A method and system for evaluating the health state of a deluge valve based on a pressure-flow characteristic map

Abstract

The present disclosure provides a rain valve health state evaluation method and system based on pressure flow feature map. To solve the problem that the prior art lacks effective evaluation of the health state of the rain valve. Its characteristics are that the system includes a pressure sensor, a flow sensor, a health state evaluation system and a display device, the pressure sensor is used to collect the pressure of the rain valve control cavity and the control pipeline, the flow sensor is used to collect the flow in the pipeline connected with the rain valve water outlet cavity, the health state evaluation system draws a characteristic map of the change rate of pressure and flow according to the pressure and flow test values, then identifies and classifies the characteristic map based on artificial intelligence method, diagnoses and evaluates the state of the rain valve, and displays the results on the display device. The overall state of the rain valve can be evaluated to give the final evaluation result, which is convenient for maintenance personnel to locate the fault rain valve, determine the repair range and reduce the cost of fault analysis.

Description

Technical fields:

[0001] This invention relates to a method and apparatus for assessing the health status of deluge valves in fire sprinkler systems, used to assess the health status of deluge valves and determine the fault type of deluge valves. Background technology:

[0002] In fire suppression systems, deluge valves are key components connecting automatic sprinkler systems and piping networks. They are typically referred to as the downstream piping of the outlet chamber. Normally, the downstream piping of the outlet chamber is dry. Only after a fire occurs, when fire detectors or smoke detectors detect signs of fire, do they activate the solenoid valve at the deluge valve's control end. This reduces the pressure in the control chamber, causing the deluge valve's diaphragm to open under the pressure of the inlet chamber. Water rapidly fills the downstream piping of the outlet chamber, and under pressure, it is quickly sprayed from each sprinkler head, achieving automatic fire suppression. Because deluge valves are often not opened, they frequently jam or fail to open completely, resulting in poor fire suppression effectiveness. Therefore, regular monitoring and evaluation of deluge valves are necessary. According to the national standard GB5135.5-2018, deluge valves should be inspected and evaluated at least quarterly.

[0003] Currently, some technicians have proposed methods for assessing the health status of certain valves, such as the technical solution disclosed in CN110068756B. However, because each valve has different operating environments and application characteristics, this assessment system is not applicable to deluge valves, and there is currently no efficient technical solution for assessing the health status of deluge valves. Summary of the Invention:

[0004] To address the problems mentioned in the background art, this specification provides a method and system for assessing the health status of a deluge valve based on a pressure-flow characteristic map, and provides specific embodiments. The solutions provided in these embodiments illustrate that, if this application is applied, it can perform functional detection and assessment of the health status of a deluge valve, identify its fault types, acquire parameters of the deluge valve during operation through sensors, and combine machine learning and image recognition methods to effectively assess the state of the deluge valve, accurately locate deluge valve faults, and remind maintenance personnel to perform repairs or replacements.

[0005] The technical solutions disclosed in this specification are as follows:

[0006] Technical Solution 1: First, a method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram is presented, characterized in that the assessment method includes the following steps:

[0007] a1: Establish an artificial intelligence recognition model based on machine learning, collect signals of various deluge valves under normal and fault conditions as rated data and establish pressure-flow characteristic curves, and use the pressure-flow characteristic curves to establish an artificial intelligence recognition model through a deep learning neural network; the signals include the pressure signal of the deluge valve control chamber and the flow signal of the deluge valve outlet chamber.

[0008] a2: The solenoid valve is used to control the deluge valve to perform the valve opening action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow change curve to be evaluated.

[0009] a3: After the flow rate of the deluge valve stabilizes, the solenoid valve is used to control the deluge valve to perform the valve closing action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow rate change curve to be evaluated.

[0010] a4: Perform data analysis and processing on the curves generated in steps a2 and a3, and plot the flow-pressure characteristic curve;

[0011] The data analysis and processing path is as follows: normalize the test data obtained through steps a2 and a3 with the rated data obtained through step a1, convert the pressure and flow data into rates of change, and then use the entire pressure curve data as the horizontal axis and the flow curve data as the vertical axis to plot the pressure-flow characteristic curve. The rate of change of the pressure curve starts from 1 and returns to 1, and the rate of change of the flow curve starts from 0 and returns to 0. The pressure-flow characteristic curve is a closed curve.

[0012] a5: Input the pressure-flow characteristic curve obtained in step a4 into the artificial intelligence recognition model established in step a1, and use a deep learning neural network to identify the shape of the pressure-flow characteristic curve and give a health or fault status judgment; if it is judged to be a fault, give the fault type; if it is judged to be an unknown fault, it is judged to be a new fault type that needs to be added to the classification feature, and this unknown fault and pressure-flow curve are labeled and put into the training library of the artificial intelligence recognition model, so that the feature library can be trained again when machine learning is performed. Repeat the machine learning process until a new artificial intelligence recognition model is established.

[0013] Further, in step a1: the acquisition of signals from various deluge valves under normal and fault conditions as rated data and the establishment of pressure-flow characteristic curves include establishing a library of pressure and flow characteristic curves of deluge valves under normal and fault conditions.

[0014] The artificial intelligence recognition model established via deep learning neural network includes using a deep learning neural network-based machine learning method based on the pressure and flow feature library of the deluge valve, performing machine learning according to the library labels, and establishing a sample classification recognition model; the model is stored in a storage medium controlled by a central control unit; the training process of establishing the sample classification recognition model by machine learning is executed at least once.

[0015] Furthermore, in steps a2 and a3, the collected pressure change data and flow change data need to be collected for at least one cycle, including one valve opening process, one flow stabilization process and one valve closing process, and the sampling frequency of all data must be consistent.

[0016] In step a4, when normalizing the collected pressure and flow rate change data, the rated values ​​of the deluge valve pressure and flow rate determined in step a1 are used as the basis for normalization. Finally, the pressure and flow rate change curves are obtained. When the valve is opened, the pressure decreases from the rated pressure to the flow pressure, and the flow rate increases from 0 to the rated flow rate. When the valve is closed, the pressure returns to the rated pressure, and the flow rate returns to 0. During the normalization process, the maximum value of the measured pressure and flow rate is used as the denominator, and other pressure and flow rate test values ​​are used as the numerator. The maximum value of the final pressure and flow rate sequence is 1, and the minimum value is 0.

[0017] In step a5, the shape of the pressure-flow characteristic curve is identified using a deep learning neural network. A confidence level of 0.7 is used as the dividing line between whether the result is usable. If the confidence level is greater than or equal to 0.7, the identification result is the diagnostic result. If the confidence level is less than 0.7, manual intervention is performed to correct the identification result.

[0018] In step a4, before normalizing the collected pressure and flow rate change data, the following conditions must be met:

[0019] Read the rated pressure and flow rate of the deluge valve to be evaluated;

[0020] Determine whether the actual pressure of the deluge valve measured by the pressure sensor is within the allowable error range of the rated pressure. If it is not within the allowable error range, refuse to perform the diagnosis and indicate the reason.

[0021] In step a5, if the confidence level of the current feature image is low (confidence level < 0.7), it indicates poor image feature matching and some doubt about the correctness of the recognition. In this case, manual intervention is required, the detection needs to be performed again, and the current pressure and flow feature map needs to be stored and its type identified. When the number of stored pressure and flow feature maps reaches the minimum requirement for model correction, model correction is performed. If the confidence level is high (confidence level ≥ 0.7), it indicates that the recognition result is normal and the result is adopted.

[0022] Set a confidence level of 0.7 as 60 points, a confidence level of 1.0 as 100 points, and set the scores proportionally for confidence levels between 0.7 and 1.0. If the confidence level is less than 0.7, it is set as unqualified.

[0023] Technical Solution 2, based on Technical Solution 1, provides the following specific path for establishing an artificial intelligence recognition model via a deep learning neural network:

[0024] (1) Prepare learning data. First, prepare pressure and flow data of the deluge valve under normal and various fault conditions, and mark the deluge valve state represented by each set of data. Data needs to be prepared for each state, and theoretically, each set of data should have no less than 1,000 sets. When the data is insufficient, the existing graphics can be scaled, rotated and grayscale changed to make up for it.

[0025] (2) Construct a deep learning network. Construct a 6-layer convolutional neural network with 3 convolutional sub-layers (C1, C2, C3) and 3 pooling layers (S2, S4, S6). The convolutional sub-layers and pooling layers of the convolutional neural network are interleaved and their order is (C1, S2, C3, S4, C5, S6) and a 3-layer fully connected deep learning neural network (F7, F8, F9). The deep learning neural network has a total of 7 layers, not counting the input and output.

[0026] The input signal to the deep learning neural network is the pressure-flow characteristic curve, and the final output of the network is the deluge valve status, including 1 normal status and 5 fault statuses.

[0027] The purpose of the convolutional sub-layer is to extract feature information from the image. The effect of convolution is determined by the convolution kernel. Multiple 3*3 convolution kernels are used to perform operations on the input. The convolution operation is used to abstract the image features into higher-order features. The activation function of the convolutional layer is the ReLU function, which is used to simplify the calculation process and avoid gradient explosion and gradient scrambling.

[0028] The pooling layer is used to reduce dimensionality while preserving image features;

[0029] The convolutional sub-layers and pooling layers are interleaved to improve the feature extraction efficiency of the input image;

[0030] The three-layer fully connected network is used to map the learned features to the feature representation space, classify them according to the features, and finally output the rain valve status.

[0031] (3) Processing of pressure and flow characteristic curve images: After measurement, based on the measured pressure and flow data during the operation of the deluge valve, a test characteristic map is generated. First, check whether the maximum pressure and flow meet the rated requirements. Then, normalize the test data by dividing each data by the maximum test value so that all data are between [0, 1]. Plot the pressure and flow characteristic map with pressure as the horizontal axis and flow as the vertical axis. The final plotted curve is a closed curve. Ideally, the pressure and flow characteristic map of the deluge valve in normal state is a square with a side length of 1. Due to measurement errors, the actual shape of the map will be partially deformed. Finally, scale the map to 92*92 pixels and put it in a specific folder for later use. Each folder stores the characteristic map of one state. These characteristic maps will eventually be used as input to the deep learning neural network.

[0032] (4) Training sample preparation.

[0033] Take 70% of the pressure flow feature map prepared in step (3) and the corresponding state labels as input to the deep learning neural network, extract the state coverage of the pressure flow feature map, requiring that at least 70% of the feature map of each state be extracted, use the remaining 30% of the images as the test set, and then perform the following operations:

[0034] 1) Initialize the network, including intermediate layers, connections, and biases;

[0035] 2) Input samples and calculate the output of each layer;

[0036] 3) Calculate the network error and the error signals of each layer;

[0037] 4) Check if the network error is less than the required error. If it is less, proceed to step 6) for testing.

[0038] 5) Calculate the error of neurons in each layer based on the error, obtain the error gradient, update and adjust the weights of each layer, and go to step 2) until the error gradient meets the requirements.

[0039] 6) Fix all weights and thresholds to obtain the model network. Save the model as a file for easy access when needed.

[0040] 7) Feed all test images into the model network, check the output status, and check the result error. If the result error meets the requirements, the training is complete; otherwise, go to step 1) to retrain. The system uses gradient descent to adjust the weight parameters.

[0041] The final AI recognition model is one that meets the requirements. The model stores the structure and parameters of the entire network and can be used directly in the subsequent implementation phase.

[0042] Furthermore, in step a1, when constructing the deep learning network, for the convolutional sub-layers (C1, C3, C5), let the number of convolutional kernels for C1, C2, and C3 be d1, d2, and d3 respectively. Then, the output of the convolutional sub-layer operation is represented as:

[0043] in:

[0044]

[0045] j = 1, 2, ..., d1 / d2 / d3 (d1 is the number of filters in layer C1, d2 is the number of filters in layer C3, and d3 is the number of filters in layer C5);

[0046] b ij It is network bias;

[0047] The summation operation is a convolution operation using a filter network. For C1, it is the convolution output of the image pixel values ​​and the filter.

[0048] For pooling layers (S2, S4, S6), with a fixed sampling size of 2*2 and a stride of 2, using max pooling, and assuming the input image size is (m, n), the output of the pooling operation is represented as: O ij (x, y) = max(x) i，j x i，j+1 x i+1，j x i+1，j+1 Where: i≤m-2; j≤n-2

[0049] Both fully connected layers F7 and F8 have 10 neurons, and their input is the output of S6. The outputs of S6 are concatenated to form a 1D sequence, and all nodes in the sequence are connected to all units of F7, thus forming a fully connected layer. F7 and F8 act as a classifier, classifying the extracted feature data into various rain valve states. The final output of F9 is obtained by performing a Softmax operation, as follows:

[0050]

[0051] The final result is the probability value of the i-th state.

[0052] Technical Solution 3: A deluge valve health status assessment system based on pressure-flow characteristic maps, comprising a central controller, an AD conversion circuit, a pressure sensor, a flow sensor, a human-machine interface, and a power supply system. This assessment system is applied to the health status assessment of deluge valves, and its unique feature is:

[0053] The evaluation system also includes:

[0054] The signal acquisition module is used to acquire the deluge valve pressure signal and flow signal from the pressure sensor and flow sensor respectively, and transmit them to the data storage module under the instruction of the central controller.

[0055] The solenoid valve control module is used to enable the deluge valve to open or close under the command of the central controller.

[0056] The data storage module is used to store or retrieve data under the instructions of the central controller;

[0057] The central controller has a built-in computer program that controls the signal acquisition module, solenoid valve control module, AD conversion circuit, pressure sensor, flow sensor, human-machine interface, data storage module, and power supply system. Under the control of the computer program, the central controller can perform the following functions:

[0058] a1: Establish an artificial intelligence recognition model based on machine learning, collect signals of various deluge valves under normal and fault conditions as rated data and establish pressure-flow characteristic curves, and use the pressure-flow characteristic curves to establish an artificial intelligence recognition model through a deep learning neural network; the signals include the pressure signal of the deluge valve control chamber and the flow signal of the deluge valve outlet chamber.

[0059] a2: The deluge valve is controlled by a solenoid valve to perform the valve opening action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the pressure value change curve and the outlet flow change curve of the deluge valve control pipeline to be evaluated are generated under the action of the built-in program of the central control unit.

[0060] a3: After the flow rate of the deluge valve stabilizes, the solenoid valve is used to control the deluge valve to perform the valve closing action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow rate change curve to be evaluated.

[0061] a4: Perform data analysis and processing on the curves generated in steps a2 and a3, and plot the flow-pressure characteristic curve;

[0062] The data analysis and processing path is as follows: normalize the test data obtained through steps a2 and a3 with the rated data obtained through step a1, convert the pressure and flow data into rates of change, and then use the entire pressure curve data as the horizontal axis and the flow curve data as the vertical axis to plot the pressure-flow characteristic curve. The rate of change of the pressure curve starts from 1 and returns to 1, and the rate of change of the flow curve starts from 0 and returns to 0. The pressure-flow characteristic curve is a closed curve.

[0063] a5: Input the pressure-flow characteristic curve obtained in step a4 into the artificial intelligence recognition model established in step a1, and use a deep learning neural network to identify the shape of the pressure-flow characteristic curve, giving a health or fault status judgment; if it is judged as a fault, give the fault type; if it is judged as an unknown fault, it is judged as a new fault type that needs to be added as classification features. Label this unknown fault and the pressure-flow curve, and put it into the training library of the artificial intelligence recognition model, so that the feature library can be trained again when machine learning is performed. Repeat the machine learning process until a new artificial intelligence recognition model is established.

[0064] The above-described at least one technical solution adopted in one or more embodiments of this specification can achieve the following:

[0065] Beneficial effects:

[0066] By analyzing the time correlation between pressure and flow, the measurement results are output as graphical images. The shape of the image represents various states of the deluge valve. Simultaneously, by recognizing the shape of the graphical image through machine learning and artificial intelligence, the location and type of fault in the deluge valve can be determined. The overall condition of the deluge valve can be evaluated, and a final evaluation result can be given. This makes it easier for maintenance personnel to locate faulty deluge valves, determine the scope of maintenance, reduce the cost of fault analysis, and realize remote unmanned measurement, effectively reducing maintenance costs. Attached image description:

[0067] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below: The drawings described below are only some of the embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a schematic diagram of the device installation of the present invention.

[0069] Figure 2 This is a schematic diagram of the device structure involved in this invention.

[0070] Figure 3 This is a pressure and flow characteristic diagram of the deluge valve under different states involved in this invention.

[0071] Figure 4 This is a diagram of a deep learning neural network model.

[0072] In the diagram: 1-Ultrasonic flow sensor, 2-Outlet chamber rear end pipe, 3-Outlet chamber, 4-Control chamber, 5-Inlet chamber, 6-Pressure sensor, 7-Control pipeline, 8-Solenoid valve, 9-Health status assessment system, 10-RS485 communication cable, 11-Ultrasonic signal converter, 12-Manual drain valve, 13-Closed diaphragm, 14-Water, 15-Battery, 16-Pressure measurement communication interface, 17-LCD display screen, 18-Touch screen, 19-Flow measurement communication interface, 20-Result output module, 21-Artificial intelligence recognition module, 22-Machine learning module. Detailed implementation method:

[0073] The specific embodiments given in this disclosure will be further described below with reference to the accompanying drawings:

[0074] The rain valve health status detection device of the present invention adopts the following technical solution:

[0075] The deluge valve health status monitoring device includes a power supply module, a pressure sensor, a flow sensor, a health status assessment system, and a touch-screen LCD display. The pressure sensor, flow sensor, and LCD display are all connected to the health status assessment system, and the power supply module supplies power to all modules.

[0076] The health status assessment system includes a data acquisition module, a data processing module, a data calculation and analysis module, and a fault location module, which are connected in sequence.

[0077] The data acquisition module can be connected to a pressure sensor and a flow sensor. The pressure sensor is installed on the control pipeline of the deluge valve, while the flow sensor is installed on the outer wall of the pipeline at the rear end of the deluge valve's outlet chamber.

[0078] The flow sensor is an ultrasonic flow meter, which is installed on the outer wall of the pipe by magnetic adsorption and does not affect the function of the pipe.

[0079] The pressure sensor and flow sensor use digital output and are connected to the data acquisition module of the health status assessment system via an RS485 bus.

[0080] The health status assessment process is divided into two phases: preparation and implementation. The preparation phase involves establishing an artificial intelligence recognition model based on machine learning. This requires collecting pressure characteristic curves of various deluge valves under various fault conditions and using these curves as data to build a recognition model through machine learning.

[0081] The preparation phase only needs to be implemented once, and the artificial intelligence recognition model can be used for a long time. If there are new pressure and flow characteristic curves that are not within the recognition range, the model can be retrained and updated.

[0082] The specific steps of the preparation stage are as follows:

[0083] (1) Prepare learning data. First, prepare pressure and flow data of the deluge valve under normal and various fault conditions, and mark the deluge valve state represented by each set of data. Data needs to be prepared for each state, and each set of data should theoretically have no less than 1000 sets. When the data is insufficient, techniques such as scaling, rotating, and changing the grayscale of existing graphics can be used to make up for it.

[0084] (2) Constructing a Deep Learning Network. A 6-layer convolutional neural network (CNN) with 3 convolutional sub-layers (C1, C2, C3) and 3 pooling layers (S2, S4, S6) is constructed. The convolutional and pooling layers are interleaved in the order (C1, S2, C3, S4, C5, S6), and a 3-layer fully connected deep learning neural network (F7, F8, F9) is also constructed. The network has a total of 7 layers, excluding input and output. The input signal of this network is a pressure-flow characteristic curve, and the final output of the network is the state of the deluge valve, including 1 normal state and 5 fault states. The purpose of the convolutional layers is to extract the feature information of the image. The effect of convolution is determined by the convolution kernel. Multiple 3*3 convolution kernels are used to operate on the input. The convolution operation can abstract the image features into higher-order features. The activation function of the convolutional layers is the ReLU function, which can effectively simplify the calculation process and avoid gradient explosion and gradient vanishing problems. The pooling layers can reduce the vida and preserve image features. The interleaved distribution of convolutional and pooling layers effectively addresses feature extraction from the input image. The final fully connected layer maps the learned features to a feature representation space, enabling classification based on these features and ultimately outputting the rain valve state.

[0085] I. For the convolutional sub-layers (C1, C3, C5), let the number of convolutional kernels for C1, C2, and C3 be d1, d2, and d3, respectively. Then the output of the convolutional sub-layer operation can be expressed as:

[0086] in:

[0087]

[0088] j = 1, 2, ..., d1 / d2 / d3 (d1 is the number of filters in layer C1, d2 is the number of filters in layer C3, and d3 is the number of filters in layer C5);

[0089] b ij It is network bias;

[0090] The summation operation (∑) utilizes the convolution operation of the filter network. For C1, it is the convolution output of the image pixel values ​​and the filter.

[0091] II. For pooling layers (S2, S4, S6), with a fixed sampling size of 2*2 and a stride of 2, using max pooling, and assuming the input image size is (m, n), the output of the pooling operation can be expressed as: O ij (x, y) = max(x) i，j x i，j+1 x i+1，j x i+1，j+1 Where: i≤m-2; j≤n-2

[0092] III. Fully connected layers F7 and F8 each have 10 neurons, with the output of S6 as their input. The outputs of S6 are concatenated to form a one-dimensional sequence. All nodes in this sequence are connected to all units of F7, thus forming a fully connected layer. F7 and F8 act as a classifier, classifying the extracted feature data into various rain valve states. The final output of F9 is calculated using a Softmax operation, as follows:

[0093]

[0094] The final result is the probability value of the i-th state.

[0095] (3) Processing of pressure-flow characteristic curve images. After measurement, a test characteristic map can be generated based on the measured pressure and flow data during the deluge valve's operation. First, check if the maximum pressure and flow meet the rated requirements. Then, normalize the test data by dividing each data point by the maximum test value, ensuring all data fall within the range [0, 1]. Plot the pressure-flow characteristic map with pressure on the x-axis and flow on the y-axis. The resulting curve is a closed curve. Ideally, the pressure-flow characteristic map of the deluge valve in normal operation is a square with a side length of 1. However, due to measurement errors, the actual shape may be partially distorted. Finally, scale the image to 92*92 pixels and place it in a specific folder for later use. Each folder stores the characteristic map for one state. These characteristic maps will ultimately serve as input to a deep learning neural network.

[0096] (4) Training sample preparation.

[0097] Take 70% of the pressure flow feature maps prepared in step (4) and their corresponding state labels as input to the deep learning neural network. Note the state coverage of the extracted feature maps, requiring that at least 70% of the feature maps for each state be extracted. Use the remaining 30% of the images as the test set:

[0098] 1) Initialize the network, including intermediate layers, connections, and biases;

[0099] 2) Input samples and calculate the output of each layer;

[0100] 3) Calculate the network error and the error signals of each layer;

[0101] 4) Check if the network error is less than the required error e. If it is less, proceed to step 6) for testing.

[0102] 5) Calculate the error of neurons in each layer based on the error, obtain the error gradient, update and adjust the weights of each layer, and go to step 2) until the error gradient meets the requirements.

[0103] 6) Fix all weights and thresholds to obtain the model network. Save the model as a file for easy access when needed.

[0104] 7) Feed all test images into the model network, check the output status, and check the result error. If the result error meets the requirements, the training is complete; otherwise, go to step 1) to retrain.

[0105] A reasonable method for adjusting weights can effectively reduce the number of training iterations, enabling the network to converge quickly and saving a significant amount of training time. The system uses gradient descent to adjust the weight parameters.

[0106] (5) The final network is an artificial intelligence recognition model that meets the requirements. The model stores the structure and parameters of the entire network, which can be used directly in the subsequent implementation stage.

[0107] The specific steps of the implementation phase are as follows:

[0108] (1) First, based on the model of the deluge valve to be tested, read the pre-stored basic data of the deluge valve, including size, material, rated control pressure and rated flow.

[0109] (2) Open the pressure relief valve of the control pipeline, open the deluge valve, and at the same time start recording the pressure change of the control pipeline and the flow change at the rear end of the outlet chamber;

[0110] (3) After the deluge valve is opened, maintain the flow rate for a period of time, and then perform the valve closing operation to close the pressure relief valve of the control pipeline until the flow rate of the outlet chamber is zero.

[0111] (4) Save the pressure change data and flow change data of the entire valve opening-holding-closing process, and normalize the data;

[0112] (5) Plot the pressure-flow characteristic curve with normalized pressure as the horizontal axis and flow rate as the vertical axis.

[0113] (6) Input the feature curve into the artificial intelligence recognition model. The model outputs the Softmax calculation result of the current deluge valve. The Softmax value is the confidence level of a certain state. Select the largest confidence level as the diagnosis result.

[0114] (7) If the confidence level is lower than 0.7, the status of the deluge valve needs to be manually checked to determine the fault type and cause. Then the characteristic curve is marked and stored in the storage area of ​​the data calculation and analysis module.

[0115] (8) When the number of low-confidence curves accumulates to a certain level, the machine learning process can be executed again to update the artificial intelligence recognition model, thereby further improving the accuracy of the evaluation. This iterative process represents the progress of machine learning and is a general processing method for artificial intelligence.

[0116] like Figure 1 , 2 As shown, the rain valve health status detection device of the present invention consists of an ultrasonic flow sensor 1, a pressure sensor 6, and a health status assessment system with an LCD screen. The health status assessment system includes a data acquisition module, a data processing module, a data calculation and analysis module, and a fault location module. The entire system is powered by a power supply module.

[0117] like Figure 1 As shown, during operation, pressure sensor 6 is installed in the deluge valve control pipeline 7, and its output signal enters the system through pressure measurement communication interface 16; ultrasonic flow sensor 1 is installed on the outer wall of the pipe 2 at the rear end of the outlet chamber, and the signal of ultrasonic flow sensor 1 enters the system's flow communication interface 19 after being processed by signal converter 11.

[0118] like Figure 2 As shown, after the condition assessment system reads signals from two sensors, the internal data processing module filters and cleans the data, normalizes it, and plots it as a pressure-flow characteristic map. This map is then sent to the artificial intelligence image recognition module 21 to complete data classification and recognition. The recognition result is the fault type and confidence level. Finally, the result is sent to the LCD display module 17 for display, completing the fault diagnosis and location.

[0119] The deluge valve evaluation system has an internal evaluation program, the workflow of which is as follows:

[0120] (1) After the software starts, it completes the initialization of the communication port and the LCD, and prepares to enter the measurement state.

[0121] (2) Select the model of the deluge valve. When selecting the model, the basic data of the corresponding deluge valve will be read from the database, including the rated control pressure and flow rate.

[0122] (3) Complete the self-test of the sensor to determine whether the two sensors are working properly. If they are not working properly, an alarm will be triggered and the system will exit.

[0123] (4) Check whether the pressure of the control pipeline is within the rated pressure allowable range;

[0124] (5) Wait for the user to click the start measurement button on the screen, and at the same time control the diaphragm chamber pressure relief valve to open;

[0125] (6) After the measurement starts, the software establishes two independent threads, each responsible for measuring one sensor data. The pressure thread measures the pressure data curve, and the flow thread measures the flow curve. The measurement always samples data according to a fixed sampling period.

[0126] (7) When the flow rate at the rear end of the deluge valve outlet chamber stabilizes, control the solenoid valve to close the deluge valve;

[0127] (8) Data acquisition is continuously performed during the deluge valve opening-stabilizing-closing process until the deluge valve is completely closed;

[0128] (9) Normalize the collected pressure curves and flow curves, and plot the pressure-flow characteristic curves;

[0129] (10) Input the curve into the artificial intelligence recognition model for recognition, and give the fault type and confidence level;

[0130] (11) If the confidence level is lower than 0.7, manual intervention is prompted, the image is manually labeled, and stored in the low confidence database according to the label;

[0131] The artificial intelligence recognition model in step (9) of the software program is established in advance through machine learning 22. During the model building process, a large number of pressure characteristic curves of different types of faults need to be prepared. The images are used as input and fed into machine learning networks such as convolutional neural networks to build the artificial intelligence recognition model. When the model is used, different pressure and flow characteristic curves are input to quickly obtain the artificial intelligence recognition results.

[0132] In step (9) of the software program, the artificial intelligence recognition model 21 can continuously improve and grow. When an unknown fault or characteristic curve is found during the state detection and evaluation process, this unknown fault or curve can be labeled and put into the training library. After accumulating for a period of time, the machine learning process is carried out again, so that the newly established model can cope with and identify new fault types. This is an iterative and repeated process. As the iteration proceeds, the recognition rate and confidence of artificial intelligence will gradually improve, and finally a perfect evaluation result will be achieved.

[0133] like Figure 3As shown, a certain model DN150 deluge valve has a rated pressure of 1.2 MPa, is started by a 24V solenoid valve remote control, has an opening time of 5 seconds, and a closing time of 15 seconds. The diagnostic process for this deluge valve is as follows.

[0134] (1) Before testing, you must first prepare an artificial intelligence recognition model that has been trained.

[0135] (2) Then install the flow sensor and pressure sensor, initialize the system, enter the measurement state, and then perform a valve opening action on the deluge valve once, observe the changes in flow and pressure. When the flow stabilizes, maintain this for at least 5 seconds, and then perform a valve closing action until the flow is 0, re-establishing the control pressure. Throughout the entire process, the changes in pressure and flow can be seen on the LCD screen. At the same time, the pressure and flow are stored in the system RAM memory.

[0136] (3) Normalize the data, using the maximum value in the data as the denominator and the other data as the numerator, and calculate the normalized flow and pressure data respectively.

[0137] (4) Plot a pressure-flow characteristic graph using pressure as the x-axis and flow rate as the y-axis. This characteristic graph will show specific closed curves. Different curves represent different states; characteristic graphs for common states are shown below. Figure 3 As shown. Here, it is assumed that the measured feature map is... Figure 3 (b) has a similar style.

[0138] (5) Save the feature map as a black and white image and reduce the image resolution to 92*92 pixels. Then send the feature map into the artificial intelligence recognition network prepared in step (1). Since the input image is a black and white image, it can be assumed that the image has only one channel. Use 6 3*3 convolution kernels for feature extraction in layer C1 and activate it with the ReLU function to obtain 6 90*90 feature results. Enter layer S2 for 2*2 stride pooling dimensionality reduction to obtain 6 45*45 feature images. In the same way, use 10 3*3 convolution kernels in layer C3 to obtain 10 44*44 feature images. After pooling in layer S4, obtain 10 22*22 feature images. Use 16 3*3 convolution kernels in layer C5 to obtain 16 20*20 feature images. After 2*2 pooling dimensionality reduction in layer S6, the image size is 16 10*10 images. These graphs are linearized and then input into fully connected layers F7 and F8, each with 10 neurons. Finally, layer F9, with 7 neurons, outputs the probabilities of the seven states of the rain valve. (The last sentence appears to be incomplete and possibly refers to a different topic.) Figure 3 (b) If the patterns are similar, the neurons in layer F9 representing slow valve closing will have the largest output values, while the output values ​​of other neurons will be relatively small. This probability can be referred to as the confidence level of this state.

[0139] (6) If the confidence level of a certain state is much higher than the probability of other states, the recognition is successful. If the confidence level of all states is low, or is distributed in a few states, the recognition result is inaccurate.

[0140] (7) If the identification is inaccurate, it is necessary to re-evaluate and manually determine the state of the rain valve. If a new state is determined after manual intervention, a new category needs to be created in the training set and the data is saved in the training set. After a certain number of new state data are accumulated, the model can be retrained so that the system can identify the new state.

[0141] The above description is merely a preferred embodiment of the present invention and is not limited to the identification and diagnosis of these seven states. Any modifications or substitutions made within the spirit and principles of the invention are included within the protection scope of the present invention.

Claims

1. A method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram, characterized in that, The evaluation method includes the following steps: a1: Establish an artificial intelligence recognition model based on machine learning, collect signals of various deluge valves under normal and fault conditions as rated data and establish pressure-flow characteristic curves, and use the pressure-flow characteristic curves to establish an artificial intelligence recognition model through a deep learning neural network; the signals include the pressure signal of the deluge valve control chamber and the flow signal of the deluge valve outlet chamber. a2: The solenoid valve is used to control the deluge valve to perform the valve opening action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow change curve to be evaluated. a3: After the flow rate of the deluge valve stabilizes, the solenoid valve is used to control the deluge valve to perform the valve closing action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow rate change curve to be evaluated. a4: Perform data analysis and processing on the curves generated in steps a2 and a3, and plot the flow-pressure characteristic curve; The data analysis and processing path is as follows: normalize the test data obtained through steps a2 and a3 with the rated data obtained through step a1, convert the pressure and flow data into rates of change, and then use the entire pressure curve data as the horizontal axis and the flow curve data as the vertical axis to plot the pressure-flow characteristic curve. The rate of change of the pressure curve starts from 1 and returns to 1, and the rate of change of the flow curve starts from 0 and returns to 0. The pressure-flow characteristic curve is a closed curve. a5: Input the pressure-flow characteristic curve obtained in step a4 into the artificial intelligence recognition model established in step a1, and use a deep learning neural network to identify the shape of the pressure-flow characteristic curve and give a health or fault status judgment; if it is judged to be a fault, give the fault type; if it is judged to be an unknown fault, it is judged to be a new fault type that needs to be added to the classification feature, and this unknown fault and pressure-flow curve are labeled and put into the training library of the artificial intelligence recognition model, so that the feature library can be trained again when machine learning is performed. Repeat the machine learning process until a new artificial intelligence recognition model is established.

2. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 1, characterized in that, In step a1: the collection of signals from various deluge valves under normal and fault conditions as rated data and the establishment of pressure-flow characteristic curves include establishing a library of pressure and flow characteristic graphs of deluge valves under normal and fault conditions. The artificial intelligence recognition model established via deep learning neural network includes using a deep learning neural network-based machine learning method based on the pressure and flow feature library of the deluge valve, performing machine learning according to the library labels, and establishing a sample classification recognition model; the model is stored in a storage medium controlled by a central control unit; the training process of establishing the sample classification recognition model by machine learning is executed at least once.

3. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 2, characterized in that: In steps a2 and a3, the collected pressure change data and flow change data need to be collected for at least one cycle, including one valve opening process, one flow stabilization process and one valve closing process, and the sampling frequency of all data must be consistent. In step a4, when normalizing the collected pressure change data and flow change data, the rated values ​​of the deluge valve pressure and flow determined in step a1 are used as the basis for normalization. Finally, the pressure and flow change rate curves are obtained. When the valve is opened, the pressure decreases from the rated pressure to the flow pressure, and the flow increases from 0 to the rated flow. When the valve is closed, the pressure returns to the rated pressure, and the flow returns to 0. During normalization, the maximum value of the measured pressure and flow rate should be used as the denominator, and the other pressure and flow rate test values ​​should be used as the numerator. The maximum value of the final pressure and flow rate sequence is 1, and the minimum value is 0. In step a5, the shape of the pressure-flow characteristic curve is identified using a deep learning neural network. A confidence level of 0.7 is used as the dividing line between whether the result is usable. If the confidence level is greater than or equal to 0.7, the identification result is the diagnostic result. If the confidence level is less than 0.7, manual intervention is performed to correct the identification result.

4. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 3, characterized in that: In step a4, before normalizing the collected pressure and flow rate change data, the following conditions must be met: Read the rated pressure and flow rate of the deluge valve to be evaluated; Determine whether the actual pressure of the deluge valve measured by the pressure sensor is within the allowable error range of the rated pressure. If it is not within the allowable error range, refuse to perform the diagnosis and indicate the reason.

5. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 4, characterized in that: In step a5, if the confidence level of the current feature image is low (confidence level < 0.7), it indicates poor image feature matching and some doubt about the correctness of the recognition. In this case, manual intervention is required, the detection needs to be performed again, and the current pressure and flow feature map needs to be stored and its type identified. When the number of stored pressure and flow feature maps reaches the minimum requirement for model correction, model correction is performed. If the confidence level is high (confidence level ≥ 0.7), it indicates that the recognition result is normal and the result is adopted. Set a confidence level of 0.7 as 60 points, a confidence level of 1.0 as 100 points, and set the scores proportionally for confidence levels between 0.7 and 1.

0. If the confidence level is less than 0.7, it is set as unqualified.

6. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 5, characterized in that: In step a1, the establishment of the artificial intelligence recognition model via a deep learning neural network is carried out according to the following path: (1) Prepare learning data. First, prepare pressure and flow data of the deluge valve under normal and various fault conditions, and mark the deluge valve state represented by each set of data. Data needs to be prepared for each state, and the number of each set of data should not be less than 1000 sets. When the data is insufficient, the existing graphics should be scaled, rotated and grayscale changed to make up for it. (2) Construct a deep learning network. Construct a 6-layer convolutional neural network with 3 convolutional sub-layers (C1, C2, C3) and 3 pooling layers (S2, S4, S6). The convolutional sub-layers and pooling layers of the convolutional neural network are interleaved and their order is (C1, S2, C3, S4, C5, S6) and a 3-layer fully connected deep learning neural network (F7, F8, F9). The deep learning neural network has a total of 7 layers, not counting the input and output. The input signal to the deep learning neural network is the pressure-flow characteristic curve, and the final output of the network is the deluge valve status, including 1 normal status and 5 fault statuses. The purpose of the convolutional sublayer is to extract feature information from the image. The effect of convolution is determined by the convolution kernel, using multiple 3D convolutional kernels.

3. Convolution kernels perform operations on the input. Convolution operations are used to abstract the features of an image into higher-order features. The activation function of the convolutional layer is the ReLU function, which simplifies the calculation process and avoids gradient explosion and gradient saturation. The pooling layer is used to reduce dimensionality while preserving image features; The convolutional sub-layers and pooling layers are interleaved to improve the feature extraction efficiency of the input image; The three-layer fully connected network is used to map the learned features to the feature representation space, classify them according to the features, and finally output the rain valve status. (3) Processing of pressure-flow characteristic curve image: After measurement, based on the measured pressure and flow data during the operation of the deluge valve, a test characteristic map is generated. First, check whether the maximum pressure and flow meet the rated requirements. Then, normalize the test data by dividing each data by the maximum test value so that all data are between [0, 1]. Plot the pressure-flow characteristic map with pressure as the horizontal axis and flow as the vertical axis. The final plotted curve is a closed curve. Ideally, the pressure-flow characteristic map of the deluge valve under normal conditions is a square with a side length of 1. Due to measurement errors, the actual shape of the graph will be partially distorted. Finally, the graph is scaled to 92. Each 92-pixel image is placed in a specific folder for later use. Each folder stores a feature map of a certain state, and these feature maps will eventually be used as input to a deep learning neural network. (4) Training sample preparation; Take 70% of the pressure flow feature map prepared in step (3) and the corresponding state labels as input to the deep learning neural network, extract the state coverage of the pressure flow feature map, requiring that 70% of the feature map of each state be extracted, use the remaining 30% of the images as the test set, and then perform the following operations: 1) Initialize the network, including intermediate layers, connections, and biases; 2) Input samples and calculate the output of each layer; 3) Calculate the network error and the error signals of each layer; 4) Check if the network error is less than the required error. If it is less, proceed to step 6) for testing. 5) Calculate the error of neurons in each layer based on the error, obtain the error gradient, update and adjust the weights of each layer, and go to step 2) until the error gradient meets the requirements; 6) Fix all weights and thresholds to obtain the model network. Save the model as a file for easy access when needed. 7) Feed all test images into the model network, check the output status, and check the result error. If the result error meets the requirements, the training is complete; otherwise, go to step 1) to retrain. The system uses gradient descent to adjust the weight parameters. (5) The final artificial intelligence recognition model is an artificial intelligence recognition model that meets the requirements. The model stores the structure and parameters of the entire network, which can be used directly in the subsequent implementation stage.

7. The method for assessing the health status of a deluge valve based on a pressure-flow characteristic diagram according to claim 6, characterized in that: In step a1, when constructing the deep learning network, for the convolutional sub-layers ( C1,C3,C5 ),set up C1, C2, C3 The number of convolution kernels are respectively d1,d2,d3 The output of the convolutional sublayer operation is then represented as: in: ; , d1 yes C1 The number of filters in the layers, d2 yes C3 The number of filters in the layers, d3 yes C5 The number of layer filters; It is network bias; The summation operation utilizes the convolution operation of the filter network. C1 In short, it is the convolution output of the image pixel values ​​and the filtered data; For pooling layers (S2,S4,S6) Fixed sampling size is 2 2. With a step size of 2, using the max pooling method, let the size of the input graphic be... (m, n) The output of the pooling operation is then represented as: ,in , ; Both fully connected layers F7 and F8 have 10 neurons, and their input is the output of S6. The outputs of S6 are concatenated to form a 1D sequence, and all nodes in the sequence are connected to all units of F7, thus forming a fully connected layer. F7 and F8 act as a classifier, classifying the extracted feature data into various rain valve states. The final output of F9 is obtained by performing a Softmax operation, as follows: The final result is the probability value of the i-th state.

8. A deluge valve health status assessment system based on pressure-flow characteristic maps, comprising a central controller, an AD conversion circuit, a pressure sensor, a flow sensor, a human-machine interface, and a power supply system, wherein the assessment system is applied to the health status assessment of deluge valves, characterized in that: The evaluation system also includes: The signal acquisition module is used to acquire the deluge valve pressure signal and flow signal from the pressure sensor and flow sensor respectively, and transmit them to the data storage module under the instruction of the central controller. The solenoid valve control module is used to enable the deluge valve to open or close under the command of the central controller. The data storage module is used to store or retrieve data under the instructions of the central controller; The central controller has a built-in computer program that controls the signal acquisition module, solenoid valve control module, AD conversion circuit, pressure sensor, flow sensor, human-machine interface, data storage module, and power supply system. Under the control of the computer program, the central controller can perform the following functions: a1: Establish an artificial intelligence recognition model based on machine learning, collect signals of various deluge valves under normal and fault conditions as rated data and establish pressure-flow characteristic curves, and use the pressure-flow characteristic curves to establish an artificial intelligence recognition model through a deep learning neural network; the signals include the pressure signal of the deluge valve control chamber and the flow signal of the deluge valve outlet chamber. a2: The solenoid valve is used to control the deluge valve to perform the valve opening action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow change curve to be evaluated. a3: After the flow rate of the deluge valve stabilizes, the solenoid valve is used to control the deluge valve to perform the valve closing action. At the same time, after obtaining the detection data through the pressure sensor in the deluge valve control chamber and the flow sensor in the deluge valve outlet chamber, the built-in program of the central control unit generates the pressure value change curve of the deluge valve control pipeline and the outlet flow rate change curve to be evaluated. a4: Perform data analysis and processing on the curves generated in steps a2 and a3, and plot the flow-pressure characteristic curve; The data analysis and processing path is as follows: normalize the test data obtained through steps a2 and a3 with the rated data obtained through step a1, convert the pressure and flow data into rates of change, and then use the entire pressure curve data as the horizontal axis and the flow curve data as the vertical axis to plot the pressure-flow characteristic curve. The rate of change of the pressure curve starts from 1 and returns to 1, and the rate of change of the flow curve starts from 0 and returns to 0. The pressure-flow characteristic curve is a closed curve. a5: Input the pressure-flow characteristic curve obtained in step a4 into the artificial intelligence recognition model established in step a1, and use a deep learning neural network to identify the shape of the pressure-flow characteristic curve and give a health or fault status judgment; if it is judged to be a fault, give the fault type; if it is judged to be an unknown fault, it is judged to be a new fault type that needs to be added to the classification feature, and this unknown fault and pressure-flow curve are labeled and put into the training library of the artificial intelligence recognition model, so that the feature library can be trained again when machine learning is performed. Repeat the machine learning process until a new artificial intelligence recognition model is established.

Images