Neural network-based derrick fatigue damage multi-dimensional diagnosis and early warning method

Abstract

The invention discloses a derrick fatigue damage multi-dimensional diagnosis and early warning method based on a neural network, and belongs to the technical field of safety monitoring. The method specifically comprises the following steps: S1, multi-source data distributed acquisition: acquiring multi-source heterogeneous data according to a set acquisition frequency, and storing and defining a unified data format by adopting a time sequence database; s2, multi-source heterogeneous data hierarchical preprocessing: performing adaptive noise cancellation and wavelet threshold denoising dual-stage processing on dynamic signals, standardizing and unifying dimensions of static data through the Z-score safety monitoring technical field, and removing abnormal data in combination with a 3 sigma safety monitoring technical field criterion and a sliding window; the limitation that only dynamic physical data are focused in the prior art is broken through, dynamic data such as vibration, stress strain and the like, environmental data such as temperature, humidity, wind speed and the like, and historical data such as accumulated use time of the derrick, hoisting load frequency and the like are synchronously collected, and the real fatigue state of the derrick can be comprehensively reflected by combining multi-dimensional data.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology, and in particular to a method for multi-dimensional diagnosis and early warning of derrick fatigue damage based on neural networks. Background Technology

[0002] As core load-bearing equipment in oil extraction, construction, and mining, derricks operate in complex and harsh environments. In oil extraction, they must withstand alternating loads from hoisting drill strings and environmental stresses from wind and temperature changes; in construction, they face impact loads from frequent material lifting and lowering. These factors make derrick steel structures prone to cumulative fatigue damage. Industry statistics show that over 60% of derrick safety accidents are caused by undetected fatigue damage, leading not only to equipment failure and production interruptions but also potential injuries and fatalities. Therefore, accurate diagnosis and early warning of derrick fatigue damage are crucial for ensuring industrial production safety. Existing technologies primarily focus on collecting dynamic physical data of the derrick (such as vibration, stress, and strain), neglecting the impact of environmental factors and historical equipment data on fatigue damage. This fails to comprehensively reflect the true fatigue state of the derrick, resulting in a "data blind spot" in subsequent diagnosis. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a multi-dimensional diagnosis and early warning method for derrick fatigue damage based on neural networks; it can solve the problems of single data acquisition dimension and incomplete information coverage in the existing technology.

[0004] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a method for multi-dimensional diagnosis and early warning of derrick fatigue damage based on neural networks, the method specifically includes the following steps:

[0005] S1. Distributed acquisition of multi-source data: Collect heterogeneous data from multiple sources according to the set acquisition frequency, store the data in a time-series database and define a unified data format;

[0006] S2. Multi-source heterogeneous data hierarchical preprocessing: Dynamic signals are processed in two stages: adaptive noise cancellation and wavelet threshold denoising. Static data are standardized and unified in the Z-score safety monitoring technology field, and abnormal data are removed by combining the criteria of the 3σ safety monitoring technology field with the sliding window.

[0007] S3. Deep feature extraction from multi-source data: The preprocessed multi-source data is sequentially subjected to single-source feature decomposition, cross-domain feature association, and feature hierarchical aggregation to extract deep feature vectors.

[0008] S4. Deep Feature Optimization and Validation: The feature information retention rate is verified by the variance explanation rate, the t-test is used to ensure the significance of features and damage labels, and the mutual information value between features is calculated to eliminate feature redundancy.

[0009] S5. Diagnostic and early warning model construction: Construct an improved CNN-LSTM diagnostic model to achieve damage status classification, build a prediction model that integrates GM(1,1) and CNN-LSTM, calculate risk values ​​according to weights and classify levels;

[0010] S6. Fatigue Damage Diagnosis and Grading Early Warning: Set two levels of real-time early warning for mild damage and severe damage. After triggering, execute the corresponding warning and control. When the model outputs that the damage risk will rise to high risk within the next three days, generate operation and maintenance suggestions including inspection nodes and maintenance measures 24 hours in advance.

[0011] Furthermore, in step S1, the multi-source heterogeneous data includes vibration data, stress-strain data, temperature and humidity data, wind speed data, cumulative usage time of the derrick, and lifting load frequency.

[0012] Furthermore, in step S1, a time-series database is used to store data in the format of timestamp, sensor ID, and data types including vibration X-axis, vibration Y-axis, vibration Z-axis, stress, and temperature.

[0013] Furthermore, step S2 specifically includes the following steps:

[0014] S21. Construct a reference noise channel to collect vibration signals from equipment including pumping units and pump trucks near the derrick, and use the LMS adaptive algorithm to adaptively cancel noise from the raw signals collected by the accelerometer and strain gauge.

[0015] S22. The db8 wavelet basis is selected to perform a 5-level decomposition on the signal after adaptive noise cancellation. After decomposition, a soft thresholding function is used to process the detail coefficients of each level.

[0016] S23. Standardize environmental data including temperature, humidity, and wind speed, as well as historical data including cumulative usage time and load frequency, and use the Z-score standardization method to unify the data dimensions to the range of [-3, 3].

[0017] S24. Abnormal data, including sensor failures and data transmission packet loss, are eliminated based on the 3σ criterion combined with sliding window statistics.

[0018] Furthermore, step S3 specifically includes the following steps:

[0019] S31. Dynamic physical characteristics: The vibration signal is decomposed into 8-12 product functions using the LMD algorithm. For each product function, the energy entropy, sample entropy, and peak factor are calculated. For the stress and strain data, the time domain mean, variance, kurtosis, frequency domain power spectrum peak value, and center frequency are calculated using a sliding window.

[0020] S32. Static physical characteristics: The environmental data is decomposed into 3-5 intrinsic mode functions and 1 trend term using the EMD algorithm. The temperature drift coefficient, humidity fluctuation range, average wind speed, and number of wind speed abrupt changes are extracted from the trend term. Based on the historical data of the derrick operation, the load coefficient, maintenance interval attenuation rate, and cumulative number of failures are calculated.

[0021] S33. Construct the correlation matrix between dynamic and static features, calculate the Pearson correlation coefficient, select feature pairs whose absolute correlation coefficient is greater than the preset value, map the selected feature pairs using the RBF kernel function to generate cross-domain interactive features, calculate the correlation score between each interactive feature and the fatigue damage label, and calculate the weight of each interactive feature using the Softmax function.

[0022] S34. The dynamic physical features, static physical features, and cross-domain interactive features are merged to form the initial aggregated features. The initial aggregated features are compressed using a 3-layer stacked autoencoder. The compressed features are then normalized using MIN-MAX to map all feature values ​​to [0, 1] to obtain the derrick fatigue damage depth features.

[0023] Furthermore, in step S4, the variance explanation rate is used to evaluate the degree of retention of compressed feature information, requiring the variance explanation rate to be no less than its threshold. The significance of depth features and fatigue damage labels is analyzed by t-test, requiring the p-value of the test result to be less than its preset value. The mutual information value between any two depth features is calculated. If the mutual information value between any two depth features exceeds its preset value, principal component analysis is used for further dimensionality reduction. By constructing the feature covariance matrix, solving for eigenvalues ​​and eigenvectors, principal components with a cumulative variance contribution rate no less than its threshold are selected as new features. Otherwise, dimensionality reduction is not required.

[0024] Furthermore, in step S5, the specific steps for constructing the diagnostic early warning model are as follows:

[0025] S51. A CNN-LSTM hybrid model is adopted, with deep features as input. The convolutional layer has 16 3×1 kernels with a stride of 1 and padding of 1. The activation function is ReLU. The pooling layer uses 2×1 max pooling with a stride of 2. The LSTM layer has 128 neurons with a dropout ratio of 0.2. The fully connected layer is divided into two layers: the first layer has 32 neurons with ReLU activation function, and the second layer has 3 neurons corresponding to no damage, mild damage, and severe damage, respectively, with softmax activation function to output the probability of each state. Detection data of derricks with different service years are collected and labeled as 100,000 samples containing deep features and corresponding damage state labels. The samples are divided into training and test sets in a 7:3 ratio. The optimizer is Adam with a learning rate of 0.001 and cross-entropy loss function. The batch size is 32, and the number of iterations is 100. An early stopping mechanism is also introduced, and training is stopped when the accuracy of the validation set does not improve for 5 consecutive times.

[0026] S52. Construct a GM(1,1) grey prediction model. The input data is a 30-day hourly time series of deep features. The time series of each feature is accumulated sequentially to generate the model. Then, a first-order linear differential equation is established. The least squares method is used to solve the development coefficient and grey action in the equation. The posterior difference test is used to test the prediction model. The fusion prediction rule is: the damage risk value of a future day is equal to the first preset weight coefficient multiplied by the damage probability corresponding to the feature predicted by the GM(1,1) model, plus the second preset weight coefficient multiplied by the probability of damage state on that day predicted by the CNN-LSTM model. The risk levels are divided into: low risk, medium risk, and high risk.

[0027] Furthermore, in step S6, the yellow warning trigger condition is: the probability of the diagnostic model outputting a mild damage state is not lower than a preset probability threshold, triggering the execution of: activating the local audible and visual alarm, popping up a warning pop-up window on the monitoring platform, and recording information such as the warning occurrence time and the corresponding derrick node location; the red warning trigger condition is: the probability of the diagnostic model outputting a severe damage state is not lower than a preset probability threshold, triggering the execution of: in addition to executing all actions for the yellow warning, additionally: pushing the warning information to the maintenance manager via SMS and a dedicated APP. If the manager does not confirm receipt within the predetermined period, the system automatically calls the manager to remind them and simultaneously sends an instruction to the derrick control system to suspend derrick operations; when the prediction model outputs that the damage risk will rise to a high-risk level within the next 3 days, the system generates maintenance suggestions 24 hours in advance, specifically combining the correlation between depth characteristics and damage types to clearly identify the recommended derrick nodes to be inspected, specific maintenance measures, and temporary load limits. After the maintenance suggestions are generated, they are pushed to the maintenance management system, automatically generating a maintenance work order. The work order includes the suggested execution time, responsible personnel, and required tools.

[0028] Beneficial effects: Breaking through the limitations of existing technologies that only focus on dynamic physical data, this technology simultaneously collects dynamic data such as vibration and stress-strain, environmental data such as temperature, humidity, and wind speed, as well as historical data such as the cumulative usage time of the derrick and the frequency of lifting loads. The combination of multi-dimensional data comprehensively reflects the true fatigue state of the derrick, eliminating "data blind spots" and providing complete data support for subsequent diagnosis and early warning. Dynamic signals undergo a two-stage processing of "adaptive noise cancellation + 5-layer soft threshold denoising using a db8 ​​wavelet base," effectively filtering equipment background and random environmental noise, significantly improving the signal-to-noise ratio. Static data is standardized using Z-score to unify dimensions, and combined with the 3σ criterion and sliding window to eliminate abnormal data such as sensor malfunctions and data transmission packet loss, ensuring accurate input data and reducing errors in subsequent feature extraction and model diagnosis. Deep features are extracted through a three-step process: "single-source feature decomposition → cross-domain feature association → feature hierarchical aggregation". First, the single-source features of dynamic and static data are fully explored. Then, key feature pairs are selected through Pearson correlation coefficient. Cross-domain interactive features are constructed through RBF kernel function mapping and Softmax weight allocation. Finally, a 3-layer stacked autoencoder is used to compress and standardize the feature vector. The resulting 8-dimensional deep feature vector has low redundancy and strong damage correlation, laying the foundation for accurate model diagnosis.

[0029] The improved CNN-LSTM diagnostic model can quickly and accurately classify damage into "no damage / minor damage / severe damage," achieving high accuracy on the test set and short response time, meeting real-time monitoring requirements. The prediction model, which integrates GM(1,1) and CNN-LSTM, calculates risk values ​​and classifies levels according to weights, predicting damage trends up to 7 days in advance. Compared to traditional models, it has a longer prediction step and higher accuracy, allowing sufficient time for maintenance. Real-time early warnings for minor and severe damage are set up, triggering corresponding alarms and operational control actions to avoid "over-response" or "under-response." When a high-risk scenario is predicted within the next 3 days, maintenance suggestions containing specific inspection nodes and maintenance measures are generated 24 hours in advance, guiding precise maintenance, reducing equipment downtime, lowering the probability of derrick safety accidents caused by fatigue damage, and ensuring industrial production safety and personnel safety. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0031] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Example 1

[0033] Fatigue damage diagnosis and early warning for ZJ70 derricks in oil extraction scenarios:

[0034] The first step is distributed data collection from multiple sources:

[0035] Data types and acquisition equipment: Multi-source heterogeneous data were collected, including: vibration data (X / Y / Z axes) of six key nodes of the derrick top crane, middle crossbeam, and bottom base were collected using an IEPE triaxial accelerometer (range ±50g, sensitivity 100mV / g); stress and strain data of the above nodes were collected using foil strain gauges (sensitivity coefficient 2.1±1%, range -2000~+2000με); ambient temperature and humidity data of the middle layer of the derrick were collected using a temperature and humidity sensor (measurement range -40~85℃, 0~100%RH); top wind speed data were collected using an ultrasonic anemometer (measurement range 0~60m / s); and the cumulative usage time of the derrick (72,600 hours as of the time of acquisition) and lifting load frequency (32 lifting operations per day on average, including 2 overload loads (>110% of rated load) per day) were synchronized from the oilfield SCADA system.

[0036] Data Acquisition Frequency and Data Storage: Set the acquisition frequency as follows: vibration data 1000Hz, stress and strain data 500Hz, temperature and humidity data 1Hz, and wind speed data 10Hz; use InfluxDB time-series database for storage, and define the data format as in claim 3: timestamp (millisecond level, such as "2024-10-01 08:00:00.123"), sensor ID (such as "ACC-01" representing the top crane accelerometer), data type (such as "vibration X-axis", "stress", "temperature") and corresponding value, with the number of bytes per data entry controlled within 32 bytes.

[0037] The second step is hierarchical preprocessing of multi-source heterogeneous data:

[0038] 1. Adaptive noise cancellation: According to claim 4, a reference noise channel is constructed, and vibration signals of the pumping unit (operating frequency 50Hz) and mud pump (operating frequency 120Hz) near the derrick are collected. The original vibration and stress signals collected by the accelerometer and strain gauge are noise-cancelled through the LMS adaptive algorithm (step size 0.001, filter order 32) to eliminate 75% of the equipment background noise.

[0039] 2: Wavelet thresholding denoising: The db8 wavelet basis is selected, and the signal after S21 processing is decomposed into 5 levels. The soft thresholding function is used to process the detail coefficients of each level. After reconstruction, the signal-to-noise ratio of the vibration signal is increased from 28dB to 36dB, and the distortion of the stress signal is controlled within 4%.

[0040] 3. Static Data Standardization: Environmental data such as temperature, humidity, and wind speed, as well as historical data such as cumulative usage time and load frequency, are standardized using the Z-score method (formula: μ is the mean of the data over the past 30 days, and σ is the standard deviation. The data dimensions are unified to the range of [-3, 3]. For example, temperature data from "-5 to 38℃" is standardized to "-1.2 to 2.8".

[0041] 4. Abnormal data removal: Based on the 3σ criterion and combined with 10-minute sliding window statistics, abnormal data was removed, such as "stress value suddenly increased to 5000με" caused by strain gauge failure (exceeding the 3σ range) and "wind speed value was 0 (lasting for 5 minutes, which does not match the actual weather)" caused by data transmission packet loss. The final abnormal data removal rate was 1.1%.

[0042] The third step is deep feature extraction from multi-source data:

[0043] 1. Dynamic Physical Feature Extraction: The noise-reduced vibration signal is decomposed into 10 product functions (PF) using the LMD algorithm. The energy entropy (e.g., PF1 energy entropy 0.82, PF2 energy entropy 0.65), sample entropy (embedding dimension 2, similarity tolerance 0.2 × signal standard deviation), and peak factor (e.g., PF3 peak factor 4.2) are calculated for each PF, resulting in 30 dimensions of vibration features. A 500ms sliding window (50% overlap) is used for the stress and strain data to calculate the time-domain mean (e.g., average stress of the middle beam 120MPa) and variance (8.5MPa). 2 The stress features include 5 dimensions: kurtosis (3.1), peak power spectrum (corresponding to a power spectrum value of 0.08 W / Hz at 15 Hz), center frequency (18 Hz), and dynamic physical features. The total number of dynamic physical features is 35.

[0044] 2. Static Physical Feature Extraction: The standardized environmental data was decomposed into four intrinsic mode functions (IMFs) and one trend term using the EMD algorithm. From the trend term, the following four environmental features were extracted: temperature drift coefficient (0.05℃ / h, reflecting the rising trend of daytime temperature), humidity fluctuation amplitude (15%RH, reflecting the diurnal humidity change), average wind speed (6.2m / s), and number of wind speed abrupt changes (3 times the wind speed exceeded 15m / s in 24 hours), for a total of four dimensions of environmental features. Based on the operating data of the past year, the following three historical features were calculated: monthly load growth coefficient (1.08, reflecting a slight increase in load month by month), maintenance interval decay rate (0.12, reflecting the maintenance interval shortening from 90 days to 78 days), and cumulative number of failures (4 minor failures in the past 6 months), for a total of seven dimensions of static physical features.

[0045] 3. Cross-domain feature association: Construct an association matrix of 35-dimensional dynamic features and 7-dimensional static features, calculate the Pearson correlation coefficient, and select feature pairs with an absolute value > 0.6 (e.g., "Vibration PF1 energy entropy - temperature drift coefficient" correlation coefficient 0.72, "Stress mean - monthly load growth coefficient" correlation coefficient 0.68), for a total of 11 key feature pairs; For each feature pair, use the RBF kernel function (K(x,y)=exp(-0.1||xy||^2)) to map and generate 11-dimensional cross-domain interactive features, calculate the correlation score between each interactive feature and the fatigue damage label ("mild damage") (e.g., "Vibration PF1 energy entropy - temperature drift coefficient" score 0.85), and calculate the weight using the Softmax function (this feature pair has a weight of 0.19, which is the highest).

[0046] 4. Feature aggregation and standardization: The 35-dimensional dynamic features, 7-dimensional static features, and 11-dimensional weighted cross-domain features are merged to form an initial 53-dimensional aggregated feature. A 3-layer stacked autoencoder (53-dimensional input layer → 32-dimensional hidden layer → 16-dimensional hidden layer → 8-dimensional output layer) is used for compression. After unsupervised pre-training (learning rate 0.001, 500 iterations), the reconstruction error is 0.008, and 92% of the feature information is retained. The compressed 8-dimensional features are standardized using MIN-MAX and mapped to the [0, 1] interval to obtain the deep feature vector: [0.32, 0.58, 0.15, 0.71, 0.29, 0.63, 0.45, 0.82].

[0047] Step 4: Deep Feature Optimization and Validation

[0048] Information retention verification: The variance explained by the 8-dimensional features is 92% (sum of variances of compressed features / sum of variances of initial aggregated features), which meets the requirement of "≥90%", indicating that the feature information is sufficiently retained.

[0049] Significance verification: A t-test was performed on the 8-dimensional features and the "mild injury" label. All feature p-values ​​were <0.05 (e.g., the p-value for the second feature was 0.023), indicating that the features and the injury label were significantly correlated.

[0050] Redundancy elimination: Calculate the mutual information value between features. The maximum value is 0.65 (4th and 6th dimension features). Since it is less than 0.8, no additional dimensionality reduction is required and the feature independence requirement is met.

[0051] Step 5: Construction of the diagnostic and early warning model:

[0052] An improved CNN-LSTM diagnostic model was developed. The model structure was designed according to claim 1. The input layer received 8-dimensional deep features (input shape (32, 8, 1), batch size = 32), convolutional layers (16 3×1 convolutional kernels, ReLU activation), pooling layers (2×1 max pooling), LSTM layers (128 neurons, dropout = 0.2), and fully connected layers (32 ReLU neurons → 3 softmax output neurons). The model was trained using 100,000 samples from 10 derricks in the oilfield (divided into training and test sets in a 7:3 ratio), with an Adam optimizer (learning rate 0.001), cross-entropy loss, and an early stopping mechanism (patience = 5). After training, the accuracy on the test set was 96.5%. The probability of outputting "mild damage" for the derrick in this embodiment was 0.83, "no damage" was 0.15, and "severe damage" was 0.02.

[0053] GM(1,1)-CNN-LSTM fusion prediction model: Using 30-day hourly time series data with 8-dimensional features as input, a GM(1,1) model is constructed, and a differential equation is established after 1-AGO processing. The solutions yield a = -0.012, b = 0.45, and the posterior difference test results are C = 0.32 < 0.35, P = 0.96 > 0.95, indicating the model is qualified. The risk values ​​for the next 7 days are calculated by weighting (GM(1,1) accounts for 0.6, CNN-LSTM accounts for 0.4), and the results are: Day 1 0.35 (medium risk), Day 3 0.42 (medium risk), Day 5 0.58 (medium risk), and Day 7 0.65 (medium risk).

[0054] Step 6: Fatigue Injury Diagnosis and Grading Early Warning:

[0055] Real-time warning: Because the diagnostic model outputs a probability of "minor damage" of 0.83≥0.7, a yellow warning is triggered: the derrick site audible and visual alarm (90dB) is activated, and a pop-up window appears on the monitoring platform (displaying "minor damage to the middle layer crossbeam, warning time 2024-10-01 10:30:00"), and the warning node is recorded as "middle layer crossbeam".

[0056] Trend warning and maintenance suggestions: The forecast for the next 7 days is medium risk, with no "upgrade to high risk in the next 3 days" scenario. No emergency maintenance suggestions are needed, but the system still generates a routine maintenance prompt: "It is recommended to check the welds of the middle layer crossbeam within 7 days, use ultrasonic testing to confirm the crack status, and reduce the lifting load by 10% in the near future to avoid overloading operations."

[0057] Example 2

[0058] A QTZ80 tower derrick that has been in service for 5 years in a construction project

[0059] The first step is distributed data collection from multiple sources:

[0060] Data types and acquisition equipment: Vibration data (X-axis, Y-axis, and Z-axis) of three nodes at the top, middle, and bottom of the derrick tower are collected using a triaxial accelerometer (range ±20g, sensitivity 50mV / g); stress and strain data of the above nodes are collected using strain gauges (sensitivity coefficient 2.0±1%, range -1500~+1500με); environmental data of the middle of the tower is collected using a temperature and humidity sensor (measurement range -20~60℃, 0~100%RH); wind speed data at the top is collected using a wind speed sensor (measurement range 0~30m / s); and the cumulative usage time of the derrick (43,800 hours) and lifting load frequency (average 45 lifting times per day, and average 5 overload loads (>110% of rated load) per day) are synchronized from the construction management system.

[0061] Data acquisition frequency and data storage: vibration data 800Hz, stress and strain data 400Hz, temperature and humidity data 1Hz, wind speed data 8Hz; InfluxDB time series database is used for storage, and the data format is defined as in claim 3: timestamp (e.g., "2024-10-05 09:15:00.456"), sensor ID (e.g., "STR-02" representing the strain gauge in the middle of the tower), data type (e.g., "vibration Y-axis", "humidity", "load frequency") and value, with the number of bytes per data entry controlled within 30 bytes.

[0062] The second step is hierarchical preprocessing of multi-source heterogeneous data:

[0063] 1. Adaptive noise cancellation: According to claim 4, the vibration signals of the tower crane (operating frequency 60Hz) and concrete pump truck (operating frequency 90Hz) at the construction site are collected by the reference noise channel. The original vibration and stress signals are processed by the LMS adaptive algorithm (step size 0.001, filter order 28) to eliminate 70% of the equipment background noise.

[0064] 2: Wavelet thresholding denoising: The signal is decomposed into 5 levels using the db8 wavelet basis. After soft thresholding, the signal-to-noise ratio of the vibration signal is increased from 26dB to 34dB, and the distortion of the stress signal is within 5%.

[0065] 3. Static data standardization: Data such as temperature and humidity, wind speed, cumulative usage time, and load frequency are standardized using Z-score, and the units are unified to [-3, 3]. For example, wind speed data is standardized from "0~22m / s" to "0~2.5".

[0066] 4. Abnormal data removal: Based on the 3σ criterion and a 10-minute sliding window, abnormal data such as "stress value of 0 (lasting for 8 minutes)" caused by strain gauge breakage and "humidity value of 100% (inconsistent with the actual weather humidity of 60%)" caused by temperature and humidity sensor failure were removed, with a removal rate of 1.3%.

[0067] The third step is deep feature extraction from multi-source data:

[0068] 1. Dynamic Physical Feature Extraction: The vibration signal is decomposed into 8 power factors (PFs) using LMD. Energy entropy (e.g., PF1 energy entropy 0.78, PF3 energy entropy 0.59), sample entropy, and peak factor are calculated, resulting in 24-dimensional vibration features. Stress data are calculated using a 400ms sliding window (50% overlap) for time-domain mean (mean stress in the middle of the tower is 95 MPa) and variance (6.8 MPa). 2 The stress features include 5 dimensions: kurtosis (2.9), peak power spectrum (corresponding to a power spectrum value of 0.06 W / Hz at 12 Hz), and center frequency (14 Hz). The dynamic features total 29 dimensions.

[0069] 2: Static physical feature extraction: Environmental data was decomposed into 3 IMFs and 1 trend term by EMD, and the following 4-dimensional environmental features were extracted: temperature drift coefficient (0.04℃ / h), humidity fluctuation range (12%RH), average wind speed (5.8m / s), and number of wind speed abrupt changes (2 times exceeding 12m / s in 24 hours); the monthly load growth coefficient (1.12), maintenance interval attenuation rate (0.15), and cumulative number of failures (6 minor failures in the past 6 months) were calculated from historical data, for a total of 3-dimensional historical features, making a total of 7-dimensional static features.

[0070] 3: Cross-domain feature association: Select dynamic-static feature pairs (such as "Vibration PF2 energy entropy-temperature drift coefficient" correlation coefficient 0.65, "Stress mean-monthly load growth coefficient" correlation coefficient 0.71), a total of 10 groups; RBF kernel function mapping generates 10-dimensional cross-domain features, after Softmax weight calculation, "Stress mean-monthly load growth coefficient" has the highest weight (0.21).

[0071] 4: Feature aggregation and standardization: 29+7+10=46-dimensional initial aggregated features, compressed into 8 dimensions by 3-layer SAE (reconstruction error 0.009, variance explained 91%), MIN-MAX standardized deep feature vector: [0.41, 0.62, 0.23, 0.78, 0.35, 0.69, 0.52, 0.88].

[0072] Step 4: Deep Feature Optimization and Validation

[0073] The variance explained is 91%, which meets the requirements; the t-test p-values ​​for the 8-dimensional features and the "mild damage" label are all <0.05 (e.g., the p-value for the 3rd dimension feature is 0.018); the maximum mutual information value between features is 0.68 <0.8, indicating no redundancy.

[0074] Step 5: Construction of the diagnostic and early warning model:

[0075] Diagnostic model: Input 8-dimensional features, after training the test set accuracy is 95.8%, and for the derrick in this embodiment, the output probability of "minor damage" is 0.87, "no damage" is 0.11, and "severe damage" is 0.02.

[0076] Prediction model: GM(1,1) model posterior difference test C = 0.33 < 0.35, P = 0.97 > 0.95; fusion prediction of risk values ​​for the next 7 days: Day 1 0.40 (medium risk), Day 3 0.55 (medium risk), Day 5 0.68 (medium risk), Day 7 0.72 (high risk).

[0077] Step 6: Fatigue Injury Diagnosis and Grading Early Warning:

[0078] Real-time warning: When a yellow warning is triggered, an on-site audible and visual alarm will be activated, and a pop-up window on the platform will record "Minor damage to the middle of the tower, warning time 2024-10-05 11:20:00".

[0079] Trend Warning and Maintenance Recommendations: Due to the predicted risk value of 0.72 (high risk) for day 7, and the fact that the risk value is close to high risk (0.68 for day 5) within 3 days of the current day, the system generates maintenance recommendations 24 hours in advance: "It is recommended to stop derrick operations within 48 hours, check the stress concentration area in the middle of the tower body (focusing on connecting bolts and welds), use magnetic particle testing to detect cracks, replace worn bolts, reduce the lifting load by 15% in the next month, and strictly prohibit overloading."

[0080] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A neural network-based derrick fatigue damage multi-dimensional diagnosis and early warning method, characterized in that, The method specifically comprises the following steps: S1, distributed collection of multi-source data: collect multi-source heterogeneous data according to a set collection frequency, and store and define a unified data format by using a time series database; S2, hierarchical preprocessing of multi-source heterogeneous data: dynamic signals are processed by two-stage processing of adaptive noise cancellation and wavelet threshold denoising, static data are standardized and unified in dimension by Z-score safety monitoring technology field, and combined with 3σ safety monitoring technology field guidelines and sliding window to eliminate abnormal data; S3, deep feature extraction of multi-source data: deep feature vectors are extracted by sequentially performing single-source feature decomposition, cross-domain feature association and feature hierarchical aggregation on the preprocessed multi-source data; S4, deep feature optimization and verification: the feature information retention is verified by variance explanation rate, the significance of the feature and the damage label is ensured by t-test, and the feature redundancy is eliminated by calculating the mutual information value between the features; S5, construction of diagnosis and early warning model: an improved CNN-LSTM diagnosis model is constructed to realize damage state classification, a prediction model of GM(1,1) and CNN-LSTM fusion is built, a risk value is calculated according to the weight, and a grade is divided; S6, fatigue damage diagnosis and grading early warning: two levels of real-time early warning of mild damage and severe damage are set, corresponding early warning and control are executed after triggering, and operation and maintenance suggestions containing inspection nodes and maintenance measures are generated 24 hours in advance when the model output damage risk within the next three days will rise to high risk.

2. The neural network-based drilling rig fatigue damage multi-dimensional diagnosis and early warning method according to claim 1, characterized in that: In the step S1, the multi-source heterogeneous data includes vibration data, stress and strain data, temperature and humidity data, wind speed data, cumulative use time of the derrick, and hoisting load frequency.

3. The neural network-based drilling rig fatigue damage multi-dimensional diagnosis and early warning method according to claim 1, characterized in that: In the step S1, the time series database is used to store the data format of time stamp, sensor ID, and data types including vibration X-axis, vibration Y-axis, vibration Z-axis, stress, and temperature.

4. The neural network-based drilling rig fatigue damage multi-dimensional diagnosis and early warning method according to claim 1, characterized in that: The step S2 specifically comprises the following steps: S21, a reference noise channel is constructed to collect vibration signals of equipment including a nearby pumping unit and a pump truck, and LMS adaptive algorithm is used to perform adaptive noise cancellation on the original signals collected by the acceleration sensor and the strain gauge; S22, db8 wavelet basis is selected to decompose the signals processed by adaptive noise cancellation for 5 layers, and soft threshold function is used to process the detail coefficients of each layer after decomposition; S23, the environmental data including temperature, humidity, and wind speed and the historical data including cumulative use time and load frequency are standardized, and Z-score standardization method is used to unify the data dimension to the interval [-3, 3]; S24, based on the 3σ criterion, combined with the sliding window statistics, abnormal data including sensor failure and data transmission packet loss are eliminated.

5. The neural network-based multi-dimensional diagnosis and early warning method for fatigue damage of a derrick according to claim 1, characterized in that: The step S3 specifically comprises the following steps: S31, dynamic physical characteristics: LMD algorithm is used to decompose the vibration signals into 8-12 product functions, energy entropy, sample entropy, and peak factor are calculated for each product function, time domain mean, variance, kurtosis, frequency domain power spectrum peak value, and center frequency are calculated for the stress and strain data by using a sliding window; S32. Static physical characteristics: The environmental data is decomposed into 3-5 intrinsic mode functions and 1 trend term using the EMD algorithm. The temperature drift coefficient, humidity fluctuation range, average wind speed, and number of wind speed abrupt changes are extracted from the trend term. Based on the historical data of the derrick operation, the load coefficient, maintenance interval attenuation rate, and cumulative number of failures are calculated. S33. Construct the correlation matrix between dynamic and static features, calculate the Pearson correlation coefficient, select feature pairs whose absolute correlation coefficient is greater than the preset value, map the selected feature pairs using the RBF kernel function to generate cross-domain interactive features, calculate the correlation score between each interactive feature and the fatigue damage label, and calculate the weight of each interactive feature using the Softmax function. S34. The dynamic physical features, static physical features, and cross-domain interactive features are merged to form the initial aggregated features. The initial aggregated features are compressed using a 3-layer stacked autoencoder. The compressed features are then normalized using MIN-MAX to map all feature values ​​to [0, 1] to obtain the derrick fatigue damage depth features.

6. The neural network-based drilling rig fatigue damage multi-dimensional diagnosis and early warning method according to claim 5, characterized in that: In step S4, the variance explained rate is used to evaluate the degree of retention of compressed feature information, requiring the variance explained rate to be no less than its threshold. The significance of depth features and fatigue damage labels is analyzed by t-test, requiring the p-value of the test result to be less than its preset value. The mutual information value between any two depth features is calculated. If the mutual information value between any two depth features exceeds its preset value, principal component analysis is used for further dimensionality reduction. By constructing the feature covariance matrix and solving for the eigenvalues ​​and eigenvectors, the principal components with a cumulative variance contribution rate no less than its threshold are selected as new features. Otherwise, dimensionality reduction is not required.

7. The neural network-based multi-dimensional diagnosis and early warning method for fatigue damage of a derrick according to claim 1, characterized in that: In step S5, the specific steps for constructing the diagnostic early warning model are as follows: S51. A CNN-LSTM hybrid model is adopted, with deep features as input. The convolutional layer has 16 3×1 kernels with a stride of 1 and padding of 1. The activation function is ReLU. The pooling layer uses 2×1 max pooling with a stride of 2. The LSTM layer has 128 neurons with a dropout ratio of 0.

2. The fully connected layer is divided into two layers: the first layer has 32 neurons with ReLU activation function, and the second layer has 3 neurons corresponding to no damage, mild damage, and severe damage, respectively, with softmax activation function to output the probability of each state. Detection data of derricks with different service years are collected and labeled as 100,000 samples containing deep features and corresponding damage state labels. The samples are divided into training and test sets in a 7:3 ratio. The optimizer is Adam with a learning rate of 0.001 and cross-entropy loss function. The batch size is 32, and the number of iterations is 100. An early stopping mechanism is also introduced, and training is stopped when the accuracy of the validation set does not improve for 5 consecutive times. S52. Construct a GM(1,1) grey prediction model. The input data is a 30-day hourly time series of deep features. The time series of each feature is accumulated sequentially to generate the model. Then, a first-order linear differential equation is established. The least squares method is used to solve the development coefficient and grey action in the equation. The posterior difference test is used to test the prediction model. The fusion prediction rule is: the damage risk value of a future day is equal to the first preset weight coefficient multiplied by the damage probability corresponding to the feature predicted by the GM(1,1) model, plus the second preset weight coefficient multiplied by the probability of damage state on that day predicted by the CNN-LSTM model. The risk levels are divided into: low risk, medium risk, and high risk.

8. The neural network-based drilling rig fatigue damage multi-dimensional diagnosis and early warning method according to claim 1, characterized in that: In step S6, the yellow warning trigger condition is: the probability of the diagnostic model outputting a mild damage state is not lower than the preset probability threshold, triggering the execution of: activating the local audible and visual alarm, popping up a warning pop-up window on the monitoring platform, and recording the warning occurrence time, the corresponding derrick node location, and other information; the red warning trigger condition is: the probability of the diagnostic model outputting a severe damage state is not lower than the preset probability threshold, triggering the execution of: in addition to executing all actions for the yellow warning, additionally: pushing the warning information to the operation and maintenance manager via SMS and a dedicated APP. If the manager does not confirm receipt within the predetermined period, the system will automatically call the manager to remind them and simultaneously send an instruction to the derrick control system to suspend derrick operations; when the prediction model outputs that the damage risk will rise to a high-risk level within the next 3 days, the system will generate maintenance suggestions 24 hours in advance. Specifically, it will combine the correlation between depth characteristics and damage type to clearly identify the recommended derrick nodes to be inspected, specific maintenance measures, and temporary load limits. After the maintenance suggestions are generated, they will be pushed to the operation and maintenance management system, automatically generating a maintenance work order. The work order includes the suggested execution time, responsible personnel, and required tools.

Images