A cutter head face instability prediction method based on an LSTM deep learning model

Abstract

The application discloses a cutter head face instability prediction method based on an LSTM deep learning model, and belongs to the technical field of tunnel construction safety early warning. The method fuses multi-source sensor data acquisition, feature saliency analysis, hybrid database architecture and time series prediction modeling technology, and constructs a complete shield construction safety early warning system. The system adopts a hybrid architecture combining a structured and document heterogeneous storage mode, establishes a double-channel processing mechanism of equipment state data and time series feature data, realizes efficient integration of data through sliding window sampling and standardization preprocessing. Based on the time series prediction algorithm of the LSTM deep learning model, in combination with the feature importance screening mechanism, the prediction accuracy and real-time performance of the cutter head face instability state are significantly improved. The experimental results show that the method can effectively capture the dynamic characteristic changes in the shield construction process, and provides a reliable intelligent early warning solution for the safe construction of tunnel projects.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent construction monitoring of underground engineering, and particularly relates to a cutter head face instability prediction method based on an LSTM deep learning model. BACKGROUND

[0002] Shield construction is the core link of tunnel engineering, and the instability of the cutter head face is one of the causes of major engineering accidents such as ground subsidence and equipment damage. With the continuous development of urban underground space and the increasing complexity of geological conditions, the safety challenges faced by shield construction are becoming more and more severe. Traditional safety warning methods rely on experience threshold judgment or single-dimensional parameter monitoring. This static evaluation mode based on artificial experience is difficult to effectively cope with the dynamic instability risk caused by the coupling of multiple physical fields (including stress field, seepage field, temperature field, etc.) under complex geological conditions, resulting in that the warning accuracy and timeliness are difficult to meet the safety needs of modern shield construction.

[0003] In recent years, with the rapid development of Internet of Things technology and artificial intelligence, the field of shield construction safety monitoring has ushered in new opportunities for technological innovation. By deploying a multi-source sensor network, key parameters such as hydraulic system pressure, thrust cylinder pressure, cutter torque, and attitude inclination can be collected in real time, providing a rich data basis for instability prediction. At the same time, the breakthrough of deep learning technology in time series data processing, especially the development of long short-term memory network (LSTM) and attention mechanism, provides a new technical path for modeling the nonlinear and time-varying characteristics in shield construction. These technological advances make it possible to establish an intelligent prediction system based on multi-source data fusion, which is expected to achieve accurate prediction and real-time warning of the instability of the cutter head face.

[0004] However, the existing technology still faces many challenges in the process of engineering application. At the data level, the traditional relational database architecture is difficult to adapt to the mixed storage needs of high-frequency time series data and unstructured equipment data in shield construction, which restricts the real-time performance of data fusion analysis; at the algorithm level, methods based on statistical process control or shallow machine learning models have obvious limitations in time series feature extraction and long-term dependency modeling; at the engineering application level, existing deep learning models generally have imperfect feature selection mechanisms, insufficient model generalization ability, and lack of engineering significance verification, resulting in high false alarm rates. SUMMARY

[0005] The application provides a cutter head face instability prediction method based on an LSTM deep learning model to solve the technical problems in the background art.

[0006] This invention is achieved through the following technical solution: a method for predicting instability of the cutting face of a cutterhead based on an LSTM deep learning model, comprising the following steps:

[0007] A data acquisition system was constructed to obtain multi-source heterogeneous data by acquiring state data from different sensors at a predetermined sampling frequency; a feature engineering model with a dual screening mechanism was established, and seven sets of key feature parameters were selected using the feature engineering model.

[0008] A hybrid database storage architecture is designed, which takes into account both state metadata and time-series characteristics, and structures the storage of seven sets of key feature parameters. 3D feature matrix, The number of time-series records selected; the... The 1D feature matrix is ​​standardized to obtain the standardized feature matrix;

[0009] Construct an LSTM time series prediction model, taking the standardized feature matrix as input, and outputting the model. dimensional prediction probability vector , The timing record corresponds to the probability of instability at the face of the cutterhead.

[0010] Based on the above dimensional prediction probability vector The percentage of abnormal results was calculated. According to the percentage of abnormal results Determine the warning level.

[0011] In a further embodiment, the status data includes non-fault status data and fault status data;

[0012] Correspondingly, the multi-source heterogeneous data includes: hydraulic system pressure, propulsion cylinder pressure, cutter head torque, and attitude tilt angle from different sensors.

[0013] In a further embodiment, the feature engineering model employs a dual screening mechanism combining independent sample testing and effect size analysis, wherein the dual screening mechanism consists of significance level screening and mean difference screening.

[0014] Correspondingly, the seven sets of key characteristic parameters include: working chamber pressure. Gear oil gas seal pressure HBW grease pressure of main bearing Cutter head torque Main drive motor torque Main drive motor speed and main drive motor speed .

[0015] In further embodiments, the screening process of the seven groups of key characteristic parameters is as follows:

[0016] The relevant parameters of the collected non-fault state data and fault state data are subjected to independent sample test analysis, the significant difference values between the non-fault state data and the fault state data of each sensor are calculated, and the analyzable samples are screened out by using the significance level screening;

[0017] The analyzable samples are further quantified by effect size analysis to obtain mean difference values, and the seven groups of key characteristic parameters are further screened out by using the mean difference screening.

[0018] In further embodiments, the design process of the mixed database storage architecture includes:

[0019] A structured metadata storage module with a mapping relationship is constructed based on state metadata, the state metadata including: ring number , propulsion rate and collection time identifier ;

[0020] The structured metadata storage module has a hierarchical index structure at the same time, and the hierarchical index structure is established based on the ring number and the collection time identifier ;

[0021] A time sequence document storage module is constructed, and a document storage structure is used to nest store the time sequence characteristics of the seven groups of key characteristic parameters; based on the ring number , a hash balancing weight factor is introduced on the document index strategy to realize distributed load balancing.

[0022] In further embodiments, The processing process of the standardization processing of the feature matrix is as follows:

[0023] For the random fluctuations and abnormal spikes of the dimension feature matrix, an offset rate index in a dynamic window period is defined; the abnormal segments are identified by using the offset rate index , and the abnormal segments are repaired by a median value reconstruction strategy;

[0024] For the sequence missing caused by the collection discontinuity or jump point of the dimension feature matrix, a weighted linear interpolation strategy based on time sequence is used for sequence missing repair;

[0025] Based on the standardization method of the sliding mean and the fluctuation amplitude , the The characteristic matrix is subjected to proportional difference standardization to obtain a normalized characteristic matrix.

[0026] In a further embodiment, the construction process of the LSTM time series prediction model comprises:

[0027] The network structure of the LSTM time series prediction model is established, and the network structure comprises an input layer, an LSTM layer, a Dropout layer and an output layer.

[0028] The activation function of the LSTM layer adopts a hyperbolic tangent function to compress the neuron output to interval to enhance the nonlinear ability of feature expression; meanwhile, the kernel weight matrix is subjected to regularization constraint;

[0029] The neuron random dropout rate of the Dropout layer is set to 0.2 to inhibit the synergistic adaptation between neurons.

[0030] The output layer adopts a Sigmoid activation function to map the prediction value to interval to obtain a prediction probability value.

[0031] In a further embodiment, the calculation formula of the abnormal result proportion value is as follows:

[0032] , wherein is the abnormal result proportion value, , , is an indicator function of .

[0033] In a further embodiment, the design process of the hybrid database storage architecture further comprises:

[0034] An incremental synchronization strategy oriented to the collection time identifier is adopted to construct a data update module, and through setting a dynamic time window, low-frequency archiving of historical state and high-frequency pushing of real-time state are realized.

[0035] The following data query mode is periodically executed to realize segmented query and extraction of multi-dimensional feature data:

[0036] The latest data query is time > last p-time , wherein time is a time stamp, last p-time is a time stamp of the last updated data.

[0037] In further embodiments, a training optimization of the LSTM time series prediction model is also included, which comprises a sample weight adjustment strategy and a training control strategy.

[0038] The present application has the following beneficial effects: The present application proposes a cutter head face instability prediction method based on an LSTM deep learning model, which effectively solves the real-time early warning problem of face instability in shield construction through multi-source data fusion and deep time series modeling technology. The following is the technical innovation and engineering application value analysis of the present application:

[0039] Firstly, in terms of data fusion and feature engineering, the present application innovatively constructs a multi-source heterogeneous data acquisition and processing system. Through the coordinated monitoring of four types of key parameters such as hydraulic system pressure and thrust cylinder pressure, multi-dimensional perception of shield construction state is achieved. Based on the dual screening mechanism of independent sample inspection and effect size analysis, seven key parameters with engineering significance are scientifically selected from multiple initial features, significantly improving the quality of model input features. The design of the hybrid database architecture realizes efficient management of high-frequency time series data and equipment state data, and the data query response time meets the timeliness requirements of real-time early warning.

[0040] Secondly, in terms of model construction and optimization, the LSTM time series prediction model designed by the present application has the following technical advantages:

[0041] 1) The LSTM layer with regularization constraint (λ = 0.001) effectively prevents model overfitting and improves generalization performance;

[0042] 2) The dynamic class weight strategy solves the problem of scarcity of fault samples in engineering data, effectively improving the recognition rate of minority class samples;

[0043] 3) The early stopping mechanism and dynamic learning rate adjustment strategy improve the model training efficiency and reduce resource consumption. According to actual engineering verification, the model achieves an accuracy of 89.7% on the test set. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a flowchart of the cutter head face instability prediction method based on the LSTM deep learning model of embodiment 1.

[0045] Figure 2 is a flowchart of the standardization process of embodiment 1.

[0046] Figure 3 is a network structure diagram of the LSTM time series prediction model of embodiment 1. DETAILED DESCRIPTION

[0047] The present application will be further described below in conjunction with the drawings and examples of the specification.​

[0048] Example 1

[0049] like Figure 1 As shown, a method for predicting instability of the cutting face of a cutterhead based on an LSTM deep learning model includes the following steps:

[0050] A data acquisition system was constructed to obtain multi-source heterogeneous data by acquiring state data from different sensors at a predetermined sampling frequency; a feature engineering model with a dual screening mechanism was established, and seven sets of key feature parameters were selected using the feature engineering model.

[0051] A hybrid database storage architecture is designed, which takes into account both state metadata and time-series characteristics, and structures the storage of seven sets of key feature parameters. 3D feature matrix, The number of time-series records selected; the... The dimensional feature matrix is ​​standardized to obtain the standardized feature matrix; in this embodiment, N is taken as 100, that is... 3D feature matrix.

[0052] Construct an LSTM time series prediction model, taking the standardized feature matrix as input, and outputting the model. dimensional prediction probability vector , The timing record corresponds to the probability of instability at the face of the cutterhead.

[0053] Based on the above dimensional prediction probability vector The percentage of abnormal results was calculated. According to the percentage of abnormal results Determine the warning level and provide the corresponding warning decision.

[0054] Furthermore, the formula for calculating the percentage of abnormal results is as follows:

[0055] In the formula, This represents the percentage of abnormal results. , , for Indicator functions, For time-series records The corresponding probability of instability of the cutting face.

[0056] Further warning level determination: If The warning level is red, and the corresponding decision is to immediately shut down the system for inspection; if The warning level is yellow, and the corresponding decision is a warning alert; if If the warning level is green, the corresponding decision is normal construction.

[0057] The state data includes non-fault state data and fault state data. Correspondingly, the multi-source heterogeneous data includes hydraulic system pressure, propelling oil cylinder pressure, cutterhead torque and attitude inclination angle from different sensors. Therefore, the data acquisition system mentioned in this embodiment is a real-time acquisition of four types of sensor data of hydraulic system pressure, propelling oil cylinder pressure, cutterhead torque and attitude inclination angle through the shield tunneling machine interface, and the sampling frequency is set to no less than 1 Hz.

[0058] The feature engineering model in this embodiment is a double screening mechanism combining independent sample test and effect size analysis, which uses the double screening mechanism to respectively perform significance and engineering significance discrimination processing. Based on the existing statistical test method, combined with the actual working condition of the tunneling site, the feature is state "characterized" processing. The seven groups of key feature parameters screened based on the double screening mechanism include: working chamber pressure , gear oil gas seal pressure , main bearing HBW grease pressure , cutterhead torque , main drive motor torque , main drive motor speed and main drive motor speed .

[0059] Based on this, the double screening mechanism of the embodiment is a significant level screening and a mean difference screening. Further, the screening process of the seven groups of key feature parameters is as follows:

[0060] The independent sample test analysis is performed on the related parameters of the collected non-fault state data and fault state data, the bilateral test is used to calculate the significant difference value between the non-fault state data and the fault state data of each sensor, and the significant level screening is used to screen out the analyzable samples. For example, the significant difference threshold is set to 0.05, then , it is considered that the difference between the non-fault state data and the fault state data is significant, and can be used as an analyzable sample.

[0061] On this basis, the effect size analysis is used to further quantify the mean difference value of the analyzable sample, and the mean difference screening is used to further screen out the seven groups of key feature parameters. In a further embodiment, the calculation formula of the mean difference value D is as follows:

[0062] ; wherein, represents the fault state mean, represents the non-fault state mean. According to the actual calculation result, the feature parameters with a difference degree exceeding 15% are screened out.

[0063] For example, the parameter main drive ring number is found to fluctuate greatly in the fault state, and the mean difference is significantly more than 15%, so it is identified as a key feature. For example, the hydraulic cylinder pressure fluctuates in a narrow range in the non-fault state, but shows an abnormal peak in the fault state, so the screening standard is adjusted to a lower difference threshold (such as 10%) to ensure that it is effectively extracted. For example, the parameter main drive ring number is found to fluctuate greatly in the fault state, and the mean difference is significantly more than 15%, so it is identified as a key feature. For example, the hydraulic cylinder pressure fluctuates in a narrow range in the non-fault state, but shows an abnormal peak in the fault state, so the screening standard is adjusted to a lower difference threshold (such as 10%) to ensure that it is effectively extracted.

[0064] For example, the parameter main drive ring number is found to fluctuate greatly in the fault state, and the mean difference is significantly more than 15%, so it is identified as a key feature. For example, the hydraulic cylinder pressure fluctuates in a narrow range in the non-fault state, but shows an abnormal peak in the fault state, so the screening standard is adjusted to a lower difference threshold (such as 10%) to ensure that it is effectively extracted. For example, the parameter main drive ring number is found to fluctuate greatly in the fault state, and the mean difference is significantly more than 15%, so it is identified as a key feature. For example, the hydraulic cylinder pressure fluctuates in a narrow range in the non-fault state, but shows an abnormal peak in the fault state, so the screening standard is adjusted to a lower difference threshold (such as 10%) to ensure that it is effectively extracted.

[0065] In order to optimize the feature set, the feature distribution analysis is also performed, and the kernel density estimation (KDE) method is used to visualize the distribution difference of each parameter in the two states. The analysis results show that the selected features have obvious distribution difference in the fault and non-fault states, verifying the rationality of the feature selection. The final retained feature set can effectively distinguish the instability state of the cutter head working surface through actual verification.

[0066] Further, the hybrid database storage architecture adopts a scheme combining structured and document type heterogeneous storage methods, and constructs an optimized storage module for device state metadata and time series feature data respectively.

[0067] Therefore, the design process of the hybrid database storage architecture of the embodiment includes:

[0068] A structured metadata storage module with a mapping relationship is constructed based on state metadata, and the state metadata includes: ring number , push rate and acquisition time identifier . Among them, the ring number adopts 64-bit integer encoding to support large interval numbering, selects double-precision floating-point type to improve the analysis accuracy of the low-speed section, and the acquisition time identifier performs millisecond-level truncation normalization processing on Unix the timestamp.

[0069] The structured metadata storage module also has a hierarchical index structure, and the hierarchical index structure is based on the ring number and the acquisition time identifier ​The structured metadata storage module table structure generates a persistent storage model through object relationship mapping, and generates a distributed storage model based on ring number and collection time identifier Combining the construction of hierarchical index structure, for improving the performance of query by time window.

[0070] The time sequence document storage module is constructed, and the time sequence characteristics of seven groups of key feature parameters are nested and stored by using the document storage structure; based on ring number , a hash balancing weight factor is introduced on the document index strategy to realize distributed load balancing.

[0071] It is worth mentioning that the nested storage of the document storage structure is that each document is bound with a unique ring number identifier, and the internal nested property The cluster stores seven types of dynamic feature trajectories, and the trajectory data is stored in key-value format, supporting floating point sequence compression and sampling interpolation processing to adapt to different sensor frequency requirements.

[0072] On the document index strategy, a hash balancing weight factor is introduced, and a dynamic sharding key is generated combined with the feature fluctuation frequency to improve the load balancing of distributed retrieval.

[0073] The structured metadata storage module and the time sequence document storage module are indexed and mapped one by one based on the ring number In other words, by constructing a double-channel index mechanism based on the ring number field, efficient association retrieval between structured and time sequence data modules is realized. Among them, the structured module establishes a joint index on the ring number and time field, enhancing the response speed of continuous process scheduling analysis; the unstructured module adopts a load balancing type distribution mapping strategy based on the ring number dimension, adapting to the horizontal expansion requirement of feature sequence.

[0074] In another embodiment, the design process of the hybrid database storage architecture also includes:

[0075] An incremental synchronization strategy oriented to collection time identifier is adopted to construct a data update module, and by setting a dynamic time window, low-frequency archiving of historical state and high-frequency pushing of real-time state are realized;

[0076] The following data query method is periodically executed to realize segmented query and extraction of multi-dimensional feature data:

[0077] The latest data query is { time > last p-time}, wherein, time is a timestamp, last p-time is the timestamp of the last updated data.

[0078] In summary, the hybrid database storage architecture adopts a separate read-write architecture as a whole, realizing decoupling of responsibilities between data collection and query. The structured metadata storage module improves retrieval capability in a high-concurrency environment by constructing a read standby channel through multi-node redundancy. The time-series document storage module enhances data persistence and recovery capability by introducing automatic replica switchover and state detection mechanisms. The measured results show that the architecture can stably support multi-channel concurrent processing and high-speed writing under typical shield construction data load, meeting the systematic needs of real-time monitoring.

[0079] In combination Figure 2 , the processing procedure of the standardization processing includes: multi-stage fluctuation detection, sequence missing repair and scale normalization processing. It is further understood that, for the random fluctuation and abnormal peak of the feature matrix of dimension, the drift rate index in the dynamic window period is defined ; the abnormal segment is identified by using the drift rate index , and the abnormal segment is repaired by a median value reconstruction strategy to maintain the consistency of local trend. The calculation formula of the drift rate index is as follows:

[0080] ; wherein, and are the neighborhood mean and standard deviation centered on the current time frame , respectively, is the sensor raw data value collected at the current time frame .

[0081] Further, if , the data collected at the current time frame t is regarded as an anomaly, is an empirical threshold, and its value is 2.5-3.0.

[0082] Secondly, for the sequence missing caused by the collection discontinuity or jump point of the feature matrix of dimension, a weighted linear interpolation strategy based on time sequence is used for sequence missing repair; wherein, the calculation formula of the weighted linear interpolation strategy is as follows:

[0083] ; wherein, is the sensor raw data value collected at the historical time frame t-1, is the distance weight of the historical time frame t-1, is the distance weight of the current time frame t. The setting of the distance weight and the distance weight combines dynamic adjustment of sampling frequency and historical stability, and enhances difference smoothness.

[0084] Finally, considering the inconsistency of different sensor parameter units and magnitudes, in order to ensure the comparability of the feature fusion model, the ratio difference standardization processing is performed on the multi-dimensional feature matrix based on the standardization method of sliding mean and fluctuation amplitude to obtain a standardized feature matrix.

[0085] Further, the calculation formula of the ratio difference standardization processing is as follows:

[0086] ; is the standardized data point, that is, the standardized feature matrix. The standardization operation is independently performed on each monitoring window to improve the sensitivity of the model to local instability signs.

[0087] In further embodiments, in combination with Figure 3 , the construction process of the LSTM time series prediction model includes:

[0088] establishing a network structure of the LSTM time series prediction model, the network structure including: an input layer, an LSTM layer, a Dropout layer, and an output layer;

[0089] The input layer is used to receive the standardized feature matrix.

[0090] The activation function of the LSTM layer adopts the hyperbolic tangent function, which compresses the neuron output to the interval to enhance the nonlinear ability of feature expression; at the same time, the kernel weight matrix is subjected to regularization constraint;

[0091] The neuron random dropout rate of the Dropout layer is set to 0.2 to inhibit the synergistic adaptation between neurons;

[0092] The output layer adopts the Sigmoid activation function to map the prediction value to the interval to obtain the prediction probability value.

[0093] In order to realize the above functions, the specific parameters of the network structure of the embodiment are shown in Table 1.

[0094] Table 1

[0095]

[0096] Further analysis in combination with Table 1 shows that the LSTM layer in the embodiment contains 100 hidden units, and the calculation formula of the regularization constraint is as follows:

[0097] , . Wherein, ​Weight matrix The weights in The number of the weight. For the total weight, For regularization strength, This is the regularization loss function.

[0098] Furthermore, to prevent overfitting, the mathematical expression for the Dropout layer is:

[0099] In the formula, This is the output result after processing by the Dropout layer. As input to the Dropout layer, The expression is based on the combination of elements. For the mask vector, Represents the mask vector The element It follows a Bernoulli distribution, and The probability of getting 0 is 0.2.

[0100] Correspondingly, the output layer uses the Sigmoid activation function as follows:

[0101] In the formula, Given the input features, For the label of the sample, The weighted sum of the input features, Given input features In the case of [condition], the conditional probability that the sample belongs to category 1.

[0102] Finally, this embodiment also includes: training optimization of the LSTM time series prediction model. The training optimization method includes: sample weight adjustment strategy and training control strategy, such as optimizing the parameters in Table 1, and optimizing the number of training iterations, etc.

[0103] In summary, this embodiment achieves high-precision prediction of the tunnel face instability state during shield tunneling by constructing a multi-source data fusion architecture, optimizing the feature engineering process, and improving the LSTM network structure. Specifically, a multi-sensor collaborative sampling mechanism is employed at the data acquisition layer to ensure the comprehensiveness and real-time nature of data acquisition; a dual screening strategy combining saliency analysis and effect size assessment is introduced at the feature processing layer, significantly improving the engineering applicability of feature selection; and at the model architecture layer, improvements to the LSTM network structure and training strategy, including the use of dynamic class weight balancing and early stopping mechanisms, effectively solve the class imbalance problem in engineering data.

Claims

1. A method for predicting instability of a cutterhead face based on an LSTM deep learning model, characterized in that, Comprise the following steps: The data acquisition system is constructed to obtain multi-source heterogeneous data by acquiring state data of different sensors at a predetermined sampling frequency; a feature engineering model with a double screening mechanism is established to screen seven groups of key characteristic parameters from non-fault state data and fault state data by using the feature engineering model; wherein, the feature engineering model is a double screening mechanism combining independent sample test and effect size analysis, and the double screening mechanism is a significant level screening and a mean difference screening; the seven groups of key characteristic parameters include: working chamber pressure , gear oil gas seal pressure , main bearing HBW grease pressure , cutter head torque , main drive motor torque , main drive motor speed , and main drive motor speed ; and the screening process of the seven groups of key characteristic parameters is as follows: The collected relevant parameters are analyzed by independent sample test, the significant difference value between the non-fault state data and the fault state data of each sensor is calculated, and the analyzable samples are screened out by using the significant level screening; The analyzable samples are further quantified by effect size analysis to obtain the mean difference value, and seven groups of key characteristic parameters are further screened out by using the mean difference value screening; The mixed database storage architecture is designed, and seven groups of key characteristic parameters are stored in a structured manner considering state metadata and time sequence characteristics dimensional feature matrix, a selected time sequence record quantity; the dimensional feature matrix is standardized to obtain a standardized feature matrix; and the design process of the mixed database storage architecture comprises: A structured metadata storage module with mapping relationship is constructed based on state metadata, the state metadata includes: ring number , pushing rate and collection time identifier ; The structured metadata storage module has a hierarchical index structure simultaneously, which is based on ring number With the acquisition time identifier Established; The timing document storage module is constructed, and the timing characteristics of seven groups of key characteristic parameters are nested and stored by using a document storage structure; and a ring number A hash balance weight factor is introduced in the document index strategy to realize distributed load balance. constructing an LSTM time series prediction model, taking the standardized feature matrix as input, outputting a stability probability vector for the time series record N , is a stability probability vector for the time series record N based on the vector of predictive probabilities calculate an abnormal result proportion value , according to the abnormal result proportion value determine an early warning level; The abnormal result proportion value The calculation formula is as follows: wherein is the fraction of abnormal results, , , is the indicator function of 2. The method according to claim 1, wherein, The state data includes non-fault state data and fault state data; Correspondingly, the multi-source heterogeneous data includes: hydraulic system pressure, propelling oil cylinder pressure, cutter torque and attitude inclination angle from different sensors.

3. The method of claim 1, wherein the method is characterized by: The processing procedure of the standardization of the feature matrix is as follows: For In view of the random fluctuations and abnormal spikes of the feature matrix, a drift rate index in a dynamic window period is defined ; using the drift rate index An abnormal section is identified, and the abnormal section is repaired through a median value reconstruction strategy For The sequence missing caused by the collection discontinuity or jump point of the feature matrix is repaired by using a time sequence-based weighted linear interpolation strategy. Based on moving average and fluctuation range The standardization method, for The proportional difference standardization process is performed on the feature matrix to obtain the standardized feature matrix.

4. The method of claim 1, wherein the method is characterized by, The construction process of the LSTM time series prediction model comprises: The network structure of the LSTM time series prediction model is established, and the network structure comprises: an input layer, an LSTM layer, a Dropout layer and an output layer; The activation function of the LSTM layer adopts a hyperbolic tangent function to compress the neuron output to the interval [-1, 1] to enhance the nonlinear ability of feature expression; meanwhile, the kernel weight matrix is subjected to a regularization constraint; The neuron random dropout rate of the Dropout layer is set to 0.2 to inhibit the synergistic adaptation between neurons; The output layer employs a Sigmoid activation function to map the predicted values to The interval gets the predicted probability value.

5. The method of claim 1, wherein the method is characterized by: The design process of the hybrid database storage architecture further comprises: Adopting the incremental synchronization strategy of collection time mark The data updating module is constructed by adopting the incremental synchronization strategy of collection time mark, and through setting a dynamic time window, low-frequency archiving of historical state and high-frequency pushing of real-time state are realized. Periodically execute the following data query mode to realize the segmented query and extraction of multi-dimensional feature data: New data search { time > last p-time}, wherein, time is a time stamp, last p-time is a time stamp of the last update of the data.

6. The method of claim 1, wherein the method is characterized by: It also includes the training and optimization of the LSTM time series prediction model, and the training and optimization method comprises: a sample weight adjustment strategy and a training control strategy.

Images