High-pressure jet pipeline pressure early warning method and system

Abstract

The invention relates to online intelligent detection, and discloses a high-pressure jet pipeline pressure early warning method and system so as to carry out early warning on possible pressure drop or fluctuation. The method comprises the following steps: training, verifying and testing a surface quality rating model based on image data of a shot target surface in a data set before and after cleaning of different jet pressure grades; according to rating data obtained before and after cleaning of any cleaning area of the shot target surface based on the trained surface quality rating model, an observation sequence is constructed in combination with pressure data of the high-pressure pump in the cleaning process; three key parameters of a jet pipe pressure estimation HMM model are estimated based on an unsupervised learning method for the observation sequences of the corresponding number; decoding the implicit state of the real-time observation sequence in a period of time window according to the key parameters of the HMM model to obtain a jet pipe pressure state sequence, and judging whether the pressure is abnormal or not according to the value change of the jet pipe pressure state sequence and executing early warning processing.

Description

Technical Field

[0001] The present invention relates to the field of online intelligent detection technology, and in particular to a high-pressure jet pipeline pressure early warning method and system. Background Art

[0002] In a high-pressure abrasive water jet surface cleaning device, the cleaning process takes place in a cleaning station. Each cleaning station contains a group of jet pipes. The terminal of each jet pipe is connected to a jet nozzle. The high-pressure pump generates high-pressure water, which is transported to the jet nozzle through the jet pipe. After mixing with the abrasive, it impacts the steel plate to remove surface impurities. Each cleaning station can complete the cleaning of a strip sub-area along the conveying direction of the steel plate. The cleaning effect depends to a large extent on the pressure of the high-pressure water in the jet pipe. In a surface cleaning device, ideally, the pressure of each jet pipe in the same group should be uniform and close. However, due to pipeline wear, foreign matter, turbulence effect, burst pipe and other problems, the pressure of some jet pipes drops or fluctuates, resulting in a decrease in cleaning quality.

[0003] In industry, open-hole pressure gauges or pressure sensors are generally used to measure the pressure of metal high-pressure pipelines. However, in high-pressure water jet devices, the jet nozzles need to be frequently disassembled, repaired, and the angles adjusted. Therefore, the jet tubes connected to the jet nozzles are made of non-metallic materials. Local openings will weaken the structural strength, and long-term use will make leakage problems more likely to occur. Therefore, pressure gauges or pressure sensors are only suitable for jet tube pressure detection in the experimental stage, and are not feasible as long-term online detection devices. Summary of the invention

[0004] The purpose of the present invention is to disclose a high-pressure jet pipeline pressure early warning method and system, so as to evaluate the pressure of the jet tube by observing the time series data of the surface cleaning quality, and to give early warning of possible pressure drops or fluctuations.

[0005] To achieve the above purpose, the high-pressure jet pipeline pressure early warning method disclosed in the present invention includes: Step S1, training, verifying and testing a surface quality rating model based on image data of the target surface before and after cleaning at different jet pressure levels in the data set; Step S2, constructing an observation sequence based on the rating data before and after cleaning obtained from any cleaning area on the surface of the target based on the trained surface quality rating model and combining it with the pressure data of the high-pressure pump during the cleaning process; Step S3, for a corresponding number of observation sequences, estimating the initial state distribution vector, state transition probability matrix and observation probability distribution of the jet tube pressure estimation HMM model based on an unsupervised learning method; Step S4, decoding the implicit state of the real-time observation sequence within a time window according to the initial state distribution vector, state transition probability matrix and observation probability distribution of the HMM (Hidden Markov Model) model to obtain the jet tube pressure state sequence, and then judging whether the pressure is abnormal according to the value change of the jet tube pressure state sequence, if so, executing early warning processing.

[0006] Preferably, the time series data formed by the corresponding number of observation sequences The form is: ;in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, Indicates the cleaning effect. represents the set pressure of the high-pressure pump, T represents the number of data samples, represents the time index of the corresponding sample in the time series data, and .

[0007] Preferably, the unsupervised learning method in step S3 adopts the Baum-Welch algorithm, and the algorithm for decoding the implicit state in step S4 adopts the Viterbi algorithm.

[0008] Optionally, the target to be shot is a steel plate, and the jet tube pressure state includes five states: low pressure, relatively low pressure, average pressure, relatively high pressure and high pressure. Further, the same steel plate corresponds to at least two camera positions during the cleaning process, and the data set of step S1 includes image data of the area corresponding to each camera position before and after cleaning, and in step S4, the real-time operating state of each camera position is synchronously detected online with parallel threads.

[0009] To achieve the above objectives, the present invention also discloses a high-pressure jet pipeline pressure warning system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method when executing the computer program.

[0010] Preferably, the high-pressure jet pipeline pressure early warning system of the present invention specifically includes: Data acquisition subsystem: including a camera device, an image acquisition module, a historical database module and a data processing module; wherein the image acquisition module is used to collect real-time image data of the steel plate surface before and after cleaning, the historical database module is used to store the steel plate surface images before and after cleaning and the high-pressure pump pressure data, and the data processing module is used to pre-process the images of the steel plate before and after cleaning; Steel plate surface quality evaluation subsystem: It includes a surface image dataset and a surface quality evaluation network. The surface image dataset consists of sub-images before and after cleaning divided according to the steel plate conveying direction and cleaning position, and the quality labels corresponding to the images. The surface quality evaluation network is constructed by a network based on a deep learning architecture. The network is trained and evaluated using the dataset. Once the requirements are met, a surface quality evaluation model is generated. Jet tube pressure interval estimation subsystem: used to train the pressure estimation model based on the time series data set. The time series data of the time series data set is ,in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, indicates the cleaning effect, and , Indicates the set pressure of the high pressure pump. represents the number of data samples, Indicates the time index of the corresponding sample in the time series data. The pressure estimation model training adopts the Baum-Welch algorithm. By maximizing the likelihood function of the observation sequence, the three key parameters of the HMM model, namely the initial state distribution vector π, the state transition probability matrix A and the observation probability distribution B, are estimated. Finally, the HMM model of the jet tube pressure state estimation is determined. The jet tube pressure warning subsystem for time series data collection and real-time pressure estimation: including time series data collection and real-time pressure estimation modules, and selecting the appropriate time window length according to the equipment operation experience , according to the window length and time sequence, the real-time observation time series data about the cleaning quality is formed , and then decode the implicit pressure state according to the Viterbi algorithm to obtain the pressure state sequence ,according to According to the changes in the medium pressure state, low pressure or pressure fluctuation alarm will be issued.

[0011] The present invention has the following beneficial effects: 1. According to the changes in the surface quality of the steel plate and other targets before and after cleaning, the pressure of the jet tube can be monitored according to the cleaning stations in the cleaning device, which is helpful to achieve rapid early warning and fault location.

[0012] 2. It overcomes the problem that the high-pressure jet tube cannot be installed with an online pressure detection instrument, and realizes the online real-time estimation of the pressure of the high-pressure jet tube, which is conducive to further ensuring the cleaning quality.

[0013] 3. In-depth analysis based on the calculated changes in the pressure state of the jet tube is conducive to further evaluating the operating status of the high-pressure water jet surface cleaning device; for example, it can provide early warnings of low pressure or pressure fluctuations based on the changes in the decoded implicit pressure state.

[0014] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a specific deployment diagram of the actual application scenario of the high-pressure jet pipeline early warning system disclosed in the embodiment of the present invention in the cleaning of wide steel plates.

[0016] Figure 2 yes Figure 1 Schematic diagram of the partitioning of medium-width steel plates.

[0017] Figure 3 It is a schematic flow chart of a high-pressure jet pipeline early warning method disclosed in an embodiment of the present invention.

[0018] Figure 4 It is a specific functional block diagram of each subsystem in the high-pressure jet pipeline early warning system disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0020] Example 1 This embodiment discloses a high-pressure jet pipeline early warning method applied to wide steel plates, comprising the following steps: Step 1: Use different pressure ranges for cleaning, collect images before and after cleaning, and create a data set for the surface quality rating model. The specific process is: (1) Based on the operating experience of the surface cleaning device, the pressure of the high-pressure pump is divided into several intervals from low to high. At each different pressure level, a set number of steel plates are cleaned, and images of each steel plate before and after cleaning are collected.

[0021] (2) According to the number m of cleaning stations in the surface cleaning device, the steel plate image is divided into m rectangular sub-images along the width direction (conveying direction) (i.e., one station is responsible for cleaning the steel plate area corresponding to one rectangular sub-image). The surface quality of the sub-image is judged based on experience, and these sub-images are manually labeled with surface quality grade labels to create a data set D1, that is, D1 consists of the sub-images segmented from the steel plate image and the surface quality grade labels corresponding to the sub-images.

[0022] Step 2: Establish a deep learning network for surface cleaning quality rating, train the network using dataset D1, and generate a deployable surface quality rating model.

[0023] Preferably, in this step, D1 can be divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to train the deep learning network. The input of the network is image data, i.e., the rectangular sub-images of the steel plate before and after cleaning described in step 1. The network output layer is a fully connected layer + Softmax activation. The output cleaning quality category is the surface quality classification, i.e., the label content of the surface quality grade described in step 1. The loss function is the cross entropy loss, and the training goal is to make the loss function At the minimum, AdamW can be used as the optimizer, and the validation set can be used to evaluate the network performance. The indicators include accuracy and F1-Score. After meeting the requirements, the test set can be used for independent testing to evaluate the generalization ability of the network. After meeting the requirements, a deployable surface quality rating model in ONNX format is generated.

[0024] Loss Function Can be: Among them, the tag , c is the total number of label categories for surface quality ratings, and the predicted probability distribution is a probability vector, each element of which represents the predicted probability of a certain surface quality grade category.

[0025] Step 3: Use the surface quality rating data output from step 2 as observable input features to construct input data sequence samples, and automatically optimize the three key parameters of the HMM model through the Baum-Welch algorithm to generate observable data. Mapping to the jet tube pressure range The HMM model of the hidden state space of . The specific process is: (1) Constructing observable input feature sequences ,in, Represents the time index of the corresponding sample in the time series data, represents the number of data samples in the sequence, is the surface quality rating before cleaning, is the surface quality rating after cleaning, , indicating the cleaning effect, Indicates the set pressure of the high-pressure pump and specifies the implicit jet pipe pressure state sequence , ~ They represent (low pressure, lower, average, higher, high), that is, the pressure in the jet tube is The value of time belongs to ~ One of the five states.

[0026] (2) Input sequence through features And unsupervised learning method is used to estimate the three key parameters of the jet tube pressure estimation HMM model, which include the initial state distribution vector π, the state transition probability matrix A and the observation probability distribution B.

[0027] During the processing, first Normalize to get the input feature input sequence represented by real values , and randomly initialize π, A and B, set the maximum number of iterations and convergence threshold, and then use the Baum-Welch algorithm to estimate the parameters π, A and B of the HMM model by maximizing the likelihood function of the observation sequence.

[0028] For example, if the range of the jet tube pressure state is specified as S = (low pressure, lower, average, higher, high), then the number of states N = 5, and the initial state distribution row vector The state transition probability matrix A is a 5×5 matrix, and the observation probability distribution B represents each implicit state of the jet tube pressure. Generate observable input features The probability that the observed variable is a continuous four-dimensional feature vector (mass before cleaning, mass after cleaning, change in cleaning mass, set pressure of high-pressure pump), so B is composed of each implicit state The mean and covariance matrix of is randomly initialized when using the Baum-Welch algorithm. , since the pressure state transfer of the jet tube is mainly concentrated in the adjacent state (pressure changes relatively slowly), the matrix A is randomly initialized and normalized, for example: .

[0029] When using the Baum-Welch algorithm, the input sequence is based on the characteristics The data set updates the values ​​of π, A, and B.

[0030] Step 4: Use the HMM model to estimate the surface quality observable sequence input in real time to obtain the jet tube pressure state sequence.

[0031] In this step, the sequence length (which can be adjusted empirically) is defined as The specific process is: use the Viterbi algorithm to decode the implicit state and find the most likely pressure state sequence , that is, maximize ,in, For the The specific process can be divided into four steps: initialization, recursion, termination and backtracking.

[0032] initialization: .

[0033] Recursion: ; .in, ; Record from the start time to the current time, and the current state is The maximum probability path when ; Record the previous state Transfer to current state To obtain the maximum probability.

[0034] Termination status ( = ): .

[0035] Backtracking = ~1): Path restoration of the maximum possible jet tube pressure state sequence ; .

[0036] The π, A, and B in the above process formula are the parameters of the HMM model finally obtained after iteration based on the training data in step 3. Among them, when the HMM model is trained with a set of input feature sequence data, the three parameters π, A, and B are fixed and remain unchanged when estimating the pressure state, but the HMM model can be retrained with new input feature sequence data, so it can be changed after retraining.

[0037] Step 5: During the online detection process, the jet tube pressure state sequence within a time window is obtained according to the HMM model The pressure abnormality is judged based on the value change and an early warning is output.

[0038] This step can divide the judgment of early warning into low pressure and pressure fluctuation. Among them, the change rule of the jet tube pressure state sequence corresponding to low pressure is: if within the current time window The value of has dropped twice in a row, and the subsequent pressure state has continued to be at a lower level or below, triggering a low pressure alarm, and the corresponding jet pressure pipe needs to be repaired. The change rule of the jet pipe pressure state sequence corresponding to the pressure fluctuation is: if within the current time window The value fluctuates repeatedly between two non-adjacent states, triggering a pressure fluctuation alarm, which requires systematic investigation.

[0039] A specific application scenario based on the above technical core is as follows Figure 1 As shown, the meaning of each index is: 1 represents the front visual detection device at the cleaning inlet end, 2 represents the rear visual detection device at the cleaning outlet end, 3 represents the high-pressure jet tube, 4 represents the high-pressure jet tube pressure warning system, 5 represents the conveying roller, 6 represents the steel plate before cleaning, 7 represents the steel plate after being cleaned by the high-pressure abrasive water jet, 8 represents the cleaning position in the high-pressure water jet surface cleaning device, and only the jet tube in position No. 1 is drawn in the figure, but the jet tubes in other positions have the same arrangement principle and method as those in position No. 1.

[0040] Figure 2 In the figure, area 1 to area 4 represent four strip areas virtually divided along the conveying direction of the steel plate, indicating that each area consists of Figure 2 The corresponding cleaning positions in the cleaning are responsible for cleaning. A set of cleaning nozzles is installed in each cleaning position, and each set of jet nozzles is connected to a set of high-pressure jet pipes, and each set of high-pressure jet pipes is connected to a high-pressure pump. The total width of the steel plate is 2 meters, and the cleaning of the entire area can be completed only after the steel plate passes through 4 cleaning positions.

[0041] The data set in step 1 above can be constructed by: keeping the set steel plate conveying speed unchanged, controlling the operating pressure of the high-pressure pump determined by experience between 40 and 50 MPa, dividing the pressure of the high-pressure pump into 20 gears, each gear is 0.5 MPa, and a total of 20 gears, starting from 40 MPa, cleaning 10 steel plates in each gear until the pressure of the high-pressure pump exceeds 50 MPa, using Figure 1 The front and rear cameras in the system shoot the surface of the steel plate, and a total of 400 steel plate images are obtained. In addition, relevant image data can be retrieved from the historical database to further increase the image sample.

[0042] Then the images before and after cleaning in the sample are processed according to Figure 2As shown, the image is divided into four equidistant strip area sub-images along the conveying direction of the steel plate. Based on experience, the quality of each sub-image is manually annotated, and the annotation result is a quality grade label. For example, a 7-category label set Q = {very poor quality, poor quality, relatively poor quality, average, good quality, good quality, very good quality} can be used to evaluate the surface quality of each sub-image. The data samples are divided into a data set D1 consisting of sub-images before and after cleaning and their corresponding quality grade labels.

[0043] Furthermore, the EfficientNet network architecture can be used to train the surface cleaning quality evaluation model. Through the model module of the keras library of the deep learning framework TensorFlow, the EfficientNet model is loaded, the network output layer is constructed through the Dense interface, and the hyperparameters and model training are set through the fit interface of the model module, and the model performance is verified. If the verification reaches the predetermined indicators, the surface cleaning quality evaluation model constructed by EfficientNet is exported as a deployable model in ONNX format.

[0044] Afterwards, the data set D1 in the embodiment is further processed to construct a data set D2= .

[0045] Through the data set D2= The three key parameters of the HMM model for jet tube pressure estimation are estimated by using the unsupervised learning method, including the initial state distribution vector π, the state transition probability matrix A and the observation probability distribution B. Normalize to get the input feature input sequence represented by real values ,For example When the evaluation value of is "poor quality", the normalized result is , randomly initialize π, A, B, set the maximum number of iterations and convergence threshold, and then use the Baum-Welch algorithm to estimate the parameters π, A, B of the HMM model by maximizing the likelihood function of the observation sequence. The parameter estimation process can create an HMM model through the GaussianHMM function provided by the hmmlearn library and train it using the model's fit function.

[0046] against Figure 1 The four groups of high-pressure jet tubes in the cleaning device are used to collect observation time series data. Assuming that the length of the collected time window is defined as 10, the collected data of each group of jet tubes is ,in , , They are The surface quality ratings and quality changes before and after cleaning corresponding to the machine positions are input into the trained HMM model to obtain four sets of pressure estimation data sequences of the jet tube. , each of which The value of is one of the five pressures: low, relatively low, average, relatively high, and high. The process can be obtained through the predict function of the trained HMM model.

[0047] In this way, the window of data observation can be slid to continuously obtain the pressure state of the jet tube. , realize the online estimation of the pressure of four sets of jet tubes. For example, in a specific solution, the pressure states of the four sets of jet tubes are obtained as follows: Group 1: .

[0048] Group 2: .

[0049] Group 3: .

[0050] Group 4: .

[0051] In the third group of jet tubes, the pressure dropped twice in succession and continued to remain in a lower pressure range. It can be analyzed that there may be a low pressure failure, so a low pressure warning is issued for the third group of jet tubes.

[0052] Example 2 This embodiment discloses a high-pressure jet pipeline pressure warning system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a similar method of the above embodiment is implemented. The core content can be summarized as follows: Figure 3 The following steps are shown: Step S1, training, verifying and testing a surface quality rating model based on the image data of the target surface before and after cleaning at different jet pressure levels in the data set. Optionally, the target includes but is not limited to the steel plate described in the above embodiment. If the same steel plate corresponds to at least two camera positions during the cleaning process, the data set includes the image data of the area corresponding to each camera position before and after cleaning.

[0053] Step S2: construct an observation sequence based on the rating data before and after cleaning obtained based on the trained surface quality rating model for any cleaning area on the target surface and the pressure data of the high-pressure pump during the cleaning process.

[0054] Preferably, the time series data formed by the corresponding number of observation sequences The form is: ;in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, Indicates the cleaning effect. represents the set pressure of the high-pressure pump, T represents the number of data samples, represents the time index of the corresponding sample in the time series data, and .

[0055] Step S3: For a corresponding number of observation sequences, estimate the initial state distribution vector, state transition probability matrix and observation probability distribution of the jet tube pressure estimation HMM model based on an unsupervised learning method.

[0056] Preferably, the unsupervised learning method of this step adopts the Baum-Welch algorithm; the jet tube pressure state includes five states of low pressure, relatively low pressure, average pressure, relatively high pressure and high pressure.

[0057] Step S4, decoding the implicit state of the real-time observation sequence within a time window according to the initial state distribution vector, state transition probability matrix and observation probability distribution of the HMM (Hidden Markov Model) model to obtain the jet tube pressure state sequence, and then judging whether the pressure is abnormal according to the value change of the jet tube pressure state sequence, if so, executing early warning processing.

[0058] Preferably, the algorithm for decoding the implicit state in this step adopts the Viterbi algorithm. Further, this step performs synchronous online detection of the real-time operating state of each machine position with parallel threads.

[0059] like Figure 4 As shown, a specific high-pressure jet pipeline pressure warning system includes: Data acquisition subsystem: includes a camera device, an image acquisition module, a historical database module and a data processing module; wherein the image acquisition module is used to collect real-time image data of the steel plate surface before and after cleaning, the historical database module is used to store the steel plate surface images before and after cleaning and the high-pressure pump pressure data, and the data processing module is used to pre-process the images of the steel plate before and after cleaning.

[0060] Steel plate surface quality evaluation subsystem: includes a surface image dataset and a surface quality evaluation network. The surface image dataset consists of sub-images before and after cleaning, which are divided according to the steel plate conveying direction and cleaning position, and the quality labels corresponding to the images. The surface quality evaluation network is constructed by a network based on a deep learning architecture. The network is trained and evaluated using the dataset. Once the requirements are met, a surface quality evaluation model is generated.

[0061] Jet tube pressure interval estimation subsystem: used to train the pressure estimation model based on the time series data set. The time series data of the time series data set is ,in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, indicates the cleaning effect, and , Indicates the set pressure of the high pressure pump. Represents the number of data samples. The Baum-Welch algorithm is used to train the pressure estimation model. The initial state distribution vector π, state transition probability matrix A and observation probability distribution B of the HMM model are estimated by maximizing the likelihood function of the observation sequence, and finally the HMM model for jet tube pressure state estimation is determined.

[0062] The jet tube pressure warning subsystem for time series data collection and real-time pressure estimation: including time series data collection and real-time pressure estimation modules, and selecting the appropriate time window length according to the equipment operation experience , according to the window length and time sequence, the real-time observation time series data about the cleaning quality is formed , and then decode the implicit pressure state according to the Viterbi algorithm to obtain the pressure state sequence ,according to According to the changes in the medium pressure state, low pressure or pressure fluctuation alarm will be issued.

[0063] In summary, the high-pressure jet pipeline pressure warning method and system disclosed in the above embodiments of the present invention respectively have at least the following beneficial effects: 1. According to the changes in the surface quality of the steel plate and other targets before and after cleaning, the pressure of the jet tube can be monitored according to the cleaning stations in the cleaning device, which is helpful to achieve rapid early warning and fault location.

[0064] 2. It overcomes the problem that the high-pressure jet tube cannot be installed with an online pressure detection instrument, and realizes the online real-time estimation of the pressure of the high-pressure jet tube, which is conducive to further ensuring the cleaning quality.

[0065] 3. In-depth analysis based on the calculated changes in the pressure state of the jet tube is conducive to further evaluating the operating status of the high-pressure water jet surface cleaning device; for example, it can provide early warnings of low pressure or pressure fluctuations based on the changes in the decoded implicit pressure state.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A high-pressure jet pipeline pressure early warning method, characterized in that: include: Step S1, training, verifying and testing a surface quality rating model based on image data of the target surface before and after cleaning at different jet pressure levels in the data set; Step S2, constructing an observation sequence based on the rating data before and after cleaning obtained from any cleaning area on the surface of the target based on the trained surface quality rating model and combining it with the pressure data of the high-pressure pump during the cleaning process; Step S3, for a corresponding number of observation sequences, estimating the initial state distribution vector, state transition probability matrix and observation probability distribution of the jet tube pressure estimation HMM model based on an unsupervised learning method; Step S4: decode the implicit state of the real-time observation sequence within a time window according to the initial state distribution vector, state transition probability matrix and observation probability distribution of the HMM model to obtain the jet tube pressure state sequence, and then judge whether the pressure is abnormal according to the value change of the jet tube pressure state sequence. If so, perform early warning processing.

2. The high-pressure jet pipeline pressure early warning method according to claim 1 is characterized in that: The corresponding number of observation sequences forms the time series data The form is: ;in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, Indicates the cleaning effect. represents the set pressure of the high-pressure pump, T represents the number of data samples, represents the time index of the corresponding sample in the time series data, and .

3. The high-pressure jet pipeline pressure early warning method according to claim 1 is characterized in that: The unsupervised learning method in step S3 adopts the Baum-Welch algorithm.

4. The high-pressure jet pipeline pressure early warning method according to claim 1 is characterized in that: The algorithm for decoding the implicit state in step S4 adopts the Viterbi algorithm.

5. The high-pressure jet pipeline pressure early warning method according to any one of claims 1 to 4, characterized in that: The target is a steel plate, and the pressure states of the jet tube include five states: low pressure, relatively low, average, relatively high and high.

6. The high-pressure jet pipeline pressure early warning method according to any one of claims 1 to 4, characterized in that: The same steel plate corresponds to at least two machine positions during the cleaning process. The data set in step S1 includes image data of the area corresponding to each machine position before and after cleaning. In step S4, the real-time operating status of each machine position is synchronously detected online with parallel threads.

7. A high-pressure jet pipeline pressure warning system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

8. The high-pressure jet pipeline pressure warning system according to claim 7 is characterized in that: Specifically include: Data acquisition subsystem: including a camera device, an image acquisition module, a historical database module and a data processing module; wherein the image acquisition module is used to collect real-time image data of the steel plate surface before and after cleaning, the historical database module is used to store the steel plate surface images before and after cleaning and the high-pressure pump pressure data, and the data processing module is used to pre-process the images of the steel plate before and after cleaning; Steel plate surface quality evaluation subsystem: It includes a surface image dataset and a surface quality evaluation network. The surface image dataset consists of sub-images before and after cleaning divided according to the steel plate conveying direction and cleaning position, and the quality labels corresponding to the images. The surface quality evaluation network is constructed by a network based on a deep learning architecture. The network is trained and evaluated using the dataset. Once the requirements are met, a surface quality evaluation model is generated. Jet tube pressure interval estimation subsystem: used to train the pressure estimation model based on the time series data set. The time series data of the time series data set is ,in, is the surface quality rating before cleaning, is the surface quality rating after cleaning, indicates the cleaning effect, and , Indicates the set pressure of the high pressure pump. represents the number of data samples, Indicates the time index of the corresponding sample in the time series data. The pressure estimation model training adopts the Baum-Welch algorithm. By maximizing the likelihood function of the observation sequence, the three key parameters of the HMM model, namely the initial state distribution vector π, the state transition probability matrix A and the observation probability distribution B, are estimated. Finally, the HMM model of the jet tube pressure state estimation is determined. The jet tube pressure warning subsystem for time series data collection and real-time pressure estimation: including time series data collection and real-time pressure estimation modules, and selecting the appropriate time window length according to the equipment operation experience , according to the window length and time sequence, the real-time observation time series data about the cleaning quality is formed , and then decode the implicit pressure state according to the Viterbi algorithm to obtain the pressure state sequence ,according to According to the changes in the medium pressure state, low pressure or pressure fluctuation alarm will be issued.

Images