Fire-fighting pipe network fault inversion model training method

Abstract

The invention discloses a fire-fighting pipe network fault inversion model training method. The method comprises the following steps: collecting operation state data of a fire-fighting pipe network in a normal state and a fault state; on the basis of the operation state data, constructing characteristic indexes based on data statistical characteristics and marking fault labels; calculating a correlation coefficient between each characteristic index and the fault tag to obtain a correlation coefficient matrix; processing the correlation coefficient matrix by using a coupling model to obtain a feature importance ranking result; in combination with the feature importance ranking result and the correlation coefficient matrix, feature combination data is selected as an independent variable to be input into the fire-fighting pipe network fault inversion model for training, and an optimal fault inversion model is screened out by adjusting parameters; according to the method, the data volume required by model training can be effectively reduced, and the model training time is shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a fire protection pipe network fault inversion model training method. Background Art

[0002] As an important power transmission and conversion facility, ensuring the normal operation of the fire protection pipe network is crucial to timely extinguishing fires in converter stations. Due to the harsh operating environment, there are many typical faults in the fire protection pipe network, such as pipe blockage, leakage, valve failure, etc. Traditional fault diagnosis methods generally rely on manual inspections and experience judgments, which are inefficient and prone to missed detections and misjudgments.

[0003] In related technologies, the master's thesis "Research on Simulation and Leakage Location Methods of Ship Fire Pipe Damage, Wu Kai" proposed the use of neural networks to locate ship fire pipeline leakage. However, this neural network-based pipeline fault diagnosis model requires a large amount of data to meet complex calculations. The accuracy and reliability of the model are highly dependent on the quality of the input data. If the data source is unreliable or the data is incomplete, the generalization ability and accuracy of the model will be greatly affected, which makes the operation and maintenance costs of fault diagnosis very high in actual applications.

[0004] A method for establishing a model based on feature importance is proposed in a patent application document with publication number CN111860630A. In this scheme, feature data is initialized; feature data is sampled to form several groups of feature data combinations, each of which is a sub-model; model parameters of each group of feature models are set; the same model parameter range is set for each group of feature combinations; the model is trained to calculate the importance of each sub-model feature; the importance of the comprehensive weighted features is calculated using all sub-models; the feature importance is sorted to obtain a new importance ranking; according to the feature order, the top-ranked feature subset is taken to set the model parameter range again, and the model is retrained for modeling; however, this scheme is for model training and feature importance sorting for a single variable, which is not suitable for multi-variable scenarios, and the data is sorted solely by feature importance, which is prone to the problem of high feature importance but low correlation with the indicator of interest. The reason is that the learning model does not involve the calculation of the correlation between data. Summary of the invention

[0005] The present invention aims to solve the problem that the fire pipe network fault diagnosis model has a lot of training data and a poor diagnosis effect.

[0006] The present invention solves the above technical problems by the following technical means:

[0007] A fire protection pipe network fault inversion model training method is proposed, the method comprising:

[0008] Collecting the operating status data of the fire protection pipe network in normal state and fault state, wherein the status data includes pressure and flow;

[0009] Based on the operating status data, construct characteristic indicators based on data statistical characteristics and mark fault labels, the characteristic indicators including pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate;

[0010] Calculate the correlation coefficient between each characteristic index and the fault label to obtain the correlation coefficient matrix;

[0011] The correlation coefficient matrix is ​​processed by using a coupling model to obtain a feature importance ranking result;

[0012] Combined with the feature importance ranking result and the correlation coefficient matrix, feature combination data are selected as independent variables and input into the fire pipe network fault inversion model for training, and the best fault inversion model is screened out by adjusting parameters.

[0013] Furthermore, the collection of operation status data of the fire protection pipe network in a normal state and a fault state includes:

[0014] By using sensors arranged on the fire protection pipe network, the operating status data of the fire protection pipe network under normal conditions and the operating status data of the fire protection pipe network under fault conditions are collected.

[0015] Furthermore, after collecting the operation status data of the fire protection pipe network in a normal state and a fault state, the method further includes:

[0016] According to the fire pipe network pressure range and flow range specified in the General Specifications for Fire Protection Facilities, the operating status data is processed to obtain pressure data within the fire pipe network pressure range and flow data within the flow range.

[0017] Furthermore, after constructing the characteristic index based on the data statistical characteristics based on the running status data, the method further includes:

[0018] The characteristic index is dimensionlessly processed so that the characteristic index is distributed between [0, 1].

[0019] Furthermore, the calculation formula of the correlation coefficient is:

[0020]

[0021] Where X is the feature index, Y is the fault label, Cov(X,Y) is the covariance of the feature index and the fault label, and std X is the standard deviation of the characteristic index, std Y is the standard deviation of the fault labels, REXY is the correlation coefficient.

[0022] Furthermore, the coupling model includes a support vector machine sub-model, a logistic regression sub-model and a gradient boosting tree connected in sequence, wherein:

[0023] The support vector machine sub-model is used to process the correlation coefficient matrix to obtain a rough fault classification result;

[0024] The logistic regression sub-model is used to process the fault rough classification result to obtain the fault fine classification result;

[0025] The gradient boosting tree is used to determine the significance of feature indicators for the classification results according to the fault sub-classification results and the correlation coefficient matrix, and obtain a feature importance ranking result.

[0026] Furthermore, the feature importance ranking result and the correlation coefficient matrix are combined to select feature combination data as independent variables and input into the fire pipe network fault inversion model for training, and the best fault inversion model is screened out by adjusting parameters, including:

[0027] After comparing the feature indicators whose correlation coefficients are within the set first range and the feature indicators whose feature importance rankings are within the set second range, select the feature indicator with a stronger correlation or a higher importance ranking;

[0028] The selected characteristic indicators are weighted and combined according to the types of the characteristic indicators to obtain the characteristic combination data;

[0029] The fire protection pipe network fault inversion model is trained using the feature combination data as an independent variable, and the best fault inversion model is screened out by adjusting parameters.

[0030] Furthermore, the first range is [-1, -0.5] and [0.5, 1]; the second range is the first 30% of the feature importance ranking.

[0031] Further, the feature combination data is expressed as:

[0032] [αX i ,βX j ,γX k ,δX l ,∈X m , ζX n ]

[0033] Where, X i ,X j ,X k ,X l ,X m ,Xn They represent pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate respectively. α, β, γ, δ,∈, ζ represent the weights corresponding to the characteristic indicators respectively.

[0034] Furthermore, after combining the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables and inputting them into the fire protection pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters, the method further includes:

[0035] The performance of the optimal fault inversion model is evaluated, and the evaluation formula is:

[0036]

[0037] Where, T 1 is the number of results that the actual operating state matches the predicted operating state, F 1 is the number of actual fault prediction results, F 2 is the number of actual normal predicted fault results, T 2 is the number of results where the actual fault location matches the predicted fault location, F 3 is the number of fault location prediction errors, S 1 is the state prediction score, S 2 is the fault location prediction score, and S represents the model performance evaluation score.

[0038] The advantages of the present invention are:

[0039] (1) The present invention converts the operating status data of the fire protection pipe network into characteristic indicators based on data statistical characteristics and marks fault labels. The fault labels include variables such as pressure and flow, and calculates the pressure mean, flow mean, pressure change rate and flow change rate, which belongs to a typical multivariate scenario. Then, the correlation coefficient between each characteristic indicator and the fault label is calculated, and the multivariate correlation coefficient matrix is ​​calculated. Then, the correlation coefficient matrix is ​​processed by using a coupling model to obtain a feature importance ranking result. Combined with the feature importance ranking result and the correlation coefficient matrix, feature combination data is selected as an independent variable and input into the fire protection pipe network fault inversion model for training. This embodiment can effectively reduce the data volume required for model training and shorten the model training time by proposing a feature extraction and selection method suitable for fire protection pipe network pressure and flow data.

[0040] (2) The present invention proposes a coupling model suitable for predicting the operation status and fault location of the pipeline network, that is, constructs an SVM-LR-GBT learning model. In order to improve the accuracy of the classification model, multiple models are connected in series to classify the data, which overcomes the shortcomings of a single classification model, improves the accuracy of data classification, effectively improves the accuracy of model prediction and the generalization ability of the model, and reduces the operation and maintenance costs in practical applications.

[0041] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a fire protection pipe network fault inversion model training method proposed in one embodiment of the present invention;

[0043] Figure 2 It is a fire protection pipe network model of a converter station in one embodiment of the present invention;

[0044] Figure 3 is a flow chart of fire protection pipe network fault location inversion in one embodiment of the present invention;

[0045] Figure 4 is a bar graph of the relevance and importance of features in one embodiment of the present invention;

[0046] Figure 5 It is a fault location map finally predicted by the inversion technology in one embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, an embodiment of the present invention proposes a fire protection pipe network fault inversion model training method, the method comprising the following steps:

[0049] S10, collecting operation status data of the fire protection pipe network in normal state and fault state, wherein the status data includes pressure and flow;

[0050] S20, based on the operating status data, construct characteristic indicators based on data statistical characteristics and mark fault labels, the characteristic indicators including pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate;

[0051] It should be noted that this embodiment converts data into characteristic index data based on the statistical characteristics of the data. For multi-physical quantity data, such as pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate, they usually have different dimensions and significant differences in order of magnitude. If the data is directly processed, it is easy to cause the problem of big data eating up small data. This embodiment converts the data into characteristic index data based on the statistical characteristics of the data, so that the data of different physical quantities can be uniformly adjusted, which is convenient for data processing and avoids the problem of big data eating up small data.

[0052] S30, calculating the correlation coefficient between each characteristic indicator and the fault label to obtain a correlation coefficient matrix;

[0053] It should be noted that, since data is sorted solely based on feature importance, it is easy to have a problem where feature importance is high but the correlation with the indicator of interest is low. The reason is that the learning model does not involve the calculation of the correlation between data. To overcome this problem, this embodiment proposes a data screening method that integrates feature importance and correlation. On the one hand, it is necessary to screen data with high feature importance, and more importantly, it is necessary to screen data with high correlation with the indicator of interest. If and only if the feature importance and correlation are both at a high level, such data is valuable for model training, and unnecessary data samples can be further reduced.

[0054] S40, using the coupling model to process the correlation coefficient matrix to obtain a feature importance ranking result;

[0055] S50, combining the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables and inputting them into the fire protection pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters.

[0056] This embodiment converts the operating status data of the fire protection pipe network into feature indicators based on data statistical characteristics and marks fault labels, then calculates the correlation coefficient between each feature indicator and the fault label, uses a coupling model to process the correlation coefficient matrix, obtains the feature importance ranking result, combines the feature importance ranking result and the correlation coefficient matrix, selects feature combination data as independent variables and inputs them into the fire protection pipe network fault inversion model for training. This embodiment can effectively reduce the data volume required for model training and shorten the model training time by proposing a feature extraction and selection method suitable for fire protection pipe network pressure and flow data.

[0057] As a further preferred technical solution, the step S10: collecting the operating status data of the fire protection pipe network in a normal state and a fault state includes:

[0058] The sensors arranged on the fire protection pipe network are used to collect the operation status data of the fire protection pipe network under normal conditions and the operation status data of the fire protection pipe network under fault conditions.

[0059] Specifically, Figure 2 As shown in the figure, the fire protection pipe network of a converter station is 400 meters long, and sensors are evenly arranged at intervals of 10 meters to monitor the pressure and flow of the fire protection pipe network in real time. The pressure / flow data is uploaded every 1 second, and the data is collected according to the time series and position series. First, the pressure and flow data under the normal operation of the pipe network are collected. Then the valve is opened artificially to simulate the leakage state. The leakage position is located in the middle of the adjacent pressure / flow sensors. Only one leakage point is set in a set of test conditions to collect the pressure and flow data under the pipeline leakage state.

[0060] As a further preferred technical solution, in step S10: after collecting the operating status data of the fire protection pipe network in a normal state and a fault state, wherein the status data includes pressure and flow, the method further includes:

[0061] According to the fire pipe network pressure range and flow range specified in the General Specifications for Fire Protection Facilities, the operating status data is processed to obtain pressure data within the fire pipe network pressure range and flow data within the flow range.

[0062] It should be noted that the uploaded pressure / flow data is pre-processed. According to the General Specifications for Fire Protection Facilities, the pressure range of the fire protection pipe network is 0.14MPa-0.2MPa, and the flow range is 15L / s-50L / s. Delete the pressure and flow data that are too low and too high to ensure the accuracy, completeness and consistency of the training data.

[0063] This embodiment establishes a database after preprocessing to ensure the accuracy and completeness of the data and the suitability of the data range.

[0064] As a further preferred technical solution, after the step S20: constructing a characteristic index based on the statistical characteristics of the data based on the operating status data, the method further includes:

[0065] The characteristic index is dimensionlessly processed so that the characteristic index is distributed between [0, 1].

[0066] Specifically, this embodiment constructs new features based on the collected data, and the extracted features include pressure (Pres), flow (Flow), pressure mean (PresA), flow mean (FlowA), pressure change rate (PresC), and flow change rate (FlowC). The purpose is to reduce the impact of data distribution and overall changes on the training effect. All data are dimensionless, and the formula is:

[0067]

[0068] Where a is the original data, a max and a min are the maximum and minimum values ​​of the data respectively.

[0069] As a further preferred technical solution, the calculation formula of the correlation coefficient is:

[0070]

[0071] Where X is the feature index, Y is the fault label, Cov(X,Y) is the covariance of the feature index and the fault label, and std X is the standard deviation of the characteristic index, std Y is the standard deviation of the fault labels, RE XY is the correlation coefficient.

[0072] This embodiment constructs a correlation coefficient matrix based on the correlation coefficient between each characteristic indicator and the fault label:

[0073]

[0074] In the formula, x 11 ,…,x 16 are the correlation coefficients of pressure to pressure, flow, pressure mean, flow mean, pressure change rate, and flow change rate, respectively. 21 ,…,x 26 They are the correlation coefficients of flow to pressure, flow, pressure mean, flow mean, pressure change rate, flow change rate, and so on.

[0075] As a further preferred technical solution, Figure 3 As shown, the coupling model includes a support vector machine sub-model, a logistic regression sub-model and a gradient boosting tree connected in sequence, wherein:

[0076] The support vector machine sub-model is used to process the correlation coefficient matrix to obtain a rough fault classification result;

[0077] The logistic regression sub-model is used to process the rough fault classification result to obtain a detailed fault classification result;

[0078] The gradient boosting tree is used to determine the significance of feature indicators for the classification results according to the fault sub-classification results and the correlation coefficient matrix, and obtain a feature importance ranking result.

[0079] It should be noted that the constructed machine learning model couples the basic model according to the required model output effect (including the binary classification problem of whether a fault occurs and the regression prediction of pressure flow data and fault location). That is, according to the required output results - the operating status of the pipeline network and the prediction of the fault location, the support vector machine, logistic regression and gradient boosting tree models are coupled to obtain a learning model suitable for predicting the operating status and fault location of the pipeline network, which effectively improves the accuracy of model prediction and the generalization ability of the model, and reduces the operation and maintenance costs in practical applications.

[0080] As a further preferred technical solution, the step S50: combining the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables to input into the fire pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters, includes the following steps:

[0081] S51, after comparing the feature indicators with correlation coefficients within the set first range and the feature indicators with feature importance rankings within the set second range, select the feature indicator with stronger correlation or higher importance ranking;

[0082] Since the traditional method of sorting data based on feature importance alone is prone to the problem that the feature importance is high but the correlation with the indicator of interest is low, the reason is that the learning model does not involve the calculation of the correlation between data. To overcome this problem, this embodiment proposes a data screening method that integrates feature importance and correlation. On the one hand, it is necessary to screen data with high feature importance, and more importantly, it is necessary to screen data with high correlation with the indicator of interest. If and only if the feature importance and correlation are both at a high level, such data is valuable for model training, and unnecessary data samples can be further reduced.

[0083] S52, weighting and combining the selected feature indicators according to the types of the feature indicators to obtain the feature combination data;

[0084] S53, using the feature combination data as independent variables to train the fire protection pipe network fault inversion model, and selecting the best fault inversion model by adjusting parameters.

[0085] Accordingly, if Figure 3As shown, according to the optimal features and their combination selection in step S50, data that meets the requirements is selected from the data uploaded by the real-time sensors of the fire protection pipe network for processing, thereby reducing the data volume for model training; the data collected and processed in real time is input into the optimal fault inversion model to obtain the real-time operating status of the fire protection pipe network and the fault location under the fault state.

[0086] As a further preferred technical solution, the step S51: after comparing the feature indicators with correlation coefficients within the set first range and the feature indicators with feature importance rankings within the set second range, select the feature indicators with stronger correlation or higher importance ranking, specifically:

[0087] By calculating the correlation between features and output results and the feature importance analysis results built into the learning model, features with correlations between [-1, -0.5] and [0.5, 1] ​​and features ranked in the top 30% in importance are compared. Based on the principle of least optimal, features with stronger correlation or higher importance are selected.

[0088] Furthermore, by analyzing the proportion of different feature types after selection, the features are weighted according to the feature type and combined to obtain the feature combination for subsequent machine learning, which is expressed as:

[0089] αX i ,βX j ,γX k ,δX l ,∈X m , ζX n ]

[0090] Where, X i ,X j ,X k ,X l ,X m ,X n They represent pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate respectively. α, β, γ, δ,∈, ζ represent the weights corresponding to the characteristic indicators respectively.

[0091] Specifically, in this embodiment, there are a total of 246 sets of feature values, and 74 sets of features with correlations between [-1, -0.5] and [0.5, 1] ​​and features ranked in the top 30% of importance are selected, such as Figure 4As shown, after analysis, the final selection is: PresA7, FlowC22, FlowC39, PresA38, PresA16, PresA12, PresA4, FlowA37, FlowA31, PresC19, PresC27, FlowA34, PresA21, PresA15, FlowA40, FlowC14, PresA17, PresC8, FlowC33, PresC18, FlowC10, PresA29, PresA24, FlowC 41,PresA35,FlowA32,FlowC6,FlowC17,Flow8,FlowC1,FlowA3,Flow12,PresA23,FlowC25,PresC11,FlowA13,FlowA1,Flow6,Pres9,Flow18,PresC37,FlowC35,FlowA28,PresA36,PresC2,Flow21,FlowC9,Pres3,FlowC31,Pres11 are used as input features. The ratio of pressure, flow, pressure mean, flow mean, pressure change rate, and flow change rate among the fifty features is 3:5:12:10:7:13. The input features are weighted and combined according to this ratio.

[0092] The model was randomly divided into 80% training set and 20% validation set. The optimal parameter setting of the basic model was determined by cross-validation method and the prediction performance of the model was evaluated. For the support vector machine model (SVM), the radial basis function kernel (RBF) was selected, the kernel function parameter was set to 0.1, and the penalty parameter (epsilon) was set to 1; for the logistic regression model (LR), L1 regularization (Lasso) was selected, and the solver parameter was liblinear; for the gradient boosted tree (GBT), n=50 was set, and the depth of the tree was set to 9. The prediction score S=0.98 under this parameter setting is the optimal fault inversion model that meets the requirements.

[0093] Based on the optimal fault inversion model, the pipeline fault data of a certain time period was selected. After selecting the input data according to the feature extraction method, the fault status and location of the pipeline network in this time period were successfully predicted. Figure 5 shown.

[0094] As a further preferred technical solution, in step S50: after combining the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables and inputting them into the fire pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters, the method further includes:

[0095] The performance of the optimal fault inversion model is evaluated, and the evaluation formula is:

[0096]

[0097] Where, T 1 is the number of results that the actual operating state matches the predicted operating state, F 1 is the number of actual fault prediction results, F 2 is the number of actual normal predicted fault results, T 2 is the number of results where the actual fault location matches the predicted fault location, F 3 is the number of fault location prediction errors, S 1 is the state prediction score, S 2 is the fault location prediction score, S represents the model performance evaluation score, and the closer the model score S is to 1, the better the model performance is.

[0098] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0099] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0100] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A fire protection pipe network fault inversion model training method, characterized in that: The method comprises: Collecting the operating status data of the fire protection pipe network in normal state and fault state, wherein the status data includes pressure and flow; Based on the operating status data, construct characteristic indicators based on data statistical characteristics and mark fault labels, the characteristic indicators including pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate; Calculate the correlation coefficient between each characteristic index and the fault label to obtain the correlation coefficient matrix; The correlation coefficient matrix is ​​processed by using a coupling model to obtain a feature importance ranking result; Combined with the feature importance ranking result and the correlation coefficient matrix, feature combination data are selected as independent variables and input into the fire pipe network fault inversion model for training, and the best fault inversion model is screened out by adjusting parameters.

2. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: The collection of operating status data of the fire protection pipe network in normal state and fault state includes: The sensors arranged on the fire protection pipe network are used to collect the operation status data of the fire protection pipe network under normal conditions and the operation status data of the fire protection pipe network under fault conditions.

3. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: After collecting the operation status data of the fire protection pipe network in a normal state and a fault state, the method further includes: According to the fire pipe network pressure range and flow range specified in the General Specifications for Fire Protection Facilities, the operating status data is processed to obtain pressure data within the fire pipe network pressure range and flow data within the flow range.

4. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: After constructing the characteristic index based on the data statistical characteristics based on the running status data, the method further includes: The characteristic index is dimensionlessly processed so that the characteristic index is distributed between [0, 1].

5. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: The calculation formula of the correlation coefficient is: Where X is the feature index, Y is the fault label, Cov(X,Y) is the covariance of the feature index and the fault label, and std X is the standard deviation of the characteristic index, std Y is the standard deviation of the fault labels, RRE XY is the correlation coefficient.

6. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: The coupling model includes a support vector machine sub-model, a logistic regression sub-model and a gradient boosting tree connected in sequence, wherein: The support vector machine sub-model is used to process the correlation coefficient matrix to obtain a rough fault classification result; The logistic regression sub-model is used to process the rough fault classification result to obtain a detailed fault classification result; The gradient boosting tree is used to determine the significance of feature indicators for the classification results according to the fault sub-classification results and the correlation coefficient matrix, and obtain feature importance ranking results.

7. The fire protection pipe network fault inversion model training method according to claim 1, characterized in that: The combining of the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables and inputting them into the fire pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters, includes: After comparing the feature indicators whose correlation coefficients are within the set first range and the feature indicators whose feature importance rankings are within the set second range, select the feature indicator with a stronger correlation or a higher importance ranking; The selected characteristic indicators are weighted and combined according to the types of the characteristic indicators to obtain the characteristic combination data; The fire protection pipe network fault inversion model is trained using the feature combination data as an independent variable, and the best fault inversion model is screened out by adjusting parameters.

8. The fire protection pipe network fault inversion model training method according to claim 7, characterized in that: The first range is [-1, -0.5] and [0.5, 1]; the second range is the first 30% of the feature importance ranking.

9. The fire protection pipe network fault inversion model training method according to claim 7, characterized in that: The feature combination data is expressed as: [αX i ,βX j ,γX k ,δX l ,∈X m ,ζX n ] Where, X i ,X j ,X k ,X l ,X m ,X n They represent pressure, flow, pressure mean, flow mean, pressure change rate and flow change rate respectively. α, β, γ, δ,∈, ζ represent the weights corresponding to the characteristic indicators respectively.

10. The fire protection pipe network fault inversion model training method according to any one of claims 1 to 9, characterized in that: After combining the feature importance ranking result and the correlation coefficient matrix, selecting feature combination data as independent variables and inputting them into the fire pipe network fault inversion model for training, and selecting the best fault inversion model by adjusting parameters, the method further includes: The performance of the optimal fault inversion model is evaluated, and the evaluation formula is: Where T1 is the number of results where the actual operating state is consistent with the predicted operating state, F1 is the number of results where the actual fault prediction is normal, F2 is the number of results where the actual normal prediction is faulty, T2 is the number of results where the actual fault location is consistent with the predicted fault location, F3 is the number of results where the fault location prediction is wrong, S1 is the state prediction score, S2 is the fault location prediction score, and S represents the model performance evaluation score.

Images