the backpropagation algorithm and generates a pre-trained model parameter file with basic feature extraction capabilities.

[0102] The second computer 208, serving as the inference host for the target domain, integrates a high-performance graphics processor and a large-capacity storage unit. It receives pre-trained parameters generated by the first computer 108 and, in conjunction with unlabeled data collected on-site, executes a multi-core maximum mean difference transfer adaptation algorithm. It calculates the distribution differences between the source and target domains in real time and updates the model. Based on the results of the new model, it maps real-time acceleration and velocity data to slurry volume concentration values ​​and provides a visual interface for operators to monitor.

[0103] like Figure 3 and Figure 4 As shown, based on the above-mentioned deep-sea mining hydraulic lift pipeline concentration monitoring system, this application also proposes a deep-sea mining hydraulic lift pipeline concentration monitoring method including the following implementation steps:

[0104] S1. Construct the training dataset;

[0105] Acceleration and velocity data of hydraulic lift pipelines from multiple indoor model experiments were collected as the source domain dataset, and acceleration and velocity data of hydraulic lift pipelines from multiple deep-sea operational environments were collected as the target domain dataset. Both the source and target domain datasets were used as training datasets.

[0106] S11. Construct the source domain dataset;

[0107] In the indoor model experiment, the acceleration data and corresponding fluid velocity data of the mining hydraulic lifting pipeline were collected using the source domain data construction and model pre-training subsystem. The data collected in this environment eliminated the unknown interference of the actual deep-sea environment and only reflected the physical mapping relationship between concentration and collision acceleration signals, thus providing good source domain samples for the subsequent neural network.

[0108] First, the source domain data construction and model pre-training subsystem is started. Clean water is injected into the circulation loop and pre-circulation is performed to remove air bubbles. The first acquisition instrument 107 records the wall acceleration noise data of the model bend 106 under pure water conditions as the reference zero point of the system.

[0109] Subsequently, mineral slurries with various particle concentrations were prepared. The rotation speed of the stirrer 101 was controlled by the first computer 108 to keep the particles suspended. The weighed solid particles were added to the system in batches. With the help of sampling and verification, the volume concentration was precisely adjusted to the preset gradient. At the same time, the water pump 102 was adjusted to cover different flow rate ranges. After the system reached steady state, multi-dimensional data synchronous acquisition was triggered. The particle collision signal on the outer arch side of the model bend 106 was obtained using the first accelerometer 105. The fluid flow rate was recorded simultaneously using the experimental flow meter 103. At the same time, in order to enhance the generalization ability of the pre-trained model, variable control experiments were further performed for different particle size distributions and different sensor installation angles to collect frequency response characteristics under multiple dimensions.

[0110] Finally, wavelet denoising was performed on the collected raw signals to remove environmental interference. The denoised acceleration time series data and flow velocity data were matched with manually precisely prepared concentration labels to construct a source domain supervised dataset containing rich physical modes and stored in the database.

[0111] S12. Construct the target domain dataset;

[0112] In the actual deep-sea mining operation environment, the target domain data acquisition and online migration monitoring subsystem is used to collect acceleration data and corresponding fluid velocity data of the mining hydraulic lifting pipeline, and construct the target domain dataset.

[0113] Because the high hydrostatic pressure at depths of thousands of meters in the deep sea alters the structural damping characteristics of steel pipes, and ocean current excitation introduces low-frequency noise, the target domain data contains environmental characteristics not present in the source domain. Although these data have few or no labels, they accurately record the frequency domain regularity of particle collision signals under deep-sea conditions. The purpose of collecting this data is to use the multi-core maximum mean difference algorithm in the subsequent S4 step to calculate the distribution difference between the target domain data and the source domain data, thereby enabling cross-domain adaptation.

[0114] S13. Preprocess the source and target domain data;

[0115] One-dimensional time series data points of a set time window length are collected and extracted by the first accelerometer 105 and the second accelerometer 206. The preferred length is 1×2048, which is sufficient to include the complete energy release process of the interaction between the particles and the pipe, ensuring the integrity of the characteristics of particle impact and friction in the time and frequency domain.

[0116] Meanwhile, this window length can limit velocity fluctuations to a small range, avoiding the impact of non-stationary flow fields on signal statistics; furthermore, the velocity data is a one-dimensional time series data point of the same length, collected synchronously with the acceleration data, and the velocity data is normalized to match the acceleration data. This is to ensure that the fluid energy background information within the exact same time period is extracted in the subsequent network, thereby achieving accurate velocity decoupling;

[0117] S2. Construct a multi-input fusion neural network model;

[0118] The multi-input fusion neural network model includes an acceleration feature extraction branch and a flow velocity feature extraction branch, which fuse the features of both and output a concentration prediction value through a fully connected layer; including,

[0119] S21. Construct a deep neural network;

[0120] The deep neural network contains two input branches, and the logic for dividing the two input branches is based on the differences in frequency response characteristics and energy contribution mechanisms under different concentrations in the deep-sea booster pipe.

[0121] The first input branch processes the acceleration time-series signal. Since the collision of ore particles with the pipe wall is random and transient, and contains information such as impact, friction and rolling, this branch is designed as the main feature extraction branch.

[0122] The second input branch processes relatively low-frequency, gently changing velocity signals. In physical characteristics, velocity cannot directly represent concentration, but it amplifies the energy of particle impact and nucleus-particle friction. Therefore, this branch mainly corrects the concentration of the first branch, preventing the model from incorrectly identifying high-energy acceleration signals caused by high velocity as high concentration.

[0123] S22. Construct the acceleration feature extraction branch;

[0124] The acceleration feature extraction branch sequentially comprises a one-dimensional multi-scale convolutional neural network, a multi-head self-attention mechanism, and a bidirectional long short-term memory network; the specific parameters are set as follows:

[0125] The first branch of the one-dimensional multi-scale convolutional neural network: The first convolutional layer has 16 kernels and a kernel size of 3×1. This branch is used to capture high-frequency impact features. In deep-sea hydraulic lifting pipes, when ore particles impact the steel pipe wall at high speed, they generate transient pulses with extremely short durations and steep energy. The 3×1 convolutional kernel provides high resolution and sensitive capture of such transient spike signals, preventing them from being smoothed out in high background noise.

[0126] The second branch of the one-dimensional multi-scale convolutional neural network: The first convolutional layer has 16 kernels with a kernel size of 5×1. This branch is used to capture mid-frequency friction features. In deep-sea hydraulic lifting pipes, the continuous random friction signal characteristics generated when ore particles slide and roll on the pipe wall surface are observed. This signal is not as sharp as an impact, but it exhibits specific statistical regularities. The medium-sized 5×1 convolutional kernel can effectively extract the frequency domain envelope features brought about by this friction, reflecting the rheological properties of the slurry.

[0127] The third branch of the one-dimensional multi-scale convolutional neural network: the first convolutional layer with 16 kernels and a kernel size of 7×1. This branch focuses on capturing low-frequency macroscopic features. In deep-sea hydraulic lifting pipes, it captures the large-scale turbulent fluctuations of the fluid within the pipe and the periodic background signal characteristics generated by pump operation. The large 7×1 convolutional kernel, by increasing the receptive field, can smooth signal fluctuations caused by local particle impacts and extract the characteristics of water flow background and mechanical noise.

[0128] The three branches of the aforementioned one-dimensional multi-scale convolutional neural network are each connected to three convolutional layers after the first large convolutional kernel. The kernel size is uniformly set to 3×1, with the number of kernels increasing sequentially from 32, 64, and 128, and a stride of 1 for each layer. As the network depth increases, the model gradually extracts higher-level features representing collision frequency and impact energy entropy, moving beyond simply extracting signal variation features. Furthermore, each convolutional layer is followed by a 2×1 pooling layer with a stride of 2. This ensures that the model ignores minute offset signals, maintaining stable concentration features even with slight fluctuations in the particle impact time. It also helps subsequent convolutional layers capture higher-level, more global signal information.

[0129] Multi-head self-attention mechanism layer: This layer, connected after the multi-scale feature extraction layer, is configured with 128 kernels and a Dropout ratio of 0.5. While the three branches of the aforementioned one-dimensional multi-scale convolutional neural network successfully extracted features of particle impact, particle friction, and macroscopic fluid background noise, these features are physically independent. Therefore, a multi-head self-attention mechanism layer was constructed. This multi-head mechanism allows different heads to operate independently in different subspaces: some heads focus on mid-frequency friction and macroscopic global features, while others focus on local particle impact features. Furthermore, in the multi-scale network, this layer can dynamically adjust the weights of features from different resolutions according to the evolution of physical laws, compensating for characteristics representing particle impact and mid-frequency friction, and reducing low-frequency water flow background and mechanical noise, ensuring that the established neural network model focuses on key temporal signal features.

[0130] Bidirectional Long Short-Term Memory (LSTM) network layer: This layer consists of two stacked layers used to extract temporal dependencies; the first layer, a BiLSTM, has 128 hidden nodes, and the second layer has 256 hidden nodes. In deep-sea hydraulic lifting pipelines, the flow of slurry over long distances exhibits spatiotemporal continuity. The current concentration state depends not only on the current acceleration signal but also on the inertia of the flow regime from the previous moment. The bidirectional LSM network layer utilizes both historical flow regimes and future trends, effectively preserving key physical information from long-term signals while automatically filtering out signal characteristics that do not conform to physical laws, enabling the model to output stable and continuous concentration monitoring curves.

[0131] S23. Construct the flow velocity feature extraction branch;

[0132] A global statistical pooling layer is used to extract steady-state and trend features of the velocity sequence and map them into a high-dimensional velocity feature vector. The velocity feature extraction branch utilizes the global statistical pooling layer to extract steady-state features of the velocity sequence, performing a dimensionality upscaling operation on the velocity data to effectively establish an energy reference system for the acceleration signal.

[0133] S231. The flow velocity data and acceleration data maintain a consistent 2048-window length to ensure the extraction of fluid energy background information within the exact same time period; simultaneously, a flow velocity sequence with a preferred length of 2048 is selected. Global average pooling is performed to obtain steady-state flow velocity characteristics. This operation ensures that the velocity characteristics obtained by the model are not affected by the random value fluctuations at a certain instant, and obtains the steady-state velocity characteristics over that time span.

[0134] S232, will The features are input to a fully connected layer, which maps them to a high-dimensional flow velocity feature vector. ;

[0135] In this embodiment, the preferred length of the vector is 32. In the subsequent neural network, the gradient contribution generated by low-dimensional features is easily masked by the gradient of high-dimensional acceleration features, causing the network to forget the influence of flow velocity during training. Mapping the flow velocity to a 32-dimensional vector ensures that the flow velocity feature occupies a reasonable weight proportion in the concatenated total feature vector, ensuring that the model can sensitively capture the influence of flow velocity on the prediction results. At the same time, the 32-dimensional vector length enables the subsequent neural network to learn the gain of different flow velocities on different frequency bands of energy.

[0136] S233. The temporal acceleration feature vector output by the bidirectional long short-term memory network in step S22. The high-dimensional flow velocity feature vector output in step S232 The features are concatenated to obtain a multi-source fused feature vector. By constructing a joint space containing acceleration and flow velocity signals, the subsequent neural network can be provided with a complete information foundation.

[0137] S24, Feature Fusion and Regression Layer;

[0138] The first fully connected layer has 256 nodes and is mainly responsible for receiving the acceleration and velocity signals from the front end. With 256 neurons, the model can retain the representation of multi-scale spatial features and velocity characteristics to the maximum extent. The second fully connected layer has 64 nodes. Through nonlinear transformation, it performs secondary filtering and combination of the 256-dimensional redundant features, extracting high-order features that affect the particle concentration in the hydraulic lift pipeline, filtering out irrelevant variable information, and finally connecting to the output layer with 1 node to output the particle concentration value in the hydraulic lift pipeline.

[0139] S3, Pre-training;

[0140] The multi-input fusion neural network model is pre-trained using source domain datasets in a supervised learning manner;

[0141] Because the actual deep-sea operating environment differs from the indoor testing environment, an initial state with basic physical feature extraction capabilities is constructed in the neural network parameter space to provide a model foundation for subsequently solving the target domain of the actual deep-sea operating environment; including,

[0142] S31. First, the source domain training set is input into the model, and forward propagation calculation is performed. The acceleration signal branch extracts the frequency domain characteristics of particle collision, particle friction and water flow background noise, and the flow velocity branch calculates the background energy gain coefficient. Then, the regression layer outputs the predicted value of ore concentration of deep-sea mining hydraulic lifting pipeline in the source domain by combining the information of both.

[0143] S32. Construct the regression loss function;

[0144] The mean squared error loss function is adopted, and the total loss function is constructed by combining the sum of squares of all weight parameters in the multi-input fusion neural network model (L2 regularization term). The total error between the predicted value output by the model and the real physical label is calculated in this way.

[0145] S33. After calculating the loss value for the current training batch in step S32, backpropagation is performed based on this loss value to calculate the gradient of the parameters of each layer of the network. The Adam optimizer is then used to iteratively update the weights and biases in the model to minimize the prediction error. In multi-scale neural networks, the contribution of loss functions at different scales to the total gradient varies. The Adam optimizer can independently calculate adaptive learning rates for neurons at different physical scales. By standardizing the step size at the parameter level for each frequency band feature channel, multi-branch, multi-scale collaborative optimization can be effectively achieved.

[0146] S34. Determine whether the training rounds have reached the preset maximum number of iterations or whether the loss function has converged; if the above termination conditions are met, stop training and save the model parameters trained in the current source domain.

[0147] S4. Introduce transfer learning for retraining and updating;

[0148] To address the data distribution drift issue caused by differences in material acoustic impedance, environmental pressure, and boundary conditions between laboratory models and the actual deep-sea mining environment, transfer learning is introduced. This allows models pre-trained in the source domain to be adaptively applied to the unlabeled deep-sea target domain. Using source and target domain datasets, a multi-kernel maximum mean difference transfer learning method is employed to retrain and update the multi-input fusion neural network model. This includes...

[0149] S41. Introduce the multi-core maximum mean difference algorithm;

[0150] The distribution distance between source domain data and target domain data in the model fusion feature layer is measured, and the source domain data and target domain data are mapped to the regenerating kernel Hilbert space;

[0151] S42. Perform transfer layer discrimination. For network layers in the model, based on the multi-core maximum mean difference algorithm described in step S41, calculate the distribution distance between the source domain data and the target domain data in the output feature space of the network layer; compare the distribution distance with a preset transfer distance threshold: if the distribution distance is less than the preset transfer distance threshold, it is determined that the layer is transferred, and the following step S43 is executed; if the distribution distance is not less than the preset transfer distance threshold, it indicates that there is a significant inter-domain difference in the feature distribution of the layer, and it is determined that the layer is not transferred, and the following step S44 is executed.

[0152] S43. Load the source domain parameters saved in step S3 as initial values; regardless of the working conditions, the physical fact that particle impact and particle friction generate acceleration signals remains unchanged, and the pre-trained convolutional kernel has learned how to capture this acceleration.

[0153] S44. Reinitialize the parameters of the non-migrating layer. Due to differences in material acoustic impedance, environmental pressure and boundary conditions between actual deep-sea mining environments, resetting this layer is to allow the model to forget the characteristics of the indoor test environment and relearn the mapping relationship under the actual deep-sea mining environment.

[0154] S45. Construct the total loss function;

[0155] The transfer learning model is obtained by training with target domain data and minimizing the total loss function through backpropagation algorithm.

[0156] By constructing a total loss function, the neural network can extract domain-invariant physical features; specifically,

[0157] Calculate source domain distribution and target domain distribution The formula for calculating the maximum mean difference of multiple kernels in the regenerating kernel Hilbert space is as follows; by minimizing this value, the neural network finds a perspective from which the mean vector in the actual deep-sea operating environment is similar to that in the indoor experimental environment:

[0158] ;

[0159] in, Represents source domain distribution and target domain distribution The square of the maximum mean difference between the multi-core samples; Indicates that it follows the source domain distribution Source domain samples; Indicates that it follows the target domain distribution Target domain samples; Represents the feature mapping function; Indicates the distribution of the source domain Find the expected value; Indicates the distribution of the target domain Find the expected value; In the regenerating nucleus Hilbert space The square of the norm in the vector is used to measure the distance between two feature mean vectors;

[0160] With feature mapping function The feature kernel is defined as the weighted sum of three independent Gaussian kernels to represent the scale range of different frequency bands. The formula for calculating the weighted sum is as follows:

[0161] ;

[0162] in, This is a combined kernel function used to measure the difference in distributions; The squared Euclidean distance between the two samples in the feature space; For the first One kernel function; For the first The bandwidth parameters of each Gaussian kernel correspond to frequency bands of different scales; For the first The weight coefficients of each core, and satisfying In order to cover the complex multi-scale physical characteristics of signals in hydraulic lifting pipelines, Using a multi-level scaling factor, the feature median distance of the dataset is defined as follows: The bandwidths of the three Gaussian kernels are then set to... (Small bandwidth core) corresponds to the high frequency particle impact characteristics and is used to eliminate the impact sound sharpness shift caused by environmental and material differences; (Mid-bandwidth kernel) corresponds to the mid-frequency particle tribotexture features and is used to align the main signal energy distribution; (Large bandwidth kernel) corresponds to low-frequency fluid background features and is used to eliminate the differences in fluid turbulence structure distribution caused by environmental differences. Through adaptive setting based on the median, it is ensured that no matter how the dimension of the feature space changes, the maximum mean difference of the multi-kernel can stably measure the distribution differences of indoor test environment and deep-sea mining actual working environment in different physical frequency bands.

[0163] The total loss function is then defined as follows:

[0164] ;

[0165] in, This is the balance coefficient; This represents the total loss function value during model training; : Indicates the loss of source domain data; This represents the multi-kernel maximum mean difference loss, used to measure the distribution difference between the source and target domains in the feature space;

[0166] The model (multi-input fusion neural network model) updates the weights by minimizing the total loss function, thereby reducing the distribution difference between the source and target domains in the feature space;

[0167] S5. Real-time measurement;

[0168] The real-time acceleration and velocity data of the deep-sea hydraulic lifting pipeline to be measured are input into the optimized multi-input fusion neural network model, and the real-time measurement results of the slurry volume concentration are output.

[0169] Specifically, the real-time acceleration and velocity data of the deep-sea hydraulic lifting pipeline to be measured are input into the multi-input fusion neural network model optimized in step S4, and flow cytometry inference and physical post-processing are performed to output the real-time measurement results of the slurry volume concentration; including,

[0170] First, perform real-time streaming inference using a sliding window approach;

[0171] The transfer learning model deployed on the monitoring host of the mining vessel is configured in online inference mode. The system continuously receives acceleration data streams and velocity data streams from the target domain data acquisition and online transfer monitoring subsystem. According to the preprocessing standard described in step S13, a sliding window mechanism is used to extract the data streams, with an update step size of 512 data points and a processing window length of 2048 data points. Through the sliding window design with a high overlap rate, the monitoring system is ensured to have a millisecond-level response speed, which can capture transient concentration fluctuations during the slurry flow and effectively prevent missed detections caused by excessively large sampling intervals.

[0172] Secondly, perform post-processing algorithms such as exponentially weighted moving average or Kalman filtering;

[0173] The original concentration prediction value is obtained at the output of the multi-input fusion neural network model. The original concentration prediction value is smoothed by exponential weighted moving average or Kalman filter algorithm. Since the change of solid phase concentration in the tube is necessarily continuous in time, high-frequency non-physical jumps can be effectively filtered out, the real concentration change trend is retained, and the final output monitoring curve is ensured to conform to the objective physical laws.

[0174] Finally, it implements multi-dimensional visualization and industrial closed-loop control.

[0175] Not only can it draw "concentration-time" trend curves in real time through a digital twin interface, intuitively displaying the evolution trajectory of solid phase distribution within the pipeline; furthermore, it can encapsulate the calculated final concentration data and early warning status into standard industrial data packets, and transmit them in real time to the mining vessel's central control system via standard communication protocols such as Modbus TCP or MQTT; the mining vessel's central control system makes intelligent judgments based on the calculated concentration change rate, and when it detects a risk of pipe blockage, it automatically triggers a safety mechanism, sending feedback commands to the mining vessel's central control system to adjust the collection speed of the seabed ore collector or open the emergency water injection valve, thereby achieving closed-loop control from passive monitoring to active defense.

[0176] Although the present invention has been shown and described in the above embodiments, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A concentration monitoring system for a deep-sea mining hydraulic lifting pipeline, characterized in that: It includes a source domain data construction and model pre-training subsystem for constructing terrestrial source domain datasets and generating pre-trained model parameters, and a target domain data acquisition and online migration monitoring subsystem for collecting deep-sea target domain data and loading pre-trained parameters to perform online monitoring. The source domain data construction and model pre-training subsystem includes a stirrer, a pump connected to the outlet of the stirrer for pumping the prepared slurry into the model bend to construct a circulation loop, a first accelerometer installed on the outer wall of the model bend, and a test flow meter installed inside the model bend; both the first accelerometer and the test flow meter are connected to a first data acquisition instrument via a first wire, and the acquired acceleration and flow velocity data are transmitted to a first computer; the first computer is configured to construct a multi-input fusion neural network model and perform supervised pre-training using the source domain data, and save the converged model parameters; The target domain data acquisition and online migration monitoring subsystem includes a relay compartment containing slurry. A relay compartment valve for controlling the slurry's entry into the actual operating environment lifting pipe is connected to the outlet end of the relay compartment. A second accelerometer is installed on the outer wall of the bend of the actual operating environment lifting pipe. A flow meter is installed in the actual operating environment lifting pipe. Both the second accelerometer and the flow meter are connected to a second data acquisition unit via a second wire to acquire real-time operating data and transmit it to a second computer. The second computer is configured to receive pre-trained model parameters generated by a first computer and, in conjunction with the real-time acquired data, perform multi-core maximum mean difference algorithm transfer learning and concentration calculation to output the final slurry volume concentration monitoring value.

2. The concentration monitoring system for deep-sea mining hydraulic lifting pipelines according to claim 1, characterized in that: Both the first and second accelerometers are installed on the outer arch side of the bend, and both are high-frequency piezoelectric accelerometers.

3. A method for monitoring the concentration of a deep-sea mining hydraulic lift pipeline using the concentration monitoring system according to claim 1 or 2, characterized in that: It includes the following steps, S1. Construct the training dataset; Acceleration and velocity data of hydraulic lifting pipelines in multiple indoor model experiments were collected as source domain datasets, and acceleration and velocity data of hydraulic lifting pipelines in multiple deep-sea actual operation environments were collected as target domain datasets. The source domain datasets and target domain datasets were used as training datasets. S2. Construct a multi-input fusion neural network model; The multi-input fusion neural network model includes an acceleration feature extraction branch and a flow velocity feature extraction branch, which fuse the features of the two and output the concentration prediction value through a fully connected layer; S3, Pre-training; The multi-input fusion neural network model is pre-trained using source domain datasets in a supervised learning manner; S4. Introduce transfer learning for retraining and updating; Using source domain datasets and target domain datasets, a multi-kernel maximum mean difference transfer learning method is used to retrain and update the multi-input fusion neural network model; S5. Real-time measurement; The real-time acceleration and velocity data of the hydraulic lifting pipeline to be measured are input into the optimized multi-input fusion neural network model, and the real-time measurement results of the slurry volume concentration are output. The real-time acceleration and velocity data of the hydraulic lifting pipeline to be measured are input into the multi-input fusion neural network model optimized in step S4, and the flow inference and physical post-processing are performed to output the real-time measurement results of the slurry volume concentration.

4. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 3, characterized in that: Step S1 includes, S11. Construct the source domain dataset; In the indoor model experiment, the acceleration data and corresponding fluid velocity data of the hydraulic lifting pipeline were collected using the source domain data construction and model pre-training subsystem. S12. Construct the target domain dataset; In the actual deep-sea mining operation environment, the target domain data acquisition and online migration monitoring subsystem is used to collect acceleration data and corresponding fluid velocity data of hydraulic lifting pipelines to construct a target domain dataset. S13. Preprocess the source and target domain data; One-dimensional time series data points of a set time window length are collected and extracted using the first and second accelerometers.

5. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 3, characterized in that: Step S2 includes, S21. Construct a deep neural network; The deep neural network contains two input branches. The logic for dividing the two input branches is based on the differences in frequency response characteristics and energy contribution mechanisms under different concentrations in the hydraulic lifting pipe. The first input branch processes the acceleration time-series signal, while the second input branch processes the relatively low-frequency, gently changing flow velocity signal. S22. Construct the acceleration feature extraction branch; The acceleration feature extraction branch sequentially includes a one-dimensional multi-scale convolutional neural network, a multi-head self-attention mechanism, and a bidirectional long short-term memory network. S23. Construct the flow velocity feature extraction branch; A global statistical pooling layer is used to extract the steady-state and trend features of the velocity sequence and map them into a high-dimensional velocity feature vector. The velocity feature extraction branch uses the global statistical pooling layer to extract the steady-state features of the velocity sequence, and performs a dimensionality upscaling operation on the velocity data, which can effectively establish an energy reference system for the acceleration signal. S24, Feature Fusion and Regression Layer; The first fully connected layer has 256 nodes and is used to receive the acceleration and velocity signals output from the front end. The second fully connected layer has 64 nodes. Through nonlinear transformation, it performs secondary screening and combination of the 256-dimensional redundant features to extract the high-order features that affect the slurry volume concentration in the hydraulic lifting pipeline, filters out irrelevant variable information, and finally connects to the output layer with 1 node to output the slurry volume concentration in the hydraulic lifting pipeline.

6. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 5, characterized in that: Step S22 includes setting the following parameters: One-dimensional multi-scale convolutional neural network scale 1st branch: The first convolutional layer has 16 kernels and a kernel size of 3×1. This branch is used to capture high-frequency collision features. The second branch of the one-dimensional multi-scale convolutional neural network: the first convolutional layer with 16 kernels and a kernel size of 5×1. This branch is used to capture mid-frequency friction features. The third branch of the one-dimensional multi-scale convolutional neural network: the first convolutional layer with 16 kernels and a kernel size of 7×1. This branch focuses on capturing low-frequency macroscopic features. Multi-head self-attention mechanism layer: This layer is connected after the multi-scale feature extraction layer, with a kernel count of 128 and a Dropout ratio of 0.5; Bidirectional Long Short-Term Memory Network Layer: Contains a two-layer stacked structure for extracting temporal dependencies; the first layer of BiLSTM has 128 hidden nodes; the second layer of BiLSTM has 256 hidden nodes.

7. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 6, characterized in that: Step S23 includes, S231. The velocity data and acceleration data maintain a consistent 2048-window length to ensure the extraction of fluid energy background information within the exact same time period; simultaneously, the velocity sequence of length 2048... Global average pooling is performed to obtain steady-state flow velocity characteristics. ; S232, will The features are input to a fully connected layer, which maps them to a high-dimensional flow velocity feature vector. ; S233. The temporal acceleration feature vector output by the bidirectional long short-term memory network in step S22. The high-dimensional flow velocity feature vector output in step S232 The features are concatenated to obtain a multi-source fused feature vector. .

8. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 3, characterized in that: The steps described Step S3 includes, S31. First, the source domain training set is input into the model, and forward propagation calculation is performed. The acceleration signal branch extracts the frequency domain characteristics of particle collision, particle friction and water flow background noise, and the flow velocity branch calculates the background energy gain coefficient. Then, the regression layer outputs the predicted value of slurry volume concentration of the deep-sea mining hydraulic lifting pipeline in the source domain by integrating the information of both. S32. Construct the regression loss function; The mean squared error loss function is used, and the total loss function is constructed by combining the sum of squares of all weight parameters in the multi-input fusion neural network model. This is used to calculate the total error between the predicted value output by the model and the real physical label. S33. After calculating the loss value of the current training batch in step S32, backpropagation is performed based on the loss value to calculate the gradient of the parameters of each layer of the network, and the Adam optimizer is used to iteratively update the weights and bias parameters in the model to minimize the prediction error. S34. Determine whether the training rounds have reached the preset maximum number of iterations or whether the loss function has converged. If so, stop training and save the model parameters trained in the current source domain.

9. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 3, characterized in that: Step S4 includes, S41. Introduce the multi-core maximum mean difference algorithm; The distribution distance between source domain data and target domain data in the model fusion feature layer is measured, and the source domain data and target domain data are mapped to the regenerating kernel Hilbert space; S42. Perform transfer layer discrimination. For network layers in the model, based on the multi-core maximum mean difference algorithm described in step S41, calculate the distribution distance between the source domain data and the target domain data in the output feature space of the network layer; compare the distribution distance with a preset transfer distance threshold to indicate that the feature distribution of the network layer is similar between the source domain and the target domain; if the distribution distance is less than the preset transfer distance threshold, determine that the layer is to be transferred, and proceed to step S43; if the distribution distance is not less than the preset transfer distance threshold, it indicates that there is a significant inter-domain difference in the feature distribution of the layer, and determine that the layer is not to be transferred, and proceed to step S44. S43. Load the source domain parameters saved in step S3 as initial values; S44. Reinitialize the parameters of the non-transfer layer; S45. Construct the total loss function; The transfer learning model is obtained by training with target domain data and minimizing the total loss function through backpropagation algorithm. By constructing a total loss function, the neural network can extract domain-invariant physical features; Calculate source domain distribution and target domain distribution The formula for calculating the maximum mean difference of multiple kernels in the regenerating kernel Hilbert space is as follows: in, Represents source domain distribution and target domain distribution The square of the maximum mean difference between the multi-core samples; Indicates that it follows the source domain distribution Source domain samples; Indicates that it follows the target domain distribution Target domain samples; Represents the feature mapping function; Indicates the distribution of the source domain Find the expected value; Indicates the distribution of the target domain Find the expected value; In the regenerating nucleus Hilbert space The square of the norm in the vector is used to measure the distance between two feature mean vectors; Feature mapping function The feature kernel is defined as the weighted sum of three independent Gaussian kernels to represent the scale range of different frequency bands. The formula for calculating the weighted sum is as follows: in, This is a combined kernel function used to measure the difference in distributions; The squared Euclidean distance between the two samples in the feature space; For the first One kernel function; For the first The bandwidth parameters of each Gaussian kernel correspond to frequency bands of different scales; For the first The weight coefficients of each core, and satisfying ; The total loss function is then defined as follows: ； in, This is the balance coefficient; This represents the total loss function value during model training; This indicates the loss of source domain data; This represents the multi-core maximum mean difference loss, used to measure the distribution difference between the source and target domains in the feature space.

10. The method for monitoring concentration in deep-sea mining hydraulic lifting pipelines according to claim 4, characterized in that: Step S5 includes, First, a sliding window-based real-time streaming inference is performed; the transfer learning model deployed on the mining vessel monitoring host is configured in online inference mode, and the system continuously receives acceleration data streams and velocity data streams from the target domain data acquisition and online migration monitoring subsystem; according to step S13, the source domain and target domain data are preprocessed, and the data streams are truncated using a sliding window mechanism; Secondly, perform post-processing algorithms such as exponentially weighted moving average or Kalman filtering; The original concentration prediction value is obtained at the output of the multi-input fusion neural network model, and the original concentration prediction value is smoothed by exponential weighted moving average or Kalman filter algorithm. Finally, multi-dimensional visualization and industrial closed-loop control are implemented; the final concentration data and early warning status after calculation are encapsulated into a standard industrial data package and transmitted in real time to the mining vessel's central control system via Modbus TCP or MQTT communication protocol; the mining vessel's central control system makes intelligent judgments based on the calculated concentration change rate.

Images