Generator set pure water system working condition abnormity monitoring method based on XGboost algorithm

Abstract

The invention relates to a generator set pure water system working condition abnormity monitoring method based on an XGboost algorithm. The method comprises the following steps: acquiring a voiceprint signal of a generator set pure water system; extracting feature data from the obtained voiceprint signal data; constructing a fault type recognition model of the pure water system of the generator set by adopting an XGboost algorithm, and taking the obtained feature vector of the voiceprint signal and the temperature data and conductivity of the pure water system as the input of the fault type recognition model; training the fault type identification model by using historical data of the pure water system of the generator set; and applying the trained fault type identification model to real-time fault prediction of the pure water system of the generator set. Accurate matching of fault types of the unit pure water system and accurate positioning of fault objects are achieved, manual fault detection is replaced, the accuracy of fault recognition of the pure water system is improved, real-time monitoring of the abnormal working condition development trend of the pure water system is achieved, and the stability and safety of a large hydraulic generator are improved.

Description

Technical Field

[0001] The present invention belongs to the field of hydropower generation monitoring, and in particular relates to a method for monitoring abnormal operating conditions of a pure water system of a generator set based on an XGboost algorithm. Background Art

[0002] The hydro-turbine generator set is equipped with a pure water system to cool the stator winding of the generator. If the pure water system fails, the heat generated by the stator winding will increase suddenly, endangering the safety of the hydro-turbine generator set and having a great impact on the stability of the power system. Therefore, it is of great significance to ensure the healthy operation of the pure water system and to promptly detect and eliminate its abnormal operating status.

[0003] The pure water system is a closed pure water circuit composed of a pure water device and a hollow stator winding, a hollow converging copper ring, a polytetrafluoroethylene connecting hose, a stator winding inlet and outlet water ring pipe, and a connecting water pipe from the ring pipe to the pure water treatment device. The cooling process of the pure water system is as follows: the pure water flows out through the generator stator winding, is pressurized by the booster pump, and is adjusted by the electric three-way valve. A part of the hot water enters the water-water heat exchanger for cooling. The cooled water is mixed with the hot water to make the temperature of the pure water meet the set value. Then, after being filtered by the filter, it enters the generator water inlet main ring pipe and flows into the stator winding. Common faults of the pure water system include abnormal pure water temperature, reduced pure water main circuit flow, excessively high pure water main circuit conductivity, large pure water leakage, reduced water level in the expansion tank, and pure water pump stopping.

[0004] The existing method of monitoring the pure water system fault of the unit by setting normal values, preset alarm values ​​and shutdown range values ​​ignores the correlation between the monitoring data of different characteristics of the pure water system, lacks in-depth analysis of the characteristic data of the monitoring data of the pure water system, and cannot timely and effectively identify the fault signs of complex faults of the pure water system. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems and provide a method for monitoring the operating abnormality of a pure water system of a generator set based on the XGboost algorithm, obtain the voiceprint signal of the pure water system of the unit, and further extract the feature vector therein, and use the temperature data and conductivity data of the pure water system as input data of a fault type identification model of the pure water system of the unit, thereby improving the accuracy of pure water system fault identification, identifying the fault signs of the pure water system as early as possible, realizing abnormal operating condition monitoring of the pure water system, and providing guarantee for the safe and stable operation of the hydropower generator set.

[0006] The technical solution of the present invention is a method for monitoring abnormal operating conditions of a pure water system of a generator set based on an XGboost algorithm, comprising the following steps: Step 1: Obtain the voiceprint signal of the pure water system of the generator set; Step 2: Extract feature data from the voiceprint signal data obtained in step 1; Step 2.1: Preprocessing the voiceprint signal data; Step 2.2: Perform fast Fourier transform on the pre-processed voiceprint signal data to obtain frequency domain feature data in the voiceprint signal; Step 2.3: Take the square of the modulus of the frequency domain feature data obtained in step 2.2; Step 2.4: Perform Mel filtering to convert the frequency in the voiceprint signal into Mel scale; Step 2.5: Take the logarithm of the filtering result obtained in step 2.4 to obtain the Fbank feature vector of the voiceprint signal; Step 3: Use the XGboost algorithm to build a fault type recognition model for the pure water system of the generator set, and use the Fbank feature vector of the voiceprint signal obtained in step 2 and the temperature data and conductivity of the pure water system as input data of the fault type recognition model; Step 4: Using historical data of the pure water system of the generator set to train the fault type identification model; Step 5: Use the trained fault type recognition model for real-time fault prediction of the pure water system of the generator set.

[0007] Preferably, in step 1, the voiceprint signal of the pure water system of the generator set is obtained by obtaining the voiceprint signal data of the expansion water tank, the pure water pipeline and the cooling water pipeline from the voiceprint collectors at the expansion water tank, the pure water pipeline and the cooling water pipeline respectively.

[0008] Furthermore, in step 2.1, the preprocessing includes pre-emphasis, framing and windowing the voiceprint signal, filtering out non-audio components in the sound and then performing truncation processing.

[0009] Preferably, in step 2.1, the windowing adopts a Hamming window, and the Hamming window function is: (1) In the formula represents the Hamming window function, I Indicates the number of frames, is the coefficient, i F Indicates the frame number.

[0010] Preferably, in step 2.2, a fast Fourier transform method is used to deconstruct the preprocessed time domain signal into a frequency domain signal, and perform fast Fourier transform processing on the signal: (2) In the formula represents the function of fast Fourier transform; Represents the input sound signal, represents the number of points of the fast Fourier transform, n F Represents the index of the sampling point in the time domain; N Indicates the total number of sampling points.

[0011] Furthermore, in step 2.3, the square of the modulus is specifically calculated as follows: (3) In the formula, P Represents the spectral line energy of the sound signal, Indicates signal No. i F frame, FFT ( ) represents fast Fourier transform.

[0012] Furthermore, in step 2.4, the Mel filter uses a Mel filter including multiple triangular filters to convert the frequency of the voiceprint signal into a Mel scale using the following calculation formula: (4) In the formula Mel ( ) represents the Mel function, f Indicates the frequency in the voiceprint signal.

[0013] Preferably, the logarithmic energy of the filter output is calculated in step 2.5 by the following formula: ; (5) In the formula S ( m ) indicates the m The logarithmic energy of the filter bank output; represents the function to perform fast Fourier transform, represents the number of points of the fast Fourier transform, M Indicates the number of triangular filters; m Indicates the filter number; N Indicates the total number of sampling points.

[0014] in

[0015] Furthermore, in step 3, the XGboost algorithm contains multiple decision trees, which combine a set of weak learners into a strong learner. , is the input vector containing m features, is the sample label, is the sample number, samples, using Indicates A regression tree, is the predicted value output by the XGboost algorithm for the parameter set of the regression tree The expression is: ; (6) In the formula F K ( ) represents the output function of the XGboost algorithm, is the number of regression trees; solve by minimizing the objective function , the objective function is based on the actual value With the predicted value The loss function is constructed. ; (7) ; In the formula, is the objective function; ( ) is the loss function; is the regularization function; , All are hyperparameters; is the number of leaves; is the weight of the leaf node.

[0016] Preferably, in step 4, when training the fault type identification model, formula (7) is rewritten as: (8) In the formula, Indicates s The objective function of round training; Indicates s The predicted value of the round; n represents the number of samples; Indicates s - Predicted values ​​for 1 round of training.

[0017] For the loss function in the objective function, use Taylor's formula to expand and get the second-order derivative: (9) In the formula , denote the first-order and second-order coefficients respectively; ; ; Refine the tree model to leaf nodes and sort it out to get: (10) In the formula, is the sample set of the leaf node, that is, all samples falling on the leaf node; the optimal weight of the leaf node is obtained by solving formula (10): and the optimal objective function value , ; (11) ; (12) To avoid overfitting, a reduction factor is set during the training iteration process to achieve small-step iterative optimization: (13) In the formula, is the reduction factor of the new spanning tree.

[0018] Compared with the prior art, the beneficial effects of the present invention include: 1) The present invention extracts the characteristic vector of the voiceprint signal of the unit's pure water system and uses the temperature data and conductivity data of the pure water system as the input data of the unit's pure water system fault type identification model, thereby achieving accurate matching of the unit's pure water system fault type and precise positioning of the fault object, replacing manual fault detection, improving the accuracy of pure water system fault identification, being able to identify the fault signs of the pure water system in advance, and achieving real-time monitoring of the abnormal development trend of the pure water system operating condition, thereby greatly reducing labor costs and improving the stability and safety of large-scale hydro-generators.

[0019] 2) The present invention calculates the FBank feature of the voiceprint signal of the pure water system for pure water system fault identification. The FBank feature can effectively represent the frequency domain characteristics of the speech signal with less feature data, thereby reducing the computational complexity of the fault monitoring method of the present invention, improving the operating efficiency of the pure water system fault monitoring system, and ensuring the real-time performance of the pure water system fault monitoring of the unit.

[0020] 3) The present invention adopts the XGboost algorithm to construct a fault type identification model for the unit pure water system, and by setting a reduction coefficient during the training process of the XGboost algorithm model, small-step iterative optimization is achieved to prevent overfitting and improve the accuracy of pure water system fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] Figure 1 It is a schematic structural diagram of a pure water system of a unit according to an embodiment of the present invention.

[0023] Figure 2 Schematic diagram of a data acquisition system for a pure water system of a unit according to an embodiment of the present invention.

[0024] Figure 3Schematic diagram of the process of extracting Fbank features according to an embodiment of the present invention.

[0025] Figure 4 Schematic diagram of a method for monitoring abnormal operating conditions of a pure water system of a unit according to an embodiment of the present invention.

[0026] Figure 5 This is a pre-emphasized waveform diagram of the voiceprint signal of the pure water system of the unit according to an embodiment of the present invention.

[0027] Figure 6 It is a frame waveform diagram of the voiceprint signal of the pure water system of the unit according to an embodiment of the present invention.

[0028] Figure 7 It is a windowed waveform diagram of the voiceprint signal of the pure water system of the unit according to an embodiment of the present invention.

[0029] Figure 8 It is a spectrum waveform diagram of the voiceprint signal of the pure water system of the unit according to an embodiment of the present invention.

[0030] Fig. 9 This is a logarithmic energy waveform diagram of the soundprint signal of the pure water system of the unit according to an embodiment of the present invention.

[0031] Explanation of the accompanying drawings: water pump 1, cooler 2, mechanical filter 3, ion exchanger 4, alkalizer unit 5, expansion water tank 6, nitrogen storage tank 7, water replenishment device 8, bypass 9, stator cooling water pipeline 10, stator winding 11, expansion water tank soundprint sensor 12, pure water main circuit soundprint sensor 13, cooler soundprint sensor 14, water pump motor temperature sensor 15, conductivity sensor 16, stator outlet water temperature sensor 17. DETAILED DESCRIPTION

[0032] like Figure 1 and Figure 2 As shown, an expansion water tank soundprint sensor 12, a pure water main circuit soundprint sensor 13 and a cooler soundprint sensor 14 are installed at the expansion water tank 6, the pure water pipeline and the outlet of the cooler 2 of the pure water system of the hydropower generator set respectively. The expansion water tank soundprint sensor 12, the pure water main circuit soundprint sensor 13 and the cooler soundprint sensor 14 all adopt ZKHY-IVOD-013 industrial diagnostic pickups with a collection frequency of 20HZ-20KHZ; a water pump motor temperature sensor 15 and a stator water outlet water temperature sensor 17 are installed at the motor and stator cooling water pipeline 10 of the water pump 1. The water pump motor temperature sensor 15 and the stator water outlet water temperature sensor 17 both adopt PT100 thermistor sensors with a measuring range of -200℃~850℃; a conductivity sensor 16 is installed at the outlet of the mechanical filter 3. The conductivity sensor 16 adopts SE625 conductivity sensor with a measuring range of 0-1000µS / cm.

[0033] like Figure 2 As shown, the output ends of the expansion tank voiceprint sensor 12, the pure water main circuit voiceprint sensor 13 and the cooler voiceprint sensor 14 are all connected to the edge computing gateway. The edge computing gateway uses the edge gateway terminal integrated machine ZKHY-IVH-002, which can collect 8-channel audio input and the input voltage is 9~36V. The edge computing gateway is connected to the server via a switch, and the server uses the cloud engine server ZKHY-VCES-2000. The water pump motor temperature sensor 15, the stator outlet water temperature sensor 17 and the conductivity sensor 16 are connected to the local control LCU, and the local control LCU is connected to the server via a switch. like Figure 4 As shown, the abnormal operating condition monitoring method of the pure water system of the generator set based on the XGboost algorithm includes: Step 1: Obtain the soundprint signal of the pure water system of the generator set, and collect the soundprint signal data of the expansion water tank, pure water pipeline and cooling water pipeline through the expansion water tank soundprint sensor, pure water main circuit soundprint sensor and cooler soundprint sensor.

[0034] Step 2: Extract Fbank feature vectors from the voiceprint signal, including: Step 1: Preprocess the voiceprint signal, pre-emphasize, frame, and window the collected 50ms sound, with a pre-emphasis coefficient of 0.97; filter out the non-audio components in the sound and then truncate it, using a Hamming window, and the Hamming window function is as follows: ; (1) In the formula represents the Hamming window function, I Indicates the number of frames, is the coefficient, i F Indicates the frame number.

[0035] Step 2: Fast Fourier Transform FFT, deconstruct the preprocessed time domain signal into a frequency domain signal, making it suitable for training in a neural network, and perform FFT processing on it: ; (2) In the formula represents the function of fast Fourier transform; Represents the input sound signal, represents the number of points of the fast Fourier transform, n F Represents the index of the sampling point in the time domain; N Indicates the total number of sampling points.

[0036] Step 3: Take the square of the modulus of the signal in the frequency domain: (3) In the formula, P Represents the spectral line energy of the sound signal, Indicates signal No. i F frame, FFT ( ) represents fast Fourier transform.

[0037] Step 4: Perform Mel filtering. The Mel filter consists of 24 triangular filters. The formula for converting frequency to Mel scale is: (4) In the formula, represents the final frequency, Indicates the actual frequency.

[0038] Step 5: Take the logarithm of the filtered result and calculate the logarithmic energy of each filter output. The formula is: (5) In the formula S ( m ) indicates the m The logarithmic energy of the filter bank output; represents the function to perform fast Fourier transform, represents the number of points of the fast Fourier transform, M Indicates the number of triangular filters; m Indicates the filter number.

[0039] After the above steps, the Fbank feature is extracted. The extraction process of the Fbank feature is as follows: Figure 3 shown.

[0040] Step 3: Use the XGboost algorithm to build a fault type identification model for the pure water system of the generator set; The fault type recognition model based on the XGboost algorithm consists of multiple decision trees. A set of weak learners are combined into a strong learner through an additive model. , To include The vector of features, is the sample label, is the sample number, samples, using Indicates A regression tree, For the parameter set of the regression tree, the model prediction value expression is: ; (6) In the formula, is the number of regression trees, The solution needs to minimize the corresponding objective function, which is based on the actual value With the predicted value The loss function is constructed. ; (7) ; In the formula, is the objective function; ( ) is the loss function; is the regularization function; , All are hyperparameters; is the number of leaves; is the weight of the leaf node.

[0041] Step 4: Select training data and training iterations, train the pure water system fault diagnosis model, and obtain a trained pure water system fault type recognition model with the ability to recognize pure water system fault types.

[0042] Assuming the sth round of training, rewrite equation (7) as: (8) In the formula, is the predicted value of round s; For the loss function in the objective function, the Taylor formula is used to expand and approximate the second-order derivative: (9) ; ; Refine the tree model to leaf nodes and sort it out to get: (10) In the formula, is the sample set of the leaf node, that is, all samples falling on the leaf node; the optimal weight of the leaf node is obtained by solving formula (10): and the optimal objective function value , ; (11) ; (12) To avoid overfitting, a reduction factor is set during the training iteration process to achieve small-step iterative optimization: ; (13) In the formula, is the reduction factor of the new spanning tree.

[0043] The Fbank feature vectors extracted from the multi-channel voiceprint signals of expansion water tank water level, pure water flow, and cooling water flow collected by the voiceprint sensor, as well as the feature vectors of conductivity, oil pump motor temperature, and pure water temperature are input into the XGboost model for training.

[0044] Step 5: Use the trained fault type recognition model for real-time fault prediction of the pure water system of the generator set. The fault type recognition model extracts the Fbank feature vector from the real-time collected voiceprint signal of the pure water system, as well as the pure water conductivity, oil pump motor temperature and pure water temperature, to obtain the prediction value for fault identification and fault location; and judges the fault type based on the prediction value to realize the fault type recognition and diagnosis of the pure water system; and judges the abnormal development trend of the pure water system operating condition based on the continuous prediction value output by the fault type recognition model.

[0045] In one embodiment, the specific process of extracting the Fbank feature vector from the voiceprint signal is as follows: 1) Voice signal reading: Use the voiceprint sensor to collect the voiceprint signal of the pure water system during operation at a sampling rate of 16kHz and a sampling accuracy of 24bit, and use the audio processing library to read the input 50ms voice signal.

[0046] 2) Perform sound sample processing; Preprocessing: Perform pre-emphasis processing on the audio signal, and the pre-emphasis coefficient is set to 0.97. Then, the audio signal is divided into frames, each frame length is 16ms, and the frame shift is 8ms. Finally, each frame is windowed, and the window function is selected as the Hamming window. The pre-emphasis waveform of the output signal of a voiceprint sensor in the embodiment is as follows Figure 5 As shown, its frame waveform is as follows Figure 6 As shown, the windowed waveform is Figure 7 shown.

[0047] 3) Fast Fourier Transform: Perform fast Fourier transform on each frame of speech signal to obtain its spectrum. Figure 5 The spectrum of the voiceprint signal shown in the figure is obtained after framing, windowing and fast Fourier transform. Figure 8 The number of fast Fourier transform points is set to 512.

[0048] 4) Mel filtering: Design a Mel filter bank with 24 filters. The filter frequency range covers the entire spectrum range of the speech signal. Filter the spectrum obtained by fast Fourier transform through the Mel filter bank to obtain the Mel spectrum.

[0049] 5) Take the logarithm: Perform a logarithmic transformation on the Mel spectrum to obtain the logarithmic Mel spectrum. The logarithmic transformation uses the natural logarithm function.

[0050] 6) Feature extraction: The logarithmic Mel spectrum is extracted as Fbank features. The extracted feature parameters include the logarithmic energy value of each filter channel. Figure 8 The logarithmic energy waveform obtained after the voiceprint signal shown in the figure is filtered by Mel and logarithmically transformed is as follows: Fig. 9 shown.

[0051] The system of the above-mentioned method for monitoring abnormal operating conditions of the pure water system of the generator set comprises: Voiceprint signal acquisition unit: used to acquire the voiceprint signal data of the expansion water tank, pure water pipeline and cooling water pipeline of the unit's pure water system, and store them in the database; Fbank feature extraction unit: communicates with the voiceprint signal acquisition unit and is used to extract Fbank feature data from the voiceprint signal of the unit's pure water system; LCU signal extraction unit: connected to the local control LCU, used to obtain the water pump motor temperature of the pure water system, the stator winding outlet water temperature and the conductivity data of the pure water system, and store them in the database; Fault diagnosis model unit: communicates with the Fbank feature extraction unit and the LCU signal extraction unit, and uses the XGboost algorithm to build a pure water system fault diagnosis model. The input of the pure water system fault diagnosis model is the Fbank feature data, the water pump motor temperature, the stator winding outlet water temperature, and the conductivity data of the pure water system; Training unit: using historical data of the pure water system of the generator set to train the pure water system fault diagnosis model of the fault diagnosis model unit, adjusting the parameters of the pure water system fault diagnosis model during the training process to obtain the optimal pure water system fault diagnosis model; Fault identification and early warning unit: used to locate the fault object according to the predicted value output by the pure water system fault diagnosis model, determine the pure water system fault type and fault level, and issue an alarm signal for the identified fault or fault sign.

Claims

1. A method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm, characterized in that: The following steps are involved: Step 1: Obtain the voiceprint signal of the pure water system of the generator set; Step 2: Extract feature data from the voiceprint signal data obtained in step 1; Step 2.1: Preprocessing the voiceprint signal data; Step 2.2: Perform fast Fourier transform on the pre-processed voiceprint signal data to obtain frequency domain feature data in the voiceprint signal; Step 2.3: Take the square of the modulus of the frequency domain feature data obtained in step 2.2; Step 2.4: Perform Mel filtering to convert the frequency in the voiceprint signal into Mel scale; Step 2.5: Take the logarithm of the filtering result obtained in step 2.4 to obtain the Fbank feature vector of the voiceprint signal; Step 3: Use the XGboost algorithm to build a fault type recognition model for the pure water system of the generator set, and use the Fbank feature vector of the voiceprint signal obtained in step 2 and the temperature data and conductivity of the pure water system as inputs of the fault type recognition model; Step 4: Using historical data of the pure water system of the generator set to train the fault type identification model; Step 5: Use the trained fault type recognition model for real-time fault prediction of the pure water system of the generator set.

2. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 1 is characterized in that: In step 1, the voiceprint signal of the pure water system of the generator set is obtained by obtaining the voiceprint signal data of the expansion water tank, the pure water pipeline and the cooling water pipeline from the voiceprint collectors at the expansion water tank, the pure water pipeline and the cooling water pipeline respectively.

3. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 1 or 2, characterized in that: In step 2.1, the preprocessing includes pre-emphasis, framing and windowing of the voiceprint signal, filtering out non-audio components in the sound and then performing truncation processing.

4. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 3 is characterized in that: In step 2.1, the windowing adopts a Hamming window, and the Hamming window function is: ；（1） In the formula represents the Hamming window function, I Indicates the number of frames, is the coefficient, i F Indicates the frame number.

5. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 4 is characterized in that: In step 2.2, the fast Fourier transform method is used to deconstruct the preprocessed time domain signal into a frequency domain signal, and then perform fast Fourier transform processing on it: ；（2） In the formula represents the function of fast Fourier transform; Represents the input sound signal, represents the number of points of the fast Fourier transform, n F Represents the index of the sampling point in the time domain; N Indicates the total number of sampling points.

6. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 5 is characterized in that: In step 2.3, the square of the modulus is specifically calculated as follows: ；（3） In the formula, P Represents the spectral line energy of the sound signal, Indicates signal No. i F frame, FFT ( ) represents fast Fourier transform.

7. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 6 is characterized in that: In step 2.4, the Mel filter uses a Mel filter including multiple triangular filters to convert the frequency of the voiceprint signal into the Mel scale using the following calculation formula: ；（4） In the formula Mel ( ) represents the Mel function, f Indicates the frequency in the voiceprint signal.

8. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 7 is characterized in that: The step 2.5 calculates the logarithmic energy of the filter output through equation (5): ； （5） In the formula S ( m ) indicates the m The logarithmic energy of the filter bank output; represents the function to perform fast Fourier transform, represents the number of points of the fast Fourier transform, M Indicates the number of triangular filters; m Indicates the filter number; N Indicates the total number of sampling points; in ； In the formula Represents the frequency response function of a triangular filter.

9. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 2 or 4 or 5 or 6 or 7 or 8, characterized in that: In step 3, the XGboost algorithm contains multiple decision trees, which combine a group of weak learners into a strong learner. , is an input vector containing multiple features, is the sample label, is the sample number, samples, using Indicates A regression tree, is the predicted value output by the XGboost algorithm for the parameter set of the regression tree The expression is: ；（6） In the formula F K ( ) represents the output function of the XGboost algorithm, is the number of regression trees; solve by minimizing the objective function , the objective function is based on the actual value With the predicted value The loss function is constructed. ；（7） ； In the formula, is the objective function; ( ) is the loss function; is the regularization function; , All are hyperparameters; is the number of leaves; is the weight of the leaf node.

10. The method for monitoring abnormal operating conditions of a pure water system of a generator set based on the XGboost algorithm according to claim 9, characterized in that: In step 4, when training the fault type identification model, formula (7) is rewritten as: ；（8） In the formula, Indicates s The objective function of round training; Indicates s The predicted value of the round; n represents the number of samples; Indicates s - Prediction value of 1 round of training; For the loss function in the objective function, use Taylor's formula to expand and get the second-order derivative: ； （9） In the formula , denote the first-order and second-order coefficients respectively; ； ； Refine the tree model to leaf nodes and sort it out to get: ；（10） In the formula, is the sample set of the leaf node, that is, all samples falling on the leaf node; the optimal weight of the leaf node is obtained by solving formula (10): and the optimal objective function value , ；（11） ；（12） To avoid overfitting, a reduction factor is set during the training iteration process to achieve small-step iterative optimization: ； （13） In the formula, is the reduction factor of the new spanning tree.

Images