A self-sensing intelligent ball valve

Abstract

The present invention discloses a self-sensing intelligent ball valve, comprising: a ball valve body; a micro sensor array; a wireless communication module; a data processing unit; and an algorithm prediction module, including a fusion system based on a decision tree algorithm and a neural network algorithm, wherein the fusion system comprises: (a) a neural network feature extraction unit for receiving and processing the sensor data to generate a feature vector F NN (b) a decision tree classification unit for constructing decision rules based on the sensor data and the feature vectors generated by the neural network to predict ball valve seal wear and potential failures; (c) an adaptive weight fusion unit for dynamically adjusting the weight coefficients of the neural network and decision tree outputs based on model performance; and (d) an early warning generation unit for generating fault early warning information 7-14 days in advance based on the prediction results of the fusion system. The present invention has the following beneficial effects: the ball valve can monitor the working status in real time and predict potential failures.

Description

Technical Field

[0001] The present invention relates to the field of valve control technology, and in particular to a self-sensing intelligent ball valve that can monitor working status in real time and predict potential failures, and is suitable for industrial Internet of Things and remote management scenarios. Background Art

[0002] Ball valves are commonly used control components in industrial control systems, used to control the flow direction or volume of fluids. Failures in traditional ball valves, such as seal wear, valve body corrosion, and sticking, are often discovered only after they occur, leading to unexpected system downtime, increased maintenance costs, and reduced system reliability.

[0003] The existing technology mainly monitors the status of ball valves in the following ways:

[0004] 1. Regular manual inspection: Relying on manual periodic inspections makes it difficult to detect problems in real time;

[0005] 2. Simple sensor monitoring: Only monitoring a single parameter, such as temperature or pressure, cannot fully reflect the valve status;

[0006] Therefore, there is a need for a self-sensing intelligent ball valve system that can overcome the above-mentioned shortcomings and realize early prediction and prevention of faults.

[0007] In summary, the technical problem to be solved by the present invention is: how to design a self-sensing intelligent ball valve that can monitor the status of the ball valve in real time and predict potential faults in advance, overcome the shortcomings of the existing technology in fault diagnosis, and improve the accuracy and timeliness of fault prediction. Summary of the Invention

[0008] The present invention provides a self-sensing intelligent ball valve based on multi-sensor data fusion and intelligent algorithm, which can monitor the working status in real time and predict potential faults.

[0009] In order to solve the above technical problems, the present invention provides a self-sensing intelligent ball valve, comprising:

[0010] The ball valve body is used to control the switching and flow regulation of the fluid;

[0011] A micro sensor array is installed on the ball valve body to collect real-time data on valve opening, medium pressure, temperature and flow rate;

[0012] a wireless communication module, electrically connected to the micro sensor array, for transmitting the collected data;

[0013] a data processing unit, communicatively connected to the wireless communication module, for processing data collected by the sensor array; and

[0014] The algorithm prediction module includes a fusion system based on a decision tree algorithm and a neural network algorithm. The fusion system includes: (a) a neural network feature extraction unit for receiving and processing the sensor data to generate a feature vector F NN (b) a decision tree classification unit for constructing decision rules based on the sensor data and the feature vectors generated by the neural network to predict ball valve seal wear and potential failures; (c) an adaptive weight fusion unit for dynamically adjusting the weight coefficients of the neural network and decision tree outputs according to model performance; and (d) an early warning generation unit for generating fault early warning information 7-14 days in advance based on the prediction results of the fusion system.

[0015] Preferably, the microsensor array comprises:

[0016] Angle sensor, used to detect valve opening;

[0017] Pressure sensor, used to detect medium pressure;

[0018] Temperature sensor, used to detect valve and medium temperature;

[0019] Flow sensors for detecting medium flow; and

[0020] Vibration sensor, used to detect vibration signals during valve operation.

[0021] Preferably, the neural network feature extraction unit adopts a three-layer feedforward neural network structure, including:

[0022] Input layer, corresponding to the preprocessed sensor data X processed ;

[0023] Hidden layer, using ReLU activation function for nonlinear mapping; and

[0024] Output layer, generates feature vector F NN ; Among them, the output calculation formula of the hidden layer is: as well as , the eigenvector calculation formula is: Among them, W1, W2, W3 are weight matrices, b1, b2, b3 are bias vectors, and ReLU is the rectified linear unit activation function.

[0025] Preferably, the decision tree classification unit selects the optimal split feature based on the information gain criterion, and the information gain calculation formula is: , where H(D) is the entropy of the data set D, and the calculation formula is , p i is the proportion of the i-th class sample in the data set D, A is the feature, D v It is the data subset when the feature A takes the value v.

[0026] Preferably, the adaptive weight fusion unit dynamically adjusts the weights according to the historical prediction accuracy of the neural network and the decision tree, and the calculation formula is: as well as , where Acc NN and Acc DT are the prediction accuracy of neural network and decision tree respectively, β is the sensitivity parameter for weight adjustment, P final is the final prediction result, and sigmoid is the S-type activation function.

[0027] Preferably, an online learning update mechanism is also included, which includes:

[0028] Regularly collect new data samples;

[0029] Update the neural network parameters every time N new samples are collected;

[0030] Reconstruct the decision tree structure every M days; and

[0031] Dynamically adjust feature extraction methods based on the latest data.

[0032] Preferably, the warning generation unit has a multi-fault mode recognition capability, and can simultaneously detect and distinguish multiple fault modes such as seal wear, valve sticking and sensor abnormality, and generate corresponding warning information and maintenance suggestions.

[0033] Preferably, the data processing unit includes a data preprocessing module for standardizing, denoising and extracting time series features from the original sensor data and converting it into a feature vector X processed : Among them, features include statistical features, frequency domain features and time series features.

[0034] In summary, the present invention has the following beneficial effects:

[0035] The present invention integrates the decision tree algorithm with the neural network algorithm, which not only effectively overcomes the defects of using each algorithm alone, but also produces a gain effect, which is specifically manifested as follows:

[0036] Overcoming the overfitting problem of decision trees: Using the abstract features extracted by the neural network as input to the decision tree significantly reduces the tree's sensitivity to noise in the original data and improves its generalization ability. Compared to using a decision tree alone, the fusion system reduces overfitting and improves accuracy on the test set.

[0037] Enhanced neural network interpretability: Features extracted by the neural network are converted into explicit rules through a decision tree, allowing maintenance personnel to clearly understand the basis for fault warnings. This improves decision transparency compared to using neural networks alone.

[0038] Optimize resource utilization: By extracting key features through lightweight neural networks and then using decision trees with less computational burden to make decisions, system resource consumption is reduced, enabling ball valves to perform intelligent analysis on edge devices.

[0039] Extended prediction time span: The fusion system not only overcomes the shortcomings of each algorithm, but also extends the warning time from 2-3 days of traditional methods to 7-14 days, providing maintenance personnel with sufficient preparation time, which is impossible to achieve using either algorithm alone.

[0040] The false alarm rate is significantly reduced: When using decision trees or neural networks alone, the false alarm rate is high, while the fusion system has a lower false alarm rate, greatly improving the reliability of maintenance decisions.

[0041] Simultaneous identification of multiple fault modes: The fusion system demonstrates the ability to identify complex faults and can simultaneously detect and distinguish multiple fault modes.

[0042] Improved overall system performance: Compared with traditional ball valves, the self-sensing intelligent ball valve of the present invention has lower maintenance costs and improved system reliability, making it suitable for industrial Internet of Things and remote management scenarios.

[0043] In summary, the present invention overcomes the shortcomings of a single algorithm through algorithm fusion, produces a synergistic gain effect, and realizes accurate monitoring of the ball valve status and early warning of potential faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a schematic diagram of the overall structure of the self-sensing intelligent ball valve of the present invention;

[0045] Figure 2 Schematic diagram of the data processing flow of the present invention;

[0046] Figure 3 This is a structural block diagram of the algorithm fusion system of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be described in further detail below with reference to the accompanying drawings.

[0048] like Figure 1 As shown, the self-sensing intelligent ball valve provided by the present invention includes a ball valve body 10, a micro sensor array 20, a wireless communication module 30, a data processing unit 40 and an algorithm prediction module 50.

[0049] The ball valve body 10 is a conventional ball valve structure, including a valve body 11 , a spherical opening and closing member 12 , a valve stem 13 , a sealing ring 14 and a driving device 15 .

[0050] The micro sensor array 20 is installed at key locations on the ball valve body, including:

[0051] Angle sensor 21: installed at the valve stem connection, used to detect the valve opening;

[0052] Pressure sensor 22: installed at the inlet and outlet of the valve body to detect the medium pressure;

[0053] Temperature sensor 23: installed on the valve body surface and the medium channel to detect temperature;

[0054] Flow sensor 24: installed at the valve outlet, used to detect the medium flow;

[0055] Vibration sensor 25: installed on the surface of the valve body, used to detect vibration signals.

[0056] Wireless communication module 30 uses Bluetooth Low Energy (BLE) or LoRa technology to transmit sensor data to data processing unit 40. Data processing unit 40, which can be installed in a control cabinet or on a remote server, includes a data preprocessing module 41 and a data storage module 42. Algorithm prediction module 50, the core of the present invention, includes an algorithm fusion system.

[0057] like Figure 3 As shown, the algorithm fusion system includes a neural network feature extraction unit, a decision tree classification unit, an adaptive weight fusion unit and an early warning generation unit.

[0058] 1. Data Preprocessing

[0059] Raw sensor data X raw First, it is preprocessed and converted into a feature vector X processed :

[0060] .

[0061] in:

[0062] Normalize is a standardization function that converts data into a standard form with a mean of 0 and a variance of 1;

[0063] Denoise is a denoising function that uses wavelet transform to remove signal noise;

[0064] Extract Features For the feature extraction function, statistical features (mean, variance, skewness, kurtosis), frequency domain features (FFT coefficient, spectrum peak) and time series features (rate of change, trend coefficient) are calculated.

[0065] For temperature sensor data, the extracted features include: 24-hour average temperature; temperature fluctuation standard deviation; temperature change rate (°C / hour); and the main frequency of periodic temperature changes.

[0066] 2. Neural Network Feature Extraction Unit

[0067] The neural network feature extraction unit adopts a three-layer feedforward neural network structure:

[0068] Input layer: The number of nodes is the same as the number of features after preprocessing, receiving X processed ;

[0069] Hidden layer 1: 64 nodes, using ReLU activation function

[0070] Hidden layer 2: 32 nodes, using ReLU activation function

[0071] Output layer: 16 nodes, generating feature vector F NN ;

[0072] The calculation process is as follows: as well as ;

[0073] Among them, W1(64×n), W2(32×64), and W3(16×32) are weight matrices, b1, b2, and b3 are bias vectors, and ReLU(x)=max(0,x) is the activation function.

[0074] The Adam optimizer is used for training, the learning rate is set to 0.001, the batch size is 64, and the number of training rounds is 200.

[0075] 3. Decision Tree Classification Unit

[0076] The decision tree classification unit receives the enhanced feature set F enhanced =[X processed ,F NN ], building a decision tree based on the information gain criterion.

[0077] The information gain calculation formula is: ;

[0078] in, is the entropy of the data set D, p i is the proportion of the i-th class sample in D, A is the feature, D v It is the data subset when the feature A takes the value v.

[0079] To prevent overfitting, the following parameter settings are used: maximum tree depth = 8; minimum number of split samples = 20; pruning strategy: cost complexity pruning, α = 0.01.

[0080] 4. Adaptive Weight Fusion Unit

[0081] The adaptive weight fusion unit dynamically adjusts the weights of the neural network and decision tree outputs: as well as

[0082] Of which: Acc NN and Acc DT are the accuracy of the neural network and decision tree on the validation set respectively; β is the weight adjustment sensitivity parameter, the default value is 5; sigmoid(x)=1 / (1+e^(-x)) is the S-type activation function; P NN and P DT are the predicted probabilities of the neural network and decision tree respectively; P final The final prediction result.

[0083] When the accuracy of the neural network is higher than that of the decision tree, the α value increases, and the system is more inclined to the prediction results of the neural network; otherwise, it is more inclined to the prediction results of the decision tree.

[0084] 5. Warning generation unit

[0085] The warning generation unit generates warning information based on the output results of the fusion system: When P final >0.85, triggers a high-risk warning; when 0.6 <P final When ≤0.85, a medium risk warning is triggered; when 0.4 <P final When ≤0.6, a low-risk warning is triggered; when P final When ≤0.4, the system is normal.

[0086] Warning information includes fault type, probability, expected time of occurrence, recommended maintenance measures, and diagnostic basis. Specifically, in this invention, for example, a warning for seal wear is set, with an 87% probability of failure and an estimated need for replacement within seven days. The recommended seal is prepared, and the diagnostic basis includes increased temperature fluctuations and decreased pressure stability.

[0087] The online learning update mechanism of the present invention includes:

[0088] Regular data collection: The system collects sensor data once an hour to form a time series record.

[0089] Neural network parameter update: perform incremental learning every time 100 new samples are collected; use the stochastic gradient descent algorithm to update the weight: W new =W old -η×∇L; the learning rate η is dynamically adjusted: η=η0 / (1+k×t), where t is the number of updates.

[0090] Decision tree structure reconstruction: The decision tree structure is re-evaluated every 30 days. If the accuracy rate drops by more than 5%, the decision tree is reconstructed. During the reconstruction, some valid branches can be retained to improve learning efficiency.

[0091] Dynamic adjustment of feature extraction: calculating feature importance scores , retain the 80% of features with the highest scores; introduce new features every quarter and evaluate the effects.

[0092] The self-sensing intelligent ball valve of the present invention not only overcomes the shortcomings of a single algorithm through the fusion algorithm of decision tree and neural network, but also improves the fault prediction time and accuracy, achieving a significant gain effect.

[0093] The present invention provides an example calculation process to demonstrate the early warning function of the self-sensing intelligent ball valve described in the present invention:

[0094] 1. Data Collection and Preprocessing

[0095] The main body of the data is a DN100 smart ball valve. Five types of sensors are installed on the ball valve, collecting data every hour. The following is a sample of raw data from 7 consecutive days (168 hours):

[0096] Table 1: Snippets of sensor raw data (partial time points)

[0097]

[0098] (1) Data preprocessing calculation

[0099] First, normalize the raw data, for example, the temperature parameter:

[0100] Normalized temperature = (original temperature - average temperature) / temperature standard deviation; calculation of valve body temperature data: average temperature = 43.1°C; temperature standard deviation = 0.5°C; therefore, the normalized valve body temperature at time t1 = (42.3-43.1) / 0.5 = -1.6°C.

[0101] Then extract the time series features and calculate the key indicators:

[0102] Table 2: Some feature results after preprocessing

[0103]

[0104] Finally, the preprocessed feature vector X is obtained processed , containing 25 features.

[0105] 2. Neural network feature extraction calculation

[0106] (1) Neural network parameter setting

[0107] The neural network feature extractor uses the following parameters: input layer: 25 nodes (corresponding to the number of preprocessed features);

[0108] Hidden layer 1: 64 nodes, ReLU activation function;

[0109] Hidden layer 2: 32 nodes, ReLU activation function; output layer: 16 nodes (generating feature vector F NN ).

[0110] (2) Forward propagation calculation

[0111] Take the preprocessed feature vector X processed As input, calculate the output of hidden layer 1:

[0112] H1=ReLU(W1×X processed +b1);

[0113] Assume that some of the weights of W1 and some of the bias values of b1 are as follows:

[0114] W1[0,0]=0.24,W1[0,1]=-0.13,...,W1[0,24]=0.31;

[0115] b1[0]=0.05;

[0116] Then the input value of the first hidden layer neuron is:

[0117] z 10 =0.24×(-1.6)+(-0.13)×0.29+...+0.31×1.33+0.05=-0.42;

[0118] Apply the ReLU activation function:

[0119] H1[0]=ReLU(-0.42)=max(0,-0.42)=0.

[0120] Similarly, calculate the outputs of all 64 neurons in hidden layer 1. For example, the 15th neuron:

[0121] z 114 =...=1.72;

[0122] H1

[14] =ReLU(1.72)=1.72.

[0123] The calculation process of hidden layer 2 is similar:

[0124] H2=ReLU(W2×H1+b2);

[0125] Assume that some values of W2 and b2 are as follows:

[0126] W2[0,0]=0.11,W2[0,1]=0.23,...,W2[0,63]=-0.08;

[0127] b2[0]=-0.14.

[0128] Then the input value of the first neuron in the second hidden layer is:

[0129] z20 =0.11×0+0.23×...+(-0.08)×...+(-0.14)=0.85;

[0130] Apply ReLU:

[0131] H2[0]=ReLU(0.85)=0.85.

[0132] Finally, calculate the output layer feature vector:

[0133] F NN =W3×H2+b3;

[0134] Assume that some values of W3 and b3 are as follows:

[0135] W3[0,0]=0.32,W3[0,1]=-0.17,...,W3[0,31]=0.26;

[0136] b3[0]=0.09;

[0137] Compute the first eigenvalue:

[0138] F NN [0]=0.32×0.85+(-0.17)×...+0.26×...+0.09=0.73.

[0139] And so on, the complete 16-dimensional feature vector F is calculated NN :

[0140] F NN =[0.73,0.41,0.65,0.82,-0.35,0.11,0.92,-0.04,0.56,-0.21,0.47,0.63,0.38,-0.15,0.89,0.27]

[0141] 3. Decision tree classification calculation

[0142] (1) Constructing enhanced feature vector

[0143] The preprocessed features X processed and neural network features F NN Combine to form an enhanced feature vector:

[0144] F enhanced =[X processed ,F NN ]

[0145] =[mean pressure difference, standard deviation pressure difference, ..., F NN [0],F NN [1],...,F NN

[15] ]

[0146] (2) Information gain calculation

[0147] To build a decision tree, the information gain of each feature needs to be calculated. Suppose a training dataset D contains 100 samples, of which 30 samples are seal wear class (labeled as 1) and 70 samples are normal class (labeled as 0).

[0148] First calculate the entropy of the data set D:

[0149] Seal wear sample ratio: p1=30 / 100=0.3;

[0150] Normal sample ratio: p0=70 / 100=0.7;

[0151] H(D)=-(p0×log2(p0)+p1×log2(p1))

[0152] =-(0.7×log2(0.7)+0.3×log2(0.3))

[0153] =-(0.7×(-0.515)+0.3×(-1.737))

[0154] =0.361+0.521=0.882.

[0155] Taking the vibration amplitude growth rate feature as an example, assuming that 50% is used as the threshold for segmentation:

[0156] D1: There are 25 samples with a vibration amplitude growth rate of >50%, of which 22 are seal wear and 3 are normal;

[0157] D2: There are 75 samples with a vibration amplitude growth rate of ≤50%, of which 8 are seal wear and 67 are normal.

[0158] Calculate the entropy of D1:

[0159] p1=22 / 25=0.88;

[0160] p0=3 / 25=0.12;

[0161] H(D1)=-(0.88×log2(0.88)+0.12×log2(0.12))

[0162] =-(0.88×(-0.184)+0.12×(-3.059))

[0163] =0.162+0.367=0.529.

[0164] Calculate the entropy of D2:

[0165] p1=8 / 75=0.107;

[0166] p0=67 / 75=0.893;

[0167] H(D2)=-(0.893×log2(0.893)+0.107×log2(0.107))

[0168] =-(0.893×(-0.163)+0.107×(-3.224))

[0169] =0.146+0.345=0.491.

[0170] Calculate information gain:

[0171] IG(D,A)=H(D)-(|D1| / |D|×H(D1)+|D2| / |D|×H(D2))

[0172] =0.882-(25 / 100×0.529+75 / 100×0.491)

[0173] =0.882-(0.132+0.368)

[0174] =0.882-0.5=0.382.

[0175] For the neural network feature F NN [7] also performs the same calculation, assuming that the threshold 0 is used as the split point:

[0176] D1:F NN [7]>0, a total of 45 samples, of which 25 are seal wear and 20 are normal;

[0177] D2:F NN There are 55 samples with [7]≤0, of which 5 are seal wear and 50 are normal.

[0178] Calculate the information gain IG(D,F NN [7])=0.343.

[0179] By calculating the information gain of all features, we can get the top 5 features after sorting:

[0180] Vibration amplitude growth rate: 0.382; F NN [7]: 0.343; Temperature rise rate: 0.315; F NN [2]: 0.287; Pressure difference change rate: 0.264.

[0181] (3) Decision tree construction

[0182] According to the information gain, the first layer of the decision tree is constructed to select the vibration amplitude growth rate as the segmentation feature:

[0183] IF vibration amplitude growth rate > 50% THEN (enter the left subtree, 22 / 25 is seal wear)

[0184] ELSE (enter right subtree, 8 / 75 is seal wear)

[0185] In the left subtree, continue to calculate the information gain of the remaining features and select F NN [2] is the segmentation feature:

[0186] IF vibration amplitude growth rate>50% THEN

[0187] IF F NN [2]>0.5 THEN (20 / 21 is seal wear, predicted to be seal wear)

[0188] ELSE (2 / 4 is seal wear, further divided)

[0189] ELSE (continue building the right subtree)

[0190] An important branch path of the complete decision tree is as follows:

[0191] IF vibration amplitude growth rate>50% THEN

[0192] IF F NN [2]>0.5 THEN prediction: seal wear (95.2% probability)

[0193] ELSE

[0194] IF Temperature rise rate > 0.05°C / h THEN Prediction: Seal wear (75% probability)

[0195] ELSE

[0196] Prediction: Normal (100% probability)

[0197] ELSE

[0198] IF F NN [7]>0 THEN

[0199] IF pressure difference change rate>0.002MPa / h THEN

[0200] Prediction: Seal wear (63.6% probability)

[0201] ELSE

[0202] Prediction: Normal (92.3% probability)

[0203] ELSE

[0204] Prediction: Normal (96.4% probability)

[0205] 4. Adaptive weight fusion calculation

[0206] (1) Model accuracy evaluation

[0207] Evaluate the accuracy of the two models on the validation set: Neural Network Accuracy: Acc NN =0.87; decision tree accuracy: Acc DT =0.82

[0208] (2) Fusion weight calculation

[0209] Set the sensitivity parameter β=5 and calculate the fusion weight:

[0210] α=sigmoid(β×(Acc NN -Acc DT ))

[0211] =sigmoid(5×(0.87-0.82))=sigmoid(5×0.05)

[0212] =sigmoid(0.25)=1 / (1+e^(-0.25))=1 / (1+0.779)=1 / 1.779=0.562.

[0213] For the current sample, the neural network model and decision tree model give the seal wear probability respectively: Neural network prediction probability: P NN =0.76; decision tree prediction probability: P DT =0.95.

[0214] Calculate the final fusion prediction result:

[0215] P final =α×P NN +(1-α)×P DT =0.562×0.76+(1-0.562)×0.95

[0216] =0.427+0.416=0.843.

[0217] (3) Early warning judgment

[0218] According to P final =0.843, compared with the warning threshold:

[0219] When P final When >0.85, a high-risk warning is triggered;

[0220] When 0.6 <P final When ≤0.85, a medium-risk warning is triggered;

[0221] When 0.4 <P final When ≤0.6, a low-risk warning is triggered;

[0222] When P final When ≤0.4, the system is normal;

[0223] At this time P final =0.843, triggering a medium-risk warning.

[0224] 5. Fault development trend prediction

[0225] (1) Historical data trend analysis

[0226] By analyzing 78 seal wear cases in historical data, the corresponding relationship between seal wear probability and failure occurrence time was established:

[0227]

[0228] (2) Failure occurrence time prediction

[0229] Based on the current P final =0.843, predicted failure time range: reference probability interval: 0.8-0.9; average failure time: 8.5 days; failure time range: 8.5±1.8 days;

[0230] Final warning result: seal wear warning, failure probability 84.3%, expected to occur within 7-10 days.

[0231] 6. Results

[0232] The ball valve continued to operate after the warning, and maintenance personnel conducted daily observations. Nine days after the warning, significant leakage was discovered at the seal. Disassembly and inspection revealed that the seal ring was worn to the point of replacement.

[0233] Comparative analysis: Predicted failure time: 7-10 days; Actual failure time: 9 days; Error: Within the predicted range; Failure type: Seal wear, consistent with the prediction.

[0234] This result validates the effectiveness of the self-sensing intelligent ball valve system proposed in this paper. Compared to using a decision tree alone (prediction time of 5-7 days) or a neural network alone (prediction time of 6-8 days), the fusion algorithm significantly improves prediction accuracy and lead time.

[0235] 7. Sensitivity Analysis

[0236] By adjusting the adaptive fusion parameter β, we can observe the impact on the prediction results:

[0237]

[0238] By adjusting the maximum depth of the decision tree, we can observe the impact of model complexity on prediction performance:

[0239]

[0240] The above analysis shows that the fusion algorithm of the present invention has good stability and can maintain stable prediction performance within the range of parameter changes.

[0241] The self-sensing intelligent ball valve system described in the present invention has a complete computational process from data preprocessing, neural network feature extraction, decision tree classification to adaptive fusion. It proves that the proposed algorithm fusion method can effectively predict ball valve seal wear failures and issue an early warning 7-10 days in advance, which is a significant improvement over using any algorithm alone, fully verifying the practicality and creativity of the present invention.

Claims

1. A self-sensing intelligent ball valve, characterized in that: include: The ball valve body is used to control the switching and flow regulation of the fluid; A micro sensor array is installed on the ball valve body to collect real-time data on valve opening, medium pressure, temperature and flow rate; a wireless communication module, electrically connected to the micro sensor array, for transmitting the collected data; a data processing unit, communicatively connected to the wireless communication module, for processing data collected by the sensor array; as well as The algorithm prediction module includes a fusion system based on a decision tree algorithm and a neural network algorithm. The fusion system includes: (a) a neural network feature extraction unit for receiving and processing the sensor data to generate a feature vector F NN (b) a decision tree classification unit for constructing decision rules based on the sensor data and the feature vectors generated by the neural network to predict ball valve seal wear and potential failures; (c) an adaptive weight fusion unit for dynamically adjusting the weight coefficients of the neural network and decision tree outputs based on model performance; and (d) an early warning generation unit for generating fault early warning information 7-14 days in advance based on the prediction results of the fusion system; The adaptive weight fusion unit dynamically adjusts the weights according to the historical prediction accuracy of the neural network and decision tree. The calculation formula is: as well as , where Acc NN and Acc DT are the prediction accuracy of neural network and decision tree respectively, β is the sensitivity parameter for weight adjustment, P final is the final prediction result, sigmoid is the S-type activation function; P NN and P DT are the predicted probabilities of the neural network and decision tree, respectively.

2. The self-sensing intelligent ball valve according to claim 1, characterized in that: The micro sensor array comprises: Angle sensor, used to detect valve opening; Pressure sensor, used to detect medium pressure; Temperature sensor, used to detect valve and medium temperature; Flow sensors for detecting medium flow; and Vibration sensor, used to detect vibration signals during valve operation.

3. The self-sensing intelligent ball valve according to claim 1, characterized in that: The neural network feature extraction unit adopts a three-layer feedforward neural network structure, including: Input layer, corresponding to the preprocessed sensor data X processed ; Hidden layer, using ReLU activation function for nonlinear mapping; and Output layer, generates feature vector F NN ; Among them, the output calculation formula of the hidden layer is: as well as , the eigenvector calculation formula is: , where W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias vectors, and ReLU is the rectified linear unit activation function.

4. The self-sensing intelligent ball valve according to claim 1, characterized in that: The decision tree classification unit selects the optimal split feature based on the information gain criterion, and the information gain calculation formula is: , where H(D) is the entropy of the data set D, and the calculation formula is , p i is the proportion of the i-th class sample in the data set D, A is the feature, D v It is the data subset when the feature A takes the value v.

5. The self-sensing intelligent ball valve according to claim 1, characterized in that: Also included is an online learning update mechanism, which includes: Regularly collect new data samples; Update the neural network parameters every time N new samples are collected; Reconstruct the decision tree structure every M days; and Dynamically adjust feature extraction methods based on the latest data.

6. The self-sensing intelligent ball valve according to claim 1, characterized in that: The warning generation unit has the ability to identify multiple fault modes, and can simultaneously detect and distinguish the fault modes of seal wear, valve sticking and sensor abnormality, and generate corresponding warning information and maintenance suggestions.

7. The self-sensing intelligent ball valve according to claim 1, characterized in that: The data processing unit includes a data preprocessing module for standardizing, denoising and extracting time series features of the original sensor data and converting it into a feature vector X processed : , where the features include statistical features, frequency domain features and time series features; X raw is the original sensor data; Normalize is a normalization function that converts the data into a standard form with a mean of 0 and a variance of 1; Denoise is a denoising function that uses wavelet transform to remove signal noise; Extract Features It is a feature extraction function that calculates statistical features, frequency domain features and time series features.

Images