Distributed surface well data processing method

Abstract

The invention discloses a distributed ground well data processing method, and relates to the technical field of data processing, and the method comprises the steps: collecting the vibration data of an air conditioning unit of an airport distributed ground well through an acceleration sensor; de-noising processing is carried out on the collected vibration data of the air conditioning unit by combining a time segmentation clustering algorithm on the basis of the wavelet decomposition layer number dynamically adjusted based on the signal-to-noise ratio and a threshold value strategy; performing segmented feature coding on the vibration data of the air conditioning unit; the deep neural network model based on feedback rewards carries out training optimization through a feedback reward mechanism, and the performance of the model in a complex fault diagnosis task is improved; and performing fault diagnosis classification on the vibration data of the air conditioning unit by using the trained deep neural network model based on the feedback rewards. According to the method, a signal-to-noise ratio driven wavelet decomposition layer number and threshold value control mechanism is adopted, the decomposition depth and the threshold value function are dynamically adjusted according to the signal local standard deviation, and the capability of retaining fault features under different working conditions is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a distributed ground well data processing method. Background Art

[0002] Ensuring the safe operation of airport ground equipment is the primary task of the airport equipment maintenance department. Timely monitoring or even prediction of equipment failures and taking effective containment measures, and analyzing the causes of equipment failures are important issues that equipment managers need to solve.

[0003] The air conditioning unit in the airport ground well is a key equipment for maintaining the temperature and humidity environment of the underground facilities in the terminal building and apron. Its stable operation directly affects the efficiency of the airport energy system and the service life of the equipment. Vibration data, as the core index of the unit's health status, can provide early warnings for faults: abnormal vibrations may indicate mechanical faults such as bearing wear, impeller imbalance, and motor misalignment. It is necessary to intervene in advance to avoid shutdown. At the same time, vibration characteristics can reflect the load status of the unit and assist in dynamically adjusting operating parameters (such as fan speed) to reduce energy consumption. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a distributed ground well data processing method.

[0005] The technical solution adopted to solve the above technical problem is: A distributed ground well data processing method includes the following steps: S1, collecting vibration data of the air conditioning unit in the airport distributed ground well through an acceleration sensor; S2, based on the wavelet decomposition layer number and threshold strategy with dynamic adjustment of the signal-to-noise ratio, combined with the time-segment clustering algorithm, denoising the collected vibration data of the air conditioning unit; S3, adopting a sliding window dynamic division strategy, combined with the autocorrelation coefficient within the window to determine the segmentation boundary, and performing segmented feature encoding on the vibration data of the air conditioning unit; S4, training and optimizing the deep neural network model based on feedback reward through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks; S5, using the trained deep neural network model based on feedback reward to perform fault diagnosis and classification on the vibration data of the air conditioning unit.

[0006] Further, the wavelet decomposition layer number and threshold strategy with dynamic adjustment of the signal-to-noise ratio in S2 are: performing wavelet decomposition on the original vibration signal, dynamically adjusting the decomposition layer number and threshold according to the signal-to-noise ratio of the signal, calculating the signal-to-noise ratio estimation value for each layer of wavelet coefficients, performing threshold processing on the wavelet coefficients through a dynamic threshold function, and reconstructing the processed wavelet coefficients to obtain the denoised vibration signal, expressed as: , In the above formula, is the vibration signal after wavelet threshold denoising, is a positive integer, is the number of dynamic decomposition layers, is the layer wavelet coefficient, is the original vibration signal, is the dynamic threshold function, is the layer wavelet coefficient signal-to-noise ratio estimate; is the rounding operation, is the logarithmic function, is the standard deviation function, is to calculate the standard deviation of the original vibration signal; is norm; is the exponential function with the natural constant as the base, is the slope parameter, is the signal-to-noise ratio threshold, = 0.1.

[0007] Furthermore, the sliding window dynamic partitioning strategy in S3 is as follows: For the denoised vibration signal, a sliding window is used for scanning, and the local autocorrelation coefficient at each window position is calculated. When the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, this position is used as the segmentation boundary, and the signal is divided into multiple segments. The signals within each segment are encoded using a preset bidirectional LSTM, and the outputs of the forward and backward LSTMs are concatenated to obtain the feature vector of each segment, expressed as: , In the above formula, is the set of vibration signals after segmentation, is the total number of dynamic segments, is a positive integer, is the starting time point of the is the ending time point of the is from to denoised vibration signal segment, is the position local autocorrelation coefficient, is the boundary determination threshold, is the global average autocorrelation coefficient; is the local autocorrelation coefficient calculation window radius, is a positive integer, is the position the The denoised signal value of a sampling point is the position the th denoised signal value of the sampling point on the right; is the total duration of the vibration signal, is a positive integer, is the local autocorrelation coefficient at the th time point.

[0008] Furthermore, the training process of the deep neural network model based on feedback reward in S4 includes the following steps: S401. Initialize the weights of the deep neural network: Based on the weight initialization strategy of pre-classification confidence, calculate the product of the gradient expectation and the reward during the pre-training stage, so that the initial weights bias towards the high-confidence feature direction; S402. Perform adaptive gating feature selection: Use a dynamic feature selection gate to achieve adaptive gating feature selection, dynamically adjust the feature retention ratio through cumulative rewards, and suppress low-contribution features; S403. Perform forward propagation of data: Adopt a spatio-temporal attention-guided residual propagation mechanism, and dynamically construct skip connection paths through the results of gating feature selection; S404. Perform class prediction: Adopt a multi-granularity prototype contrast classification strategy, construct learnable class prototype vectors in the feature space, and determine the final class through the multi-scale matching scores of the gating features and the prototype vectors; S405. Calculate the loss function: Adopt a hierarchical reward-weighted cross-entropy loss to dynamically adjust the loss weights of each class, and strengthen the attention to difficult-to-classify faults; S406. Perform dynamic sparse connection optimization of the deep neural network: Based on the connection pruning strategy of reward gradient, correct the gradient direction through cumulative rewards, and enhance the ability to capture continuous fault patterns; S407. Calculate the parameter update amount: Adopt the deep neural network parameter update rule based on the reward accumulation amount; S408. Update the parameters of the deep neural network: Adopt a momentum-accelerated reward-aware update strategy; S409. Repeat the above steps until the preset stop iteration condition is met, which means the model training is completed.

[0009] Furthermore, the weight initialization strategy of pre-classification confidence in S401 is as follows: In the pre-training stage, for each layer of the deep neural network, calculate the gradient of its pre-classification cross-entropy loss with respect to the weights, multiply the gradient by the pre-classification confidence reward factor to obtain the gradient expectation, multiply the gradient expectation by the initial learning rate, and then add it to the initial weight of the He normal distribution of this layer to obtain the initial weight of this layer.

[0010] Furthermore, the method for adaptive gating feature selection in S402 is as follows: for the feature vector of each segment, calculate the activation value of its dynamic feature selection gate, map the activation value to the interval [0, 1] through the Sigmoid activation function to obtain the output of the feature selection gate, use this output to weight the feature vector element by element, retain the important features, and at the same time, weight and fuse the unselected features with the selected features of the previous moment to obtain the final feature selection result.

[0011] Furthermore, the spatio-temporal attention-guided residual propagation mechanism in S403 is as follows: calculate the spatio-temporal attention weight for the output features of the previous layer through the weight matrix, multiply it element by element with the conventional linear transformation, perform non-linear projection transformation on the output features of the previous layer through multiple residual paths and sum them up, linearly interpolate the attention-weighted result and the residual path output according to the attention weight, and dynamically construct the skip connection path.

[0012] Furthermore, the multi-granularity prototype contrast classification strategy in S404 is as follows: for each segment feature, calculate the exponentially normalized score of its negative L2 distance from all category prototype vectors, weight the scores of each segment through the gating coefficient, calculate the classification score for the global average feature using the fully connected layer, and finally fuse the prototype matching score and the traditional classification score proportionally to select the category with the maximum score, which is expressed as: , In the formula, is the multi-scale prototype matching score of the th class, is the total number of dynamic segments, is a positive integer, is the gating coefficient of the th segment, is the scale factor, is the th vibration signal segment's final retained feature vector obtained after passing through the adaptive gating feature selection mechanism, is the th class's trainable prototype vector, is norm, is the total number of fault classes, is the th class's trainable prototype vector, is the currently calculated class index, is the temporary index variable when traversing all classes, is the exponential function with the natural constant as the base; is the predicted probability distribution output in the pre-classification stage, Denote the category index that takes the maximum value within the parentheses, is the fusion coefficient, is the Sigmoid activation function, is the fully connected classification weight, is the feature vector after global average pooling.

[0013] Furthermore, the calculation method of the loss function in S405 is as follows: Calculate the accuracy of each category on the validation set, as well as the F1 scores on the training set and the validation set. Assign a weight factor to each category, multiply the weight factor of each category by the corresponding cross-entropy loss to obtain the weighted loss value, and sum the weighted loss values of all categories to obtain the final total loss function value.

[0014] Furthermore, the connection pruning strategy of the reward gradient in S406 is as follows: Calculate the gradient value of each connection, multiply it by the corresponding weight value to obtain an index measuring the importance of the connection, correct the gradient direction according to the cumulative reward to enhance the ability to capture continuous failure modes, calculate the threshold of each connection, generate a mask matrix according to the calculated threshold and the importance index of the connection, and set the unimportant connections to zero to achieve dynamic sparse connection optimization, expressed as: , In the formula, is the layer mask matrix, is a positive integer, is the indicator function, is the partial derivative symbol, is the weighted cross-entropy loss value, is the layer, the th neuron to the th neuron connection weight, is the th iteration dynamic pruning threshold; is the first adjustment parameter, is the median calculation function, is the element-wise product, is the layer weight matrix, is the second adjustment parameter, is the recent average reward; The parameter update rule of the deep neural network for the reward accumulation in S407 is as follows: Calculate the gradient of the loss function with respect to the weight, add it to the gradient direction processed by the sign function, then multiply by the learning rate and the reward gain coefficient, and combine the decay weighted accumulation of the historical reward to obtain the parameter update amount; The reward perception update strategy for momentum acceleration in S408 is as follows: record the historical update direction through the momentum vector, decay the old momentum with the momentum decay rate each time an update is made, and superimpose the current update amount, then apply the momentum vector to weight update.

[0015] The beneficial effects of the present invention are as follows: (1) The present invention adopts a signal-to-noise ratio-driven wavelet decomposition layer number and threshold control mechanism, dynamically adjusts the decomposition depth and threshold function according to the local standard deviation of the signal, and improves the ability to retain fault characteristics under different working conditions.

[0016] (2) The present invention dynamically determines the segmentation boundary by comparing local autocorrelation and global average autocorrelation, accurately captures mutation characteristics such as bearing faults, and effectively retains the complete structure of the faults.

[0017] (3) The present invention uses a feedback reward-driven gating mechanism, adopts historical rewards to dynamically adjust the retention weights, and realizes the enhancement of feature extraction.

[0018] (4) The present invention constructs a dynamically variable skip connection path, enables the fault propagation path to adapt to the signal characteristics, and improves the modeling ability of fault propagation.

[0019] (5) The present invention uses a reward-driven gradient correction and momentum accumulation strategy to realize sparse connection optimization and stable parameter update, and improves the long-term stability and deployability of the model under complex tasks. Description of the Drawings

[0020] Figure 1 It is the boundary detection effect diagram of the dynamic segmentation strategy of the present invention.

[0021] Figure 2 It is the overlapping distribution diagram of the features extracted by the traditional method.

[0022] Figure 3 It is the clear star-shaped radiation distribution diagram of the features extracted by the method of the present invention. Detailed Embodiment

[0023] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] A distributed underground well data processing method in this embodiment includes the following steps: S1, collect the vibration data of the air conditioner unit of the airport distributed underground well through an acceleration sensor.

[0025] During the on-site deployment phase, sensors are installed at key locations such as the unit bearings, compressor housing, and motor base. The data acquisition unit uploads data in real-time via wireless or wired means to an edge computing device or a central server for processing and storage. The sampling frequency of the acceleration sensor is from 1 kHz to 10 kHz, and the data format is a single-channel or multi-channel time series. The raw data is stored in the form of a continuous waveform and is accompanied by timestamp information to indicate its acquisition timing.

[0026] The annotation of vibration data is carried out by combining on-site inspection records, fault repair logs, and expert diagnosis results. The annotation is completed by combining manual confirmation and rule-assisted judgment. The annotation categories include "normal operation", "slight imbalance", "bearing fault", "loose anomaly", and "structural resonance". Each sample corresponds to a label, and the label participates in model training in the form of one-hot encoding. In terms of sample distribution, it is ensured that each category contains at least hundreds of samples to avoid the problem of class imbalance during model training.

[0027] S2. Based on the wavelet decomposition layer number and threshold strategy dynamically adjusted according to the signal-to-noise ratio, combined with the time-segment clustering algorithm, the collected vibration data of the air-conditioning unit is denoised.

[0028] The vibration data of the air-conditioning unit has high-frequency noise interference and non-stationary time-series characteristics. Conventional techniques use fixed-threshold wavelet denoising, which cannot adapt to the noise intensity differences of vibration signals under different working conditions and is prone to insufficient feature extraction.

[0029] The wavelet decomposition layer number and threshold strategy with dynamically adjusted signal-to-noise ratio in this embodiment are as follows: perform wavelet decomposition on the original vibration signal, dynamically adjust the decomposition layer number and threshold according to the signal-to-noise ratio of the signal, calculate the signal-to-noise ratio estimation value for each layer of wavelet coefficients, perform threshold processing on the wavelet coefficients through a dynamic threshold function, and reconstruct the processed wavelet coefficients to obtain the denoised vibration signal, which is expressed as: , In the above formula, is the vibration signal after wavelet threshold denoising, is a positive integer, is the dynamically adjusted decomposition layer number. During the calculation process, considering that the noise intensity of the air-conditioning vibration signal changes with the working conditions, such as load fluctuations and speed changes, the standard deviation reflects the local fluctuation intensity of the signal, and the greater the fluctuation, the more decomposition layers are required to remove high-frequency noise. is the layer wavelet coefficient, which is used to decompose the signal and extract different frequency components. is the original vibration signal, is the dynamic threshold function, which represents the adaptive threshold shrinkage intensity based on the signal-to-noise ratio. During the calculation process, considering the overlap between the noise and the fault characteristic frequency bands in the air conditioner vibration signal, such as structural resonance and high-frequency noise, the dynamic threshold retains the effective frequency bands, such as the characteristic frequency band of bearing faults, and suppresses the noise. is the layer wavelet coefficient signal-to-noise ratio estimation. By dynamically adjusting the threshold according to the signal-to-noise ratio, the noise can be effectively removed while retaining the fault characteristics.

[0030] is the rounding operation. is the logarithmic function, with the default base being 10. is the standard deviation function. is to calculate the standard deviation of the original vibration signal, which is used to adjust the decomposition depth according to the local fluctuation intensity of the signal.

[0031] is norm. is the exponential function with the natural constant as the base. is the slope parameter. is set to 0.3. is the signal-to-noise ratio threshold, which ensures that weak fault characteristics can still be retained at low signal-to-noise ratios. is set to 0.1.

[0032] The signal-to-noise ratio threshold is directly compared with the value of the wavelet coefficient signal-to-noise ratio estimation to determine the output behavior of the threshold function.

[0033] Considering that in the fault diagnosis task, the vibration energy of early faults is weak. That is to say, the fault characteristics of early faults are very unobvious and are easily submerged by noise. The role of the signal-to-noise ratio threshold is to control the retention and suppression degree of wavelet coefficients, so as to realize the adaptive adjustment of the denoising intensity in wavelet denoising. When specifically setting, if is too high (such as >0.2), the low signal-to-noise ratio frequency band will be overly suppressed, resulting in missed detection. Therefore, setting =0.1 can achieve the optimal balance between noise suppression and feature retention.

[0034] S3. Adopt a sliding window dynamic partitioning strategy, combine the autocorrelation coefficient within the window to determine the segmentation boundary, and perform segmented feature coding on the vibration data of the air conditioner unit.

[0035] The time correlation of the vibration signal presents local mutation characteristics under different fault modes. The conventional fixed window segmentation will destroy the continuity of key fault characteristics.

[0036] The dynamic sliding window partitioning strategy in this embodiment is as follows: For the denoised vibration signal, a sliding window is used for scanning, and the local autocorrelation coefficient at each window position is calculated. When the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, this position is taken as the segmentation boundary, and the signal is divided into multiple segments. The signals within each segment are encoded using a preset bidirectional LSTM, and the outputs of the forward and backward LSTMs are concatenated to obtain the feature vector of each segment, which is expressed as: , In the above formula, is the set of vibration signals after segmentation, is the total number of dynamic segments, is a positive integer, is the starting time point of the th segment, is the ending time point of the th segment.

[0037] is the denoised vibration signal segment from to , is the local autocorrelation coefficient at position . The calculation method takes into account that the fault characteristics of the air conditioner vibration signal often show sudden changes in local correlation, such as non-periodic vibrations of looseness abnormalities, and captures the mutation points of the vibration signal, such as the periodic impacts of bearing faults, is the boundary determination threshold, which controls the sensitivity of segmentation. A high threshold can avoid false segmentation caused by noise and ensure that the segmentation boundary corresponds to real fault events, is set to 0.6, is the global average autocorrelation coefficient, which is used as the threshold for determining the segmentation boundary.

[0038] is the window radius for calculating the local autocorrelation coefficient, which needs to adapt to the time scale of the fault characteristics, such as the impact interval of bearing faults, to avoid destroying the feature continuity with a fixed window, is set to 128, is a positive integer, is the denoised signal value of the th sampling point to the left of position , is the denoised signal value of the th sampling point to the right of position ; is the total duration of the vibration signal, is a positive integer, is the local autocorrelation coefficient at the th time point.

[0039] Each segment is encoded into a feature vector by a preset bidirectional LSTM, expressed as: , In the formula, is the bidirectional feature vector of the th segment, is the forward long short-term memory network of the preset bidirectional LSTM, is the th segment of the vibration signal, represents the vector concatenation operation, is the backward long short-term memory network of the preset bidirectional LSTM.

[0040] S4. The deep neural network model based on feedback reward is trained and optimized through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks.

[0041] The structure of the deep neural network model based on feedback reward is: It consists of multiple layers, including an input layer, multiple hidden layers, and an output layer. For example, the number of hidden layers is 5.

[0042] The input layer receives the vibration signal feature vector after denoising and segment feature encoding. The hidden layer adopts a spatio-temporal attention-guided residual propagation mechanism, combined with an adaptive gating feature selection module, to dynamically adjust the feature retention ratio and capture the temporal dependence relationship of the vibration signal.

[0043] The output features of each layer are propagated through a skip connection path to enhance the model's ability to capture key fault features.

[0044] The output layer adopts a multi-granularity prototype contrast classification strategy, and performs multi-scale matching through gating features and learnable class prototype vectors, and finally outputs the classification result of fault diagnosis.

[0045] The deep neural network structure based on feedback reward is trained and optimized through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks.

[0046] The training process of the deep neural network model based on feedback reward includes the following steps: S401. Initialize the weights of the deep neural network: Based on the weight initialization strategy of pre-classification confidence, calculate the product of the gradient expectation and the reward in the pre-training stage to make the initial weights bias towards the high-confidence feature direction.

[0047] Conventional deep neural networks usually adopt the method of randomly initializing parameters, which is prone to falling into local optima in vibration data classification.

[0048] The weight initialization strategy for the pre-classification confidence in this embodiment is as follows: In the pre-training stage, for each layer of the deep neural network, calculate the gradient of its pre-classification cross-entropy loss with respect to the weights, multiply the gradient by the pre-classification confidence reward factor to obtain the expected gradient, multiply the expected gradient by the initial learning rate, and then add it to the He normal distribution initial weights of that layer to obtain the initial weights of that layer, which is expressed as: , In the above formula is the weight matrix after initialization of the -th layer of the deep neural network, is the He normal distribution initial weights of the -th layer of the deep neural network, is the initial learning rate, which is set to 0.01, is the pre-classification cross-entropy loss, is the mathematical expectation operation, is the gradient of the pre-classification loss with respect to the weights, is the one-hot encoding of the true fault class label, is the predicted probability distribution output in the pre-classification stage, is the pre-classification confidence reward factor.

[0049] is the pre-classification correct probability, is the pre-classification error probability.

[0050] S402. Perform adaptive gated feature selection: Use a dynamic feature selection gate to implement adaptive gated feature selection, dynamically adjust the feature retention ratio through cumulative rewards, and suppress low-contribution features.

[0051] There are a large number of redundant features in the vibration signal. These features contribute little to fault diagnosis but can cause model overfitting. Conventional techniques usually use fixed feature selection methods and cannot be adjusted according to the dynamic characteristics of the data.

[0052] The method for performing adaptive gated feature selection in this embodiment is as follows: For each segmented feature vector, calculate the activation value of its dynamic feature selection gate. The activation value is jointly determined by the feature vector, the gated weight matrix, the gated bias term, and the historical average reward value. Map the activation value to the interval [0, 1] through the Sigmoid activation function to obtain the output of the feature selection gate. Use this output to weight the feature vector element by element, retain important features, and at the same time weight and fuse the unselected features with the selected features at the previous moment to obtain the final feature selection result, which is expressed as: , In the formula, is the The gating coefficient of the feature vector, which is used to determine the importance of each feature in the feature vector, is the Sigmoid activation function, is the gating weight matrix, which is used to calculate the gating coefficient, is the th segmented bidirectional feature vector, which is encoded by the bidirectional LSTM, is the gating bias term, is the reward influence coefficient, which is used to adjust the influence of the historical reward on the gating coefficient, is set to 0.2, is the historical average reward value, which is used to dynamically adjust the feature retention ratio.

[0053] is the th final retained feature vector obtained after the th vibration signal segment passes through the adaptive gating feature selection mechanism, th vibration signal segment passes through the adaptive gating feature selection mechanism, is the element-wise product.

[0054] is the average reward value of the previous is the th sample classification reward value, is the one-hot encoding of the true fault class label, is the predicted probability distribution output in the pre-classification stage.

[0055] S403. Perform the forward propagation of the data: Adopt the spatio-temporal attention-guided residual propagation mechanism to dynamically construct the skip connection path through the gating feature selection result.

[0056] The temporal dependence relationship of the vibration signal shows non-uniform distribution characteristics in different fault stages. When the conventional fully connected neural network performs forward propagation, it is difficult to capture the propagation path of the key fault features.

[0057] The spatio-temporal attention-guided residual propagation mechanism of this embodiment is: Calculate the spatio-temporal attention weight for the output features of the previous layer through the weight matrix, multiply it element-wise with the conventional linear transformation, perform non-linear projection transformation on the output features of the previous layer through multiple residual paths and sum them up, linearly interpolate the attention-weighted result and the residual path output according to the attention weight, and dynamically construct the skip connection path, which is expressed as: , In the formula, is the The spatio-temporal attention weights of the layer are used to dynamically construct the skip connection path. is the Sigmoid activation function. is the attention weight matrix of the layer. is the final retained feature vector obtained after the

[0058] adaptive gated feature selection mechanism processes the vibration signal segments. is the element-wise product. is the conventional propagation weight. is the output feature of the layer. is a positive integer, M is the maximum number of residual paths. is the normalized non-linear activation function used to normalize the output of the residual path. Let its input be , is the projection matrix of the th residual path.

[0059] is the norm. is the hyperbolic tangent function.

[0060] S404. Perform class prediction: Adopt the multi-granularity prototype contrast classification strategy, construct learnable class prototype vectors in the feature space, and determine the final class through the multi-scale matching scores between the gated features and the prototype vectors to solve the inter-class similarity problem among fault patterns.

[0061] The multi-granularity prototype contrast classification strategy of this embodiment is as follows: For each segmented feature, calculate the exponentially normalized score of its negative L2 distance from all class prototype vectors, weight the scores of each segment through the gating coefficient, use the fully connected layer to calculate the classification score for the global average feature, and finally fuse the prototype matching score and the traditional classification score proportionally to select the class with the maximum score, which is expressed as: , In the formula, is the multi-scale prototype matching score of the th class. is the total number of dynamic segments. is a positive integer. is the gating coefficient of the th segment. is the scale factor used to adjust the sensitivity of the prototype matching score and is set to 2.0. is the final retained feature vector obtained after the vibration signal segments pass through the adaptive gating feature selection mechanism. is the th class of trainable prototype vectors. To address the intra-class differences in air conditioner faults, such as bearing faults of different severities, the prototype vectors align the samples of the same class in the feature space, enhancing the classification robustness. is the norm. is the th class of trainable prototype vectors. is the currently calculated class index. is a positive integer. is a temporary index variable when traversing all classes. is a positive integer. is the exponential function with the natural constant as the base.

[0062] is the predicted probability distribution output in the pre-classification stage. denotes taking the class index that maximizes the value within the parentheses. is the fusion coefficient, used to balance the prototype matching score and the traditional classification score. is set to 0.7. is the Sigmoid activation function. is the fully connected classification weight. is the feature vector after global average pooling.

[0063] S405. Calculate the loss function: Adopt the hierarchical reward weighted cross-entropy loss to dynamically adjust the loss weights of each class, strengthening the attention to difficult-to-classify faults.

[0064] In the fault diagnosis task of air conditioner units, there are differences in the diagnosis difficulty of different fault types. The data of some fault types are relatively complex and are prone to misclassification, while the data of other fault types are relatively easy to distinguish. Traditional loss functions usually adopt a unified weight for all classes and cannot effectively distinguish these difficulty differences, easily resulting in insufficient attention to difficult-to-classify faults during the training process of the model, thus affecting the overall diagnostic performance.

[0065] The calculation method of the loss function in this embodiment is as follows: Calculate the accuracy rate of each category on the validation set, as well as the F1 scores on the training set and the validation set. Assign a weight factor to each category. This weight factor consists of two parts. One part is the normalized value based on the category accuracy rate, which is used to reflect the performance of the category on the validation set. The other part is an adjustment term based on the category F1 score, which is used to further strengthen the attention to difficult-to-classify faults. Multiply the weight factor of each category by the corresponding cross-entropy loss to obtain the weighted loss value, and sum up the weighted loss values of all categories to obtain the final total loss function value, which is expressed as: , In the above formula, is the weighted cross-entropy loss value, is the total number of fault categories, is the class dynamic reward weight, is the one-hot encoding of the true label of the th class, is the logarithmic function, is the th class prediction probability value, represents the category index calculated currently.

[0066] is the accuracy rate of the th category on the validation set. The validation set is an independent evaluation data set divided from the training data and not participating in training. is the accuracy rate of the th category on the validation set, is the balance coefficient. For example, it is set to 0.2, is the F1 score of the th category in the training set, is the F1 score of the th category in the validation set, is the temporary index variable when traversing all categories.

[0067] S406. Perform dynamic sparse connection optimization of the deep neural network: Based on the connection pruning strategy of the reward gradient, correct the gradient direction by accumulating rewards to enhance the ability to capture continuous fault patterns and reduce the model complexity.

[0068] In a deep neural network, there are a large number of redundant connections. These connections not only increase the computational complexity of the model but also may lead to overfitting. Conventional techniques usually adopt random pruning or pruning methods based on fixed thresholds to reduce the number of connections. However, these methods cannot be adjusted according to the dynamic performance of the model during training, which may lead to a decline in the performance of the pruned model.

[0069] The connection pruning strategy for the reward gradient in this embodiment is as follows: Calculate the gradient value of each connection, multiply it by the corresponding weight value to obtain an index measuring the importance of the connection, correct the gradient direction according to the cumulative reward to enhance the ability to capture continuous failure modes, calculate the threshold for each connection, which consists of two parts. One part is the median based on the absolute value of the gradient, used to reflect the importance of the connection, and the other part is an adjustment term based on the recent average reward, used to dynamically adjust the strictness of pruning. Generate a mask matrix according to the calculated threshold and the importance index of the connection, set the unimportant connections to zero, and achieve dynamic sparse connection optimization, which is expressed as: , In the formula, is the mask matrix of the th layer, is a positive integer, is the indicator function, is the partial derivative symbol, is the connection weight from the th neuron to the th neuron in the th layer, is the dynamic pruning threshold at the th iteration, is the weighted cross-entropy loss value. is the first adjustment parameter, is set to 0.3, is the median calculation function, is the element-wise product, is the weight matrix of the th layer, is the second adjustment parameter, is set to 0.5, is the recent average reward.

[0070] S407. Calculate the parameter update amount: Adopt the parameter update rule of the deep neural network based on the reward accumulation amount.

[0071] The parameter update rule of the deep neural network for the reward accumulation amount in this embodiment is as follows: Calculate the gradient of the loss function with respect to the weight, add it to the gradient direction processed by the sign function, then multiply by the learning rate and the reward gain coefficient, and combine with the decay weighted cumulative amount of the historical reward to obtain the parameter update amount, which is expressed as: , In the formula, is the weight update amount of the th layer, is the weight update learning rate, is set to 0.01, is the gradient of the loss function with respect to the weight, is the reward gain coefficient, set to 0.3, is the historical reward accumulation with decay weighting, is the sign function.

[0072] is the current training cycle number, is the reward decay factor, set to 0.9, is the decay weight of the iteration, representing the reward weighted sum decayed over time, represents the true label of the iteration, and represents the predicted output of the model at the iteration, is the pre-classification confidence reward factor for the

[0073] S408. Perform parameter update of the deep neural network: Adopt a reward-aware update strategy with momentum acceleration.

[0074] Traditional parameter update methods are prone to model oscillations due to data distribution drift in online scenarios.

[0075] The reward-aware update strategy with momentum acceleration in this embodiment is: Record the historical update direction through the momentum vector. At each update, decay the old momentum using the momentum decay rate and superimpose the current update amount. Apply the momentum vector to the weight update, expressed as: , where, is the updated weight matrix of the layer, is the weight matrix of the layer before update, is the momentum vector of the layer, is the element-wise product, is the weight update amount of the

[0076] is the momentum decay rate, set to 0.9.

[0077] In the deep neural network of the present invention, model training parameters including a gated weight matrix, a gated bias term, an attention weight matrix, a conventional propagation weight, and a fully connected classification weight, as well as trainable prototype vectors, are optionally optimized using the reward accumulation update rule of step S407 and the momentum acceleration strategy of step S408, or can be updated using the traditional stochastic gradient descent method according to computational resource limitations.

[0078] S409. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. The preset iteration stop condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0079] S5. Use the trained deep neural network model based on feedback rewards to perform fault diagnosis and classification on the vibration data of the air-conditioning unit.

[0080] After the real-time collected air-conditioning vibration signal is processed by dynamic wavelet denoising, sliding window autocorrelation analysis is used for adaptive segmentation. Each segment extracts temporal features through a pre-trained bidirectional LSTM, and then key features are screened by a dynamic gating module. The model captures the propagation path of fault features through a spatio-temporal attention mechanism, calculates the matching degree between each segment feature and the stored class prototype vector in a multi-scale space, and combines the global classification results of the fully connected layer for weighted decision-making, and finally outputs the fault category with the highest confidence.

[0081] As Figure 1 shown, the experimental verification of this embodiment is as follows: To compare the protection effect of adaptive segmentation and traditional fixed windows on feature continuity, sequence data with various typical state changes is constructed to characterize various working condition fluctuations that may occur in the ground well air-conditioning equipment. In the experiment, the time series t is defined as 2000 uniform sampling points from 0 to 2 seconds, and the corresponding vibration signal is composed of the following four segments of signals with different modes spliced together: ① The first 500 sampling points (0 - 0.5 seconds): stable vibration data when the equipment is operating normally; ② The middle 500 sampling points (0.5 - 1.0 seconds): impact response data during a sudden fault; ③ The third segment of 500 sampling points (1.0 - 1.5 seconds): mechanical looseness or periodic deviation data; ④ The last 500 sampling points (1.5 - 2.0 seconds): data with external noise interference.

[0082] The four spliced signals form a composite signal with typical feature changes, which is used to verify the response ability of the segmentation method at signal mutation points.

[0083] In this experiment, the determination of the segmentation boundary is completed using a dynamic threshold strategy, and the main calculation process is as follows: ① Calculate the autocorrelation coefficient, perform autocorrelation analysis on the complete signal, and obtain the symmetric autocorrelation intensity sequence corresponding to each time point, which is used to measure the repeatability or stability of the local waveform structure; ② Based on the global mean of the autocorrelation sequence, set the segmentation determination threshold to 0.65 times of it to filter out the points with significant structural mutations; ③ Traverse all time points, find the positions where their autocorrelation values are higher than this threshold as possible paragraph boundary points. The boundary is regarded as the inflection point where the signal state changes from stable to impact or from ordered to disordered, characterizing the mutation characteristics of the signal structure.

[0084] To compare the protection effects of adaptive segmentation and traditional fixed window on feature continuity, the purple curve in the figure is a composite signal containing steady-state vibration, transient impact, modulation waveform and random noise, the cyan curve is the change of autocorrelation coefficient, the red dashed line is the dynamic determination threshold, and the orange vertical lines mark the detected segmentation boundaries. It can be seen from the experiment that in the transition region where the signal changes from steady state to impact response, the autocorrelation coefficient rises rapidly and breaks through the threshold, and the boundary detection accurately locks the moment of state mutation. At the junction of the modulation signal and random noise, the traditional fixed window will cause feature truncation, while the dynamic segmentation automatically adjusts the window length according to the local correlation change, so that the features within each segment are consistent. The adaptive division based on the intrinsic characteristics of the signal solves the problem that sudden features are easily segmented in mechanical fault diagnosis.

[0085] Respectively through the visualization method after feature dimensionality reduction, compare the feature vectors extracted by the traditional method and the method of the present invention, and show the separability performance of the two feature extraction methods in the feature space. The five typical faults in the dataset are "normal", "unbalanced", "fault", "loose", and "resonance". 300 sample points are selected for each category and projected into a two-dimensional space.

[0086] The traditional method uses a deep neural network based on gradient descent and error backpropagation for feature extraction.

[0087] The experiment maps the high-dimensional features to a two-dimensional plane through the t-SNE algorithm to observe their distribution characteristics. The main calculation process is as follows: ① Respectively perform feature extraction on two groups of 1500 two-dimensional sample data (5 categories × 300 pieces); ② Use the sklearn.manifold.TSNE module for non-linear dimensionality reduction, and set the parameters n_components = 2 and perplexity = 30; ③ Respectively perform dimensionality reduction mapping on the feature data of the traditional method and the method of the present invention, and compress them into a two-dimensional space for intuitive visualization; ④ Render the coordinate values of the processed two-dimensional data points in different colors, and use colors to mark each type of fault, reflecting the distribution pattern and classification potential.

[0088] In the feature distribution after feature extraction by traditional methods, the distribution boundaries of each type of sample are not obvious, the means are slightly different, and the covariance is large, indicating that the features extracted by traditional methods have strong inter-class overlap, fuzzy boundaries, and poor clustering.

[0089] For the feature data of the method of the present invention, the distribution of each type of sample is compact, the means of each type are far apart, and the covariance is small, indicating that after the present invention uses the prototype contrast learning mechanism and gated feature selection, the extracted features have strong intra-class compactness and inter-class separability.

[0090] As Figure 2 and Figure 3 shown, by visualizing the feature space distribution to compare the classification potential of different methods, the manifold learning algorithm is used to project the high-dimensional features onto a two-dimensional plane. Figure 2 It shows that the features extracted by traditional methods are overlapping, especially there are a large number of overlapping areas between the bearing fault and looseness abnormality categories. Figure 3 It shows that the features extracted by the method of the present invention present a clear star-shaped radiation distribution. The five categories each form a tight cluster with distinct boundaries. The color of the points in the figure represents different fault types, and the distribution pattern reflects the discrimination ability of the model, proving that the multi-granularity prototype contrast strategy constructs learnable category prototype vectors and establishes a more discriminative metric relationship in the feature space, significantly improving the inter-class separation.

[0091] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A distributed underground well data processing method, characterized in that: It includes the following steps: S1. Collect the vibration data of the air conditioning unit of the airport distributed ground well through an acceleration sensor; S2. Based on the wavelet decomposition layer number and threshold strategy with dynamic SNR adjustment, combined with the time segmentation clustering algorithm, denoise the collected vibration data of the air conditioning unit; S3. Adopt a sliding window dynamic division strategy, combined with the autocorrelation coefficient within the window to determine the segmentation boundary, and perform segmented feature encoding on the vibration data of the air conditioning unit; S4. The deep neural network model based on feedback reward is trained and optimized through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks; S5. Use the trained deep neural network model based on feedback reward to perform fault diagnosis and classification on the vibration data of the air conditioning unit.

2. The distributed wellbore data processing method according to claim 1, wherein The wavelet decomposition layer number and threshold strategy with dynamic SNR adjustment in S2 are as follows: Perform wavelet decomposition on the original vibration signal, dynamically adjust the decomposition layer number and threshold according to the SNR of the signal, calculate the SNR estimation value of each layer of wavelet coefficients, perform threshold processing on the wavelet coefficients through a dynamic threshold function, and reconstruct the processed wavelet coefficients to obtain the denoised vibration signal, which is expressed as: ， In the above formula, is the vibration signal after wavelet threshold denoising, is a positive integer, is the number of dynamic decomposition levels, is the -th layer wavelet coefficient, is the original vibration signal, is the dynamic threshold function, is the -th layer wavelet coefficient signal-to-noise ratio estimation; is a rounding operation, is a logarithmic function, is a standard deviation function, is to calculate the standard deviation of the original vibration signal; is norm; is the exponential function with the natural constant as the base, is the slope parameter, is the signal-to-noise ratio threshold, = 0.

1.

3. The distributed wellbore data processing method according to claim 1, wherein The sliding window dynamic division strategy in S3 is as follows: For the denoised vibration signal, use a sliding window to scan, calculate the local autocorrelation coefficient at each window position, and when the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, use this position as the segmentation boundary, divide the signal into multiple segments, use a preset bidirectional LSTM to perform feature encoding on the signal within each segment, and splice the outputs of the forward and reverse LSTMs to obtain the feature vector of each segment, which is expressed as: ， In the above formula, is the set of vibration signals after segmentation, is the total number of dynamic segmentations, is a positive integer, is the starting time point of the th segmentation, is the ending time point of the th segmentation, is the denoised vibration signal segment from to , is the local autocorrelation coefficient at position , is the boundary determination threshold, is the global average autocorrelation coefficient; is the window radius for calculating the local autocorrelation coefficient, is a positive integer, is the position the denoised signal value of the th sampling point on the left side of the position, is the position the denoised signal value of the th sampling point on the right side of the position; is the total duration of the vibration signal, is a positive integer, is the local autocorrelation coefficient at the 4. The distributed wellbore data processing method according to claim 1, wherein The training process of the deep neural network model based on feedback reward in S4 includes the following steps: S401. Initialize the weights of the deep neural network: Based on the weight initialization strategy of pre-classification confidence, calculate the product of the gradient expectation and the reward during the pre-training stage to make the initial weights bias towards the high-confidence feature direction; S402. Perform adaptive gating feature selection: Use a dynamic feature selection gate to achieve adaptive gating feature selection, dynamically adjust the feature retention ratio through cumulative rewards, and suppress low-contribution features; S403. Perform forward propagation of data: Adopt a spatio-temporal attention-guided residual propagation mechanism to dynamically construct a skip connection path through the results of gating feature selection; S404. Perform class prediction: Adopt a multi-granularity prototype contrast classification strategy, construct a learnable class prototype vector in the feature space, and determine the final class through the multi-scale matching score between the gating feature and the prototype vector; S405. Calculate the loss function: Adopt a hierarchical reward weighted cross-entropy loss to dynamically adjust the loss weights of each class, and strengthen the attention to difficult-to-classify faults; S406. Perform dynamic sparse connection optimization of the deep neural network: Based on the connection pruning strategy of the reward gradient, correct the gradient direction through cumulative rewards to enhance the ability to capture continuous fault patterns; S407. Calculate the parameter update amount: Adopt a deep neural network parameter update rule based on the reward accumulation amount; S408. Update the parameters of the deep neural network: adopt a reward-aware update strategy with momentum acceleration; S409. Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed.

5. The distributed wellbore data processing method according to claim 4, characterized in that, The weight initialization strategy of the pre-classification confidence in S401 is as follows: in the pre-training stage, for each layer of the deep neural network, calculate the gradient of its pre-classification cross-entropy loss with respect to the weight, multiply the gradient by the pre-classification confidence reward factor to obtain the gradient expectation, multiply the gradient expectation by the initial learning rate, and then add it to the He normal distribution initial weight of this layer to obtain the initial weight of this layer.

6. The distributed downhole data processing method according to claim 4, wherein The method for performing adaptive gating feature selection in S402 is as follows: for each segmented feature vector, calculate the activation value of its dynamic feature selection gate, map the activation value to the interval [0, 1] through the Sigmoid activation function to obtain the output of the feature selection gate, use this output to weight the feature vector element by element, retain the important features, and at the same time weight and fuse the unselected features with the selected features at the previous moment to obtain the final feature selection result.

7. The distributed downhole data processing method according to claim 4, wherein The spatio-temporal attention-guided residual propagation mechanism in S403 is as follows: calculate the spatio-temporal attention weights for the output features of the previous layer through the weight matrix, multiply it element by element with the conventional linear transformation, perform non-linear projection transformation on the output features of the previous layer through multiple residual paths and sum them, linearly interpolate the attention-weighted result and the residual path output according to the attention weights, and dynamically construct the skip connection path.

8. The distributed wellbore data processing method according to claim 4, wherein The multi-granularity prototype contrast classification strategy in S404 is as follows: for each segmented feature, calculate the exponential normalized score of its negative L2 distance from all category prototype vectors, weight the scores of each segment through the gating coefficient, use the fully connected layer to calculate the classification score for the global average feature, and finally fuse the prototype matching score and the traditional classification score proportionally, and select the category with the maximum score, which is expressed as: ， Wherein, is the multi-scale prototype matching score of the th class, is the total number of dynamic segments, is a positive integer, is the gating coefficient of the th segment, is the scale factor, is the th final retained feature vector obtained after the vibration signal segment passes through the adaptive gating feature selection mechanism, is the th class of trainable prototype vectors, is norm, is the total number of fault classes, is the th class of trainable prototype vectors, is the currently calculated class index, is a temporary index variable when traversing all classes, is the exponential function with the natural constant as the base; is the predicted probability distribution output in the pre-classification stage, denotes taking the class index that maximizes the value within the brackets, is the fusion coefficient, is the Sigmoid activation function, is the fully connected classification weight, is the feature vector after global average pooling.

9. The distributed wellbore data processing method according to claim 4, wherein The calculation method of the loss function in S405 is as follows: calculate the accuracy of each category on the validation set, as well as the F1 scores on the training set and the validation set, assign a weight factor to each category, multiply the weight factor of each category by the corresponding cross-entropy loss to obtain the weighted loss value, and sum the weighted loss values of all categories to obtain the final total loss function value.

10. The distributed downhole data processing method according to claim 4, characterized in that, The connection pruning strategy of the reward gradient in S406 is as follows: calculate the gradient value of each connection, multiply it by the corresponding weight value to obtain an index measuring the importance of the connection, correct the gradient direction according to the cumulative reward to enhance the ability to capture continuous failure modes, calculate the threshold of each connection, generate a mask matrix according to the calculated threshold and the importance index of the connection, and set the unimportant connections to zero to achieve dynamic sparse connection optimization, which is expressed as: ， In the formula, is the layer mask matrix, is a positive integer, is the indicator function, is the partial derivative symbol, is the weighted cross-entropy loss value, is the th neuron in the th layer to the th neuron connection weight; is the first adjustment parameter, is the median calculation function, is the element-wise product, is the layer weight matrix, is the second adjustment parameter, is the recent average reward; The parameter update rule of the deep neural network for the reward accumulation in S407 is as follows: calculate the gradient of the loss function with respect to the weight, add it to the gradient direction processed by the sign function, then multiply by the learning rate and the reward gain coefficient, and combine the decay-weighted cumulative amount of the historical reward to obtain the parameter update amount; The reward perception update strategy for momentum acceleration in S408 is as follows: record the historical update direction through the momentum vector, decay the old momentum with the momentum decay rate each time an update is made, and superimpose the current update amount, then apply the momentum vector to the weight update.

Images