A distributed well data processing method

Abstract

The present invention discloses a distributed ground well data processing method, which relates to the field of data processing technology. The method includes collecting vibration data of air conditioning units in distributed ground wells at airports using acceleration sensors; denoising the collected vibration data of the air conditioning units using a wavelet decomposition layer number and threshold strategy based on dynamic signal-to-noise ratio adjustment, combined with a time segment clustering algorithm; performing segmented feature encoding on the vibration data of the air conditioning units; training and optimizing a deep neural network model based on feedback rewards through a feedback reward mechanism to improve the model's performance in complex fault diagnosis tasks; and using the trained deep neural network model based on feedback rewards to perform fault diagnosis and classification on the vibration data of the air conditioning units. The present invention adopts a signal-to-noise ratio-driven wavelet decomposition layer number and threshold control mechanism, dynamically adjusting the decomposition depth and threshold function based on the local standard deviation of the signal, thereby improving the ability to retain fault characteristics under different operating conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a distributed 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 and even predicting equipment failures and taking effective containment measures, as well as analyzing the causes of equipment failures are important issues that equipment managers need to address.

[0003] Air conditioning units in airport underground shafts are critical for maintaining the temperature and humidity environment in terminal buildings and apron underground facilities. Their stable operation directly impacts the efficiency of airport energy systems and equipment lifespan. Vibration data, a core indicator of unit health, can provide early warning of faults. Abnormal vibrations may indicate mechanical failures such as bearing wear, impeller imbalance, and motor misalignment, requiring proactive intervention to avoid downtime. Furthermore, vibration characteristics can reflect unit load status, assisting in the dynamic adjustment of 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 shortcomings of the above-mentioned prior art and provide a distributed well data processing method.

[0005] The technical solution adopted to solve the above technical problems is: a distributed ground well data processing method, comprising the following steps: S1, collecting vibration data of air-conditioning units in distributed ground wells of airports through acceleration sensors; S2, denoising the collected vibration data of air-conditioning units based on the wavelet decomposition layer number and threshold strategy with dynamic adjustment of signal-to-noise ratio, combined with a time segmentation clustering algorithm; S3, adopting a sliding window dynamic division strategy, combining the autocorrelation coefficient within the window to determine the segmentation boundary, and performing segmentation feature encoding on the vibration data of the air-conditioning units; S4, training and optimizing the deep neural network model based on feedback reward through the feedback reward mechanism, so as 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 units.

[0006] Furthermore, the wavelet decomposition layer number and threshold strategy for dynamic adjustment of the signal-to-noise ratio in S2 are as follows: perform wavelet decomposition on the original vibration signal, dynamically adjust the decomposition layer number and threshold value according to the signal-to-noise ratio of the signal, calculate the signal-to-noise ratio estimate for each layer of wavelet coefficients, perform threshold processing on the wavelet coefficients using a dynamic threshold function, and reconstruct the processed wavelet coefficients to obtain the denoised vibration signal, which is expressed as:

[0007] ,

[0008] In the above formula, is the vibration signal after wavelet threshold denoising, is a positive integer, is the number of dynamic decomposition layers, For the Layer wavelet coefficients, is the original vibration signal, is the dynamic threshold function, For the Layer wavelet coefficient signal-to-noise ratio estimation;

[0009] For rounding operation, is a logarithmic function, is the standard deviation function, It is to find the standard deviation of the original vibration signal;

[0010] for norm;

[0011] is an exponential function with a natural constant as its base, is the slope parameter, is the signal-to-noise ratio threshold, =0.1.

[0012] Furthermore, the dynamic partitioning strategy of the sliding window in S3 is as follows: the denoised vibration signal is scanned using a sliding window, and the local autocorrelation coefficient of each window position is calculated. When the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, the position is used as the segmentation boundary to divide the signal into multiple segments. The signal in each segment is feature encoded using a preset bidirectional LSTM, and the forward and reverse LSTM outputs are spliced to obtain the feature vector of each segment, which is expressed as:

[0013] ,

[0014] In the above formula, is the segmented vibration signal set, is the total number of dynamic segments, is a positive integer, For the The starting time point of each segment, For the The end time point of each segment, For arrive The denoised vibration signal fragment, For location The local autocorrelation coefficient of is the boundary determination threshold, is the global average autocorrelation coefficient;

[0015] The window radius for calculating the local autocorrelation coefficient, is a positive integer, For location Left side The denoised signal value of the sampling point, For location Right side The denoised signal value of each sampling point;

[0016] is the total duration of the vibration signal, is a positive integer, For the The local autocorrelation coefficient at each time point.

[0017] Furthermore, the training process of the deep neural network model based on feedback reward in S4 includes the following steps:

[0018] S401. Initialize the weights of the deep neural network: Based on the weight initialization strategy of the pre-classification confidence, calculate the gradient expectation and reward product in the pre-training phase, so that the initial weights are biased towards the direction of high-confidence features;

[0019] S402, perform adaptive gated feature selection: Adopt 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;

[0020] S403. Perform forward propagation of data: Use a residual propagation mechanism guided by spatiotemporal attention to dynamically construct skip connection paths based on gated feature selection results.

[0021] S404. Perform category prediction: Use a multi-granularity prototype comparison classification strategy to construct a learnable class prototype vector in the feature space, and determine the final category through the multi-scale matching score between the gated features and the prototype vector;

[0022] S405. Calculate the loss function: Use layered reward-weighted cross entropy loss to dynamically adjust the loss weights of each category, and strengthen the focus on difficult-to-classify faults;

[0023] S406. Perform dynamic sparse connection optimization of deep neural networks: A connection pruning strategy based on reward gradients modifies the gradient direction through cumulative rewards to enhance the ability to capture persistent failure modes.

[0024] S407, calculating the parameter update amount: using a deep neural network parameter update rule based on the reward accumulation amount;

[0025] S408. Update the parameters of the deep neural network: adopt a momentum-accelerated reward-aware update strategy;

[0026] S409: Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0027] Furthermore, the weight initialization strategy of the pre-classification confidence in S401 is: in the pre-training stage, for each layer of the deep neural network, the gradient of its pre-classification cross entropy loss with respect to the weight is calculated, the gradient is multiplied by the pre-classification confidence reward factor to obtain the gradient expectation, the gradient expectation is multiplied by the initial learning rate, and then added to the initial weight of the He normal distribution of the layer to obtain the initial weight of the layer.

[0028] Furthermore, the method for performing adaptive gated feature selection in S402 is: for each segmented feature vector, calculate the activation value of its dynamic feature selection gate, map the activation value to the [0, 1] interval through the Sigmoid activation function, obtain the output of the feature selection gate, use the output to weight the feature vector element by element, retain important features, and at the same time weightedly fuse the unselected features with the selected features at the previous moment to obtain the final feature selection result.

[0029] Furthermore, the residual propagation mechanism guided by spatiotemporal attention in S403 is as follows: the spatiotemporal attention weights are calculated for the output features of the previous layer through a weight matrix, which is multiplied element-by-element with a conventional linear transformation, and the output features of the previous layer are nonlinearly projected and summed through multiple residual paths. The attention weighted results and the residual path outputs are linearly interpolated according to the attention weights, and a jump connection path is dynamically constructed.

[0030] Furthermore, the multi-granularity prototype comparison classification strategy in S404 is as follows: for each segment feature, the exponentially normalized score of its negative L2 distance with all category prototype vectors is calculated, the score of each segment is weighted by the gating coefficient, and the classification score of the global average feature is calculated using a fully connected layer. Finally, the prototype matching score and the traditional classification score are proportionally fused, and the category with the largest score is selected, which is expressed as:

[0031] ,

[0032] Where, For the Multi-scale prototype matching score of the class, is the total number of dynamic segments, is a positive integer, For the The gating coefficient of each segment, is the scale factor, For the The final retained feature vector obtained by the adaptive gated feature selection mechanism for each vibration signal segment is: For the class trainable prototype vector, for norm, is the total number of fault categories, For the class trainable prototype vector, is the category index currently being calculated, It is a temporary index variable when traversing all categories. is an exponential function with a natural constant as its base;

[0033] is the predicted probability distribution output from the pre-classification stage, Indicates taking the category index that maximizes the value in 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.

[0034] Furthermore, the calculation method of the loss function in S405 is as follows: calculate the accuracy of each category on the validation set, and the F1 score 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 a weighted loss value, and sum the weighted loss values of all categories to obtain the final total loss function value.

[0035] Furthermore, the connection pruning strategy of the reward gradient in S406 is as follows: the gradient value of each connection is calculated and multiplied by the corresponding weight value to obtain an indicator to measure the importance of the connection, the gradient direction is corrected according to the cumulative reward to enhance the ability to capture persistent failure modes, the threshold of each connection is calculated, and a mask matrix is generated based on the calculated threshold and the connection importance indicator to set unimportant connections to zero, thereby achieving dynamic sparse connection optimization, which is expressed as:

[0036] ,

[0037] Where, For the layer mask matrix, is a positive integer, is the indicator function, is the symbol of partial derivative, is the weighted cross entropy loss value, For the Tier neurons to the The connection weights of neurons, For the Dynamic pruning threshold at iterations; is the first adjustment parameter, is the median calculation function, is the element-wise product, For the layer weight matrix, is the second adjustment parameter, is the recent average reward;

[0038] The deep neural network parameter update rule of the reward accumulation amount in S407 is as follows: calculate the gradient of the loss function with respect to the weight, add it to the gradient direction after the sign function processing, multiply it by the learning rate and the reward gain coefficient, and combine it with the decay weighted accumulation of historical rewards to obtain the parameter update amount;

[0039] The reward perception update strategy of momentum acceleration in S408 is as follows: the historical update direction is recorded through the momentum vector, the old momentum is decayed by the momentum decay rate at each update, and the current update amount is superimposed, and the momentum vector is applied to the weight update.

[0040] The beneficial effects of the present invention are as follows: (1) The present invention adopts a signal-to-noise ratio-driven wavelet decomposition layer 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.

[0041] (2) The present invention dynamically determines the segmentation boundary by comparing the local autocorrelation with the global average autocorrelation, accurately captures the sudden change characteristics of bearing faults, and effectively preserves the complete structure of the fault.

[0042] (3) The present invention uses a feedback reward-driven gating mechanism to dynamically adjust the retention weight using historical rewards to enhance feature extraction.

[0043] (4) The present invention constructs a dynamically changeable jump connection path, so that the fault propagation path adapts to changes in signal characteristics, thereby improving the modeling capability of fault propagation.

[0044] (5) The present invention uses a reward-driven gradient correction and momentum accumulation strategy to achieve sparse connection optimization and stable parameter update, thereby improving the long-term stability and deployability of the model under complex tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a diagram showing the boundary detection effect of the dynamic segmentation strategy of the present invention.

[0046] Figure 2 The features extracted by traditional methods are in overlapping distribution.

[0047] Figure 3 The features extracted by the method of the present invention present a clear star-shaped radiation distribution diagram. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] A distributed well data processing method of this embodiment includes the following steps:

[0050] S1, collects vibration data of the air conditioning units in the airport's distributed ground wells through acceleration sensors.

[0051] During field deployment, sensors are installed in key locations such as unit bearings, compressor housings, and motor bases. Data acquisition units upload data in real time to edge computing devices or central servers via wireless or wired means for processing and storage. The accelerometer sampling frequency ranges from 1kHz to 10kHz, and the data format is single-channel or multi-channel time series. Raw data is stored as a continuous waveform with a timestamp to indicate acquisition timing.

[0052] Vibration data is labeled by combining on-site inspection records, fault repair logs, and expert diagnosis results. This is accomplished through a combination of manual verification and rule-based judgment. The labeling categories include "normal operation," "minor imbalance," "bearing fault," "looseness abnormality," and "structural resonance." Each sample is assigned a label, and the labels are one-hot encoded for model training. Sample distribution is ensured to contain at least several hundred samples per category to avoid class imbalance during model training.

[0053] S2, based on the wavelet decomposition layer number and threshold strategy with dynamic adjustment of signal-to-noise ratio, combined with the time segment clustering algorithm, the collected vibration data of the air-conditioning unit is denoised.

[0054] The vibration data of air-conditioning units has high-frequency noise interference and non-stationary time series characteristics. Conventional technology uses fixed-threshold wavelet denoising, which cannot adapt to the differences in noise intensity of vibration signals under different working conditions and easily leads to insufficient feature extraction.

[0055] The wavelet decomposition layer number and threshold value strategy for dynamic adjustment of the signal-to-noise ratio in this embodiment are as follows: the original vibration signal is subjected to wavelet decomposition, the decomposition layer number and threshold value are dynamically adjusted according to the signal-to-noise ratio of the signal, the signal-to-noise ratio estimation value is calculated for each layer of wavelet coefficients, the wavelet coefficients are thresholded using a dynamic threshold function, and the processed wavelet coefficients are reconstructed to obtain the denoised vibration signal, which is expressed as:

[0056] ,

[0057] In the above formula, is the vibration signal after wavelet threshold denoising, is a positive integer, In order to dynamically decompose the number of layers, during the calculation process, the noise intensity of the air conditioner vibration signal changes with the working conditions, such as load fluctuations, speed changes, etc. The standard deviation reflects the intensity of the local fluctuation of the signal. The greater the fluctuation, the more decomposition layers are required to remove high-frequency noise. For the Layer wavelet coefficients are used to decompose the signal and extract different frequency components. is the original vibration signal, is a dynamic threshold function, which characterizes the adaptive threshold contraction intensity based on the signal-to-noise ratio. During the calculation process, the overlap of noise and fault characteristic frequency bands in the air conditioner vibration signal, such as structural resonance and high-frequency noise, is considered. The dynamic threshold retains the effective frequency band, such as the characteristic frequency band of bearing fault, and suppresses noise. For the Layer wavelet coefficient signal-to-noise ratio estimation, by dynamically adjusting the threshold based on the signal-to-noise ratio, can effectively remove noise while retaining fault characteristics.

[0058] For rounding operation, For logarithmic function, the default base is 10. is the standard deviation function, It is to find the standard deviation of the original vibration signal and adjust the decomposition depth according to the local fluctuation intensity of the signal.

[0059] for norm;

[0060] is an exponential function with a natural constant as its base, is the slope parameter, Set to 0.3, is the signal-to-noise ratio threshold, ensuring that weak fault features are retained even at low signal-to-noise ratios. Set to 0.1.

[0061] Signal-to-noise ratio threshold Direct SNR estimation with wavelet coefficients The numerical comparison determines the output behavior of the threshold function.

[0062] Considering that in the fault diagnosis task, the vibration energy of the early fault is weak, that is, the fault characteristics of the early fault are not obvious and are easily submerged by noise, the signal-to-noise ratio threshold is The function of is to control the retention and suppression degree of wavelet coefficients, so as to realize adaptive denoising intensity adjustment in wavelet denoising. Too high (such as >0.2), low signal-to-noise ratio frequency bands will be over-suppressed, resulting in missed detection, so setting =0.1 can achieve the best balance between noise suppression and feature preservation.

[0063] S3 adopts the sliding window dynamic partitioning strategy, combines the autocorrelation coefficient within the window to determine the segmentation boundary, and performs segmentation feature encoding on the vibration data of the air-conditioning unit.

[0064] The temporal correlation of vibration signals presents local mutation characteristics under different fault modes, and conventional fixed window segmentation will destroy the continuity of key fault characteristics.

[0065] The sliding window dynamic partitioning strategy of this embodiment is as follows: the denoised vibration signal is scanned using a sliding window, and the local autocorrelation coefficient of each window position is calculated. When the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, the position is used as the segmentation boundary to divide the signal into multiple segments. The signal in each segment is feature encoded using a preset bidirectional LSTM, and the forward and reverse LSTM outputs are spliced to obtain the feature vector of each segment, which is expressed as:

[0066] ,

[0067] In the above formula, is the segmented vibration signal set, is the total number of dynamic segments, is a positive integer, For the The starting time point of each segment, For the The end time point of each segment.

[0068] For arrive The denoised vibration signal fragment, For location The local autocorrelation coefficient is calculated by considering that the fault characteristics of the air conditioner vibration signal are often manifested as local correlation mutations, such as aperiodic vibrations caused by looseness abnormalities, and capturing the mutation points of the vibration signal, such as the periodic impact of bearing faults. The threshold for boundary determination controls the sensitivity of the segmentation. A high threshold can avoid false segmentation caused by noise and ensure that the segmentation boundary corresponds to the real fault event. Set to 0.6, is the global average autocorrelation coefficient, which is used to determine the threshold of segment boundaries.

[0069] The window radius for calculating the local autocorrelation coefficient needs to be adapted to the time scale of the fault characteristics, such as the impact interval of the bearing fault, to avoid the fixed window destroying the continuity of the characteristics. Set to 128, is a positive integer, For location Left side The denoised signal value of the sampling point, For location Right side The denoised signal value of each sampling point;

[0070] is the total duration of the vibration signal, is a positive integer, For the The local autocorrelation coefficient at each time point.

[0071] Each segment is encoded into a feature vector through a preset bidirectional LSTM, expressed as:

[0072] ,

[0073] Where, For the piecewise bidirectional eigenvectors, It is a preset bidirectional LSTM forward long short-term memory network. For the Segmented vibration signals, Represents vector concatenation operation, It is a reverse long short-term memory network of the preset bidirectional LSTM.

[0074] S4, the deep neural network model based on feedback reward is trained and optimized through the feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks.

[0075] The structure of the deep neural network model based on feedback reward is:

[0076] 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.

[0077] The input layer receives the vibration signal feature vector after denoising and segmented feature encoding. The hidden layer adopts the residual propagation mechanism guided by spatiotemporal attention, combined with an adaptive gated feature selection module, to dynamically adjust the feature retention ratio and capture the temporal dependency of the vibration signal.

[0078] The output features of each layer are propagated through skip connection paths to enhance the model's ability to capture key fault features.

[0079] The output layer adopts a multi-granularity prototype comparison classification strategy, performs multi-scale matching between gated features and learnable class prototype vectors, and finally outputs the classification results of fault diagnosis.

[0080] The feedback-reward-based deep neural network structure is trained and optimized through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks.

[0081] The training process of the deep neural network model based on feedback reward includes the following steps:

[0082] S401. Initialize the weights of the deep neural network: Based on the weight initialization strategy of the pre-classification confidence, calculate the gradient expectation and reward product through the pre-training stage, so that the initial weights are biased towards the direction of high-confidence features.

[0083] Conventional deep neural networks usually use random parameter initialization methods, which can easily fall into local optimality in vibration data classification.

[0084] The weight initialization strategy for the pre-classification confidence of this embodiment is as follows: in the pre-training stage, for each layer of the deep neural network, the gradient of the pre-classification cross entropy loss with respect to the weight is calculated, the gradient is multiplied by the pre-classification confidence reward factor to obtain the expected gradient, the expected gradient is multiplied by the initial learning rate, and then added to the initial weight of the He normal distribution of the layer to obtain the initial weight of the layer, which is expressed as:

[0085] ,

[0086] In the above formula For deep neural network The weight matrix after layer initialization, For deep neural network The He normal distribution initial weight of the layer, To initialize the learning rate, Set to 0.01, is the cross entropy loss for pre-classification, is the mathematical expectation operation, is the gradient of the pre-classification loss with respect to the weight, is the one-hot encoding of the true fault category label, is the predicted probability distribution output from the pre-classification stage, is the pre-classification confidence reward factor.

[0087] is the correct probability of pre-classification, is the pre-classification error probability.

[0088] S402. Perform adaptive gated feature selection: Adopt 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.

[0089] Vibration signals contain a large number of redundant features that contribute little to fault diagnosis but can lead to model overfitting. Conventional techniques typically use fixed feature selection methods that cannot be adjusted to the dynamic characteristics of the data.

[0090] The method for adaptive gated feature selection in this embodiment is as follows: for each segmented feature vector, the activation value of its dynamic feature selection gate is calculated. The activation value is jointly determined by the feature vector, the gate weight matrix, the gate bias term, and the historical average reward value. The activation value is mapped to the interval [0, 1] using the Sigmoid activation function to obtain the output of the feature selection gate. The output is used to weight the feature vector element by element, retaining important features. At the same time, the unselected features are weighted and fused with the selected features at the previous moment to obtain the final feature selection result, which is expressed as:

[0091] ,

[0092] Where, For the The gating coefficient of the feature vector is used to determine the importance of each feature in the feature vector. is the Sigmoid activation function, is the gating weight matrix, used to calculate the gating coefficient, For the The bidirectional feature vector of each segment is obtained by bidirectional LSTM encoding. is the gate bias term, is the reward influence coefficient, which is used to adjust the impact of historical rewards on the gating coefficient. Set to 0.2, It is the historical average reward value, which is used to dynamically adjust the feature retention ratio.

[0093] For the The final retained feature vector obtained by the adaptive gated feature selection mechanism for each vibration signal segment is: For the The final retained feature vector obtained by the adaptive gated feature selection mechanism for each vibration signal segment is: is the element-wise product.

[0094] For the front The average reward value of samples, For the Sample classification reward value, is the one-hot encoding of the true fault category label, It is the predicted probability distribution output by the pre-classification stage.

[0095] S403. Perform forward propagation of data: Adopt the residual propagation mechanism guided by spatiotemporal attention, and dynamically construct the skip connection path through the gated feature selection results.

[0096] The temporal dependency of vibration signals presents non-uniform distribution characteristics at different fault stages. Conventional fully connected neural networks have difficulty capturing the propagation path of key fault features during forward propagation.

[0097] The residual propagation mechanism guided by spatiotemporal attention in this embodiment is as follows: the spatiotemporal attention weights are calculated for the output features of the previous layer through a weight matrix, and the weights are multiplied element-by-element with a conventional linear transformation. The output features of the previous layer are subjected to nonlinear projection transformation through multiple residual paths and summed up. The attention weighted results and the residual path outputs are linearly interpolated according to the attention weights, and a skip connection path is dynamically constructed, which can be expressed as:

[0098] ,

[0099] Where, For the Layer spatiotemporal attention weights, used to dynamically construct skip connection paths, is the Sigmoid activation function, For the layer attention weight matrix, For the The final retained feature vector of each vibration signal segment is obtained after the adaptive gated feature selection mechanism.

[0100] For the The output features of the layer, is the element-wise product, is the conventional propagation weight, For the The output features of the layer, is a positive integer, M is the maximum number of residual paths, is a normalized nonlinear activation function used to normalize the output of the residual path. Let its input for , For the The projection matrix of the residual path, is the input of the normalized nonlinear activation function.

[0101] for norm, is the hyperbolic tangent function.

[0102] S404. Perform category prediction: Adopt a multi-granularity prototype comparison classification strategy, construct a learnable class prototype vector in the feature space, determine the final category through the multi-scale matching score between the gated features and the prototype vector, and solve the inter-class similarity problem between fault modes.

[0103] The multi-granularity prototype comparison classification strategy of this embodiment is as follows: for each segment feature, the exponentially normalized score of its negative L2 distance with all category prototype vectors is calculated, the scores of each segment are weighted by the gating coefficient, and the classification score of the global average feature is calculated using a fully connected layer. Finally, the prototype matching score and the traditional classification score are proportionally fused, and the category with the largest score is selected, which is expressed as:

[0104] ,

[0105] Where, For the Multi-scale prototype matching score of the class, is the total number of dynamic segments, is a positive integer, For the The gating coefficient of each segment, is a scaling factor used to adjust the sensitivity of the prototype matching score, Set to 2.0, For the The final retained feature vector obtained by the adaptive gated feature selection mechanism for each vibration signal segment is: For the The prototype vector can be trained to target the intra-class differences of air conditioner faults, such as bearing faults of different severities. The prototype vector aligns similar samples in the feature space to enhance classification robustness. for norm, is the total number of fault categories, For the class trainable prototype vector, is the category index currently being calculated, is a positive integer, It is a temporary index variable when traversing all categories. is a positive integer, is an exponential function with a natural constant as its base.

[0106] is the predicted probability distribution output from the pre-classification stage, Indicates taking the category index that maximizes the value in the brackets. is the fusion coefficient, which is used to balance the prototype matching score and the traditional classification score. Set to 0.7, is the Sigmoid activation function, is the fully connected classification weight, is the feature vector after global average pooling.

[0107] S405. Calculate the loss function: Use layered reward weighted cross entropy loss to dynamically adjust the loss weights of each category and strengthen the focus on difficult-to-classify faults.

[0108] In the fault diagnosis task of air conditioning units, different fault types have varying degrees of difficulty. Some fault types are complex and prone to misclassification, while others are relatively easy to distinguish. Traditional loss functions typically apply uniform weights to all categories, failing to effectively distinguish these differences in difficulty. This can lead to insufficient attention to difficult-to-classify faults during model training, thus affecting overall diagnostic performance.

[0109] The loss function of this embodiment is calculated as follows: the accuracy of each category on the validation set and the F1 score on the training set and validation set are calculated, and a weight factor is assigned to each category. The weight factor consists of two parts: one is a normalized value based on the category accuracy, which is used to reflect the performance of the category on the validation set; the other is an adjustment item based on the category F1 score, which is used to further strengthen the focus on difficult-to-classify faults. The weight factor of each category is multiplied by the corresponding cross-entropy loss to obtain a weighted loss value, and the weighted loss values of all categories are summed to obtain the final total loss function value, which is expressed as:

[0110] ,

[0111] In the above formula, is the weighted cross entropy loss value, is the total number of fault categories, For the Class dynamic reward weight, For the One-hot encoding of the true class label, is a logarithmic function, For the Class prediction probability value, Indicates the category index of the current calculation.

[0112] For the The accuracy of each category in the validation set, which is an independent evaluation data set divided from the training data and not involved in the training, For the The accuracy of the categories in the validation set is is the balance coefficient, for example, set it to 0.2, For the training set The F1 score of each category, For the validation set The F1 score of each category, It is a temporary index variable when traversing all categories.

[0113] S406. Perform dynamic sparse connection optimization of deep neural networks: A connection pruning strategy based on reward gradients corrects the gradient direction through cumulative rewards, enhances the ability to capture persistent failure modes, and is used to reduce model complexity.

[0114] Deep neural networks contain a large number of redundant connections, which not only increase the computational complexity of the model but can also lead to overfitting. Conventional techniques typically use random pruning or fixed-threshold pruning to reduce the number of connections. However, these methods cannot adapt to the dynamic performance of the model during training, which can lead to performance degradation after pruning.

[0115] The reward gradient connection pruning strategy of this embodiment is as follows: the gradient value of each connection is calculated and multiplied by the corresponding weight value to obtain an indicator to measure the importance of the connection. The gradient direction is corrected according to the cumulative reward to enhance the ability to capture persistent failure modes. The threshold of each connection is calculated. The threshold consists of two parts: one is based on the median of the absolute value of the gradient, which is used to reflect the importance of the connection, and the other is an adjustment term based on the recent average reward, which is used to dynamically adjust the strictness of pruning. Based on the calculated threshold and the connection importance indicator, a mask matrix is generated to set unimportant connections to zero, realizing dynamic sparse connection optimization, which is expressed as:

[0116] ,

[0117] Where, For the layer mask matrix, is a positive integer, is the indicator function, is the symbol of partial derivative, For the Tier neurons to the The connection weights of neurons, For the Dynamic pruning threshold at iterations, is the weighted cross entropy loss value. is the first adjustment parameter, Set to 0.3, is the median calculation function, is the element-wise product, For the layer weight matrix, is the second adjustment parameter, Set to 0.5, The average reward in recent time.

[0118] S407. Calculate the parameter update amount: adopt the deep neural network parameter update rule based on the reward accumulation amount.

[0119] The deep neural network parameter update rule for the reward accumulation in this embodiment is: calculate the gradient of the loss function with respect to the weight, add it to the gradient direction after the sign function processing, multiply it by the learning rate and the reward gain coefficient, and combine it with the decaying weighted accumulation of historical rewards to obtain the parameter update amount, which is expressed as:

[0120] ,

[0121] Where, For the The layer weight update amount, Update the learning rate for the weights, 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 cumulative amount of historical rewards weighted by decay, is a symbolic function.

[0122] is the current training cycle number, is the reward decay factor, Set to 0.9, For the The decay weight of the iteration, represents the weighted sum of rewards that decay over time, Indicates the The true label of the iteration, Indicates the The predicted output of the model for the iteration, For the The pre-classification confidence reward factor for the iteration.

[0123] S408. Update the parameters of the deep neural network: adopt a momentum-accelerated reward-aware update strategy.

[0124] Traditional parameter updating methods are prone to model oscillation due to data distribution drift in online scenarios.

[0125] The reward-aware update strategy for momentum acceleration in this embodiment is as follows: the historical update direction is recorded through the momentum vector. The old momentum is decayed using the momentum decay rate during each update, and the current update amount is added. The momentum vector is then applied to the weight update, which can be expressed as:

[0126] ,

[0127] Where, After the update layer weight matrix, Before the update layer weight matrices, For the The momentum vector of the layer, is the element-wise product, For the Amount of layer weight update.

[0128] is the momentum decay rate, Set to 0.9.

[0129] In the deep neural network of the present invention, the model training parameters including the gated weight matrix, the gated bias term, the attention weight matrix, the conventional propagation weight, the fully connected classification weight, and the trainable prototype vector can be 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 based on computing resource limitations.

[0130] S409. Repeat the above steps until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0131] S5, using the trained feedback-reward based deep neural network model to perform fault diagnosis and classification on the vibration data of the air-conditioning unit.

[0132] After dynamic wavelet denoising, the real-time air conditioner vibration signal is adaptively segmented using sliding window autocorrelation analysis. A pre-trained bidirectional LSTM extracts temporal features from each segment, and a dynamic gating module filters key features. The model uses a spatiotemporal attention mechanism to capture the propagation path of fault features. It calculates the degree of match between each segment's features and the stored class prototype vector in multi-scale space, then makes a weighted decision based on the global classification results of the fully connected layer, ultimately outputting the fault category with the highest confidence.

[0133] like Figure 1 As shown, this embodiment is experimentally verified as follows:

[0134] To compare the preservation of feature continuity using adaptive segmentation and traditional fixed windows, we constructed a series of data with various typical state changes to characterize the various operating fluctuations that can occur in underground well air conditioning equipment. In this experiment, the time series t is defined as 2000 uniformly sampled points from 0 to 2 seconds, and the corresponding vibration signal is composed of the following four different signal patterns:

[0135] ① The first 500 sampling points (0-0.5 seconds): stable vibration data when the equipment is operating normally;

[0136] ② The middle 500 sampling points (0.5-1.0 seconds): impulse response data when a fault occurs;

[0137] ③ The third section of 500 sampling points (1.0-1.5 seconds): mechanical looseness or periodic deviation data;

[0138] ④The last 500 sampling points (1.5-2.0 seconds): data with external noise interference.

[0139] The four spliced signals form a composite signal with typical characteristic changes, which is used to verify the responsiveness of the segmentation method at signal mutations.

[0140] In this experiment, the segment boundary is determined using a dynamic threshold strategy. The main calculation process is as follows:

[0141] ① Calculate the autocorrelation coefficient and perform autocorrelation analysis on the complete signal to obtain the symmetrical autocorrelation intensity sequence corresponding to each time point, which is used to measure the repeatability or stability of the local waveform structure;

[0142] ② Taking the global mean of the autocorrelation sequence as the benchmark, the segmentation threshold is set to 0.65 times of it to filter out points with significant structural mutations;

[0143] ③ Traverse all time points and find the positions where the autocorrelation value is higher than the threshold. These positions are used as possible paragraph boundary points. The boundaries are regarded as the turning points where the signal state changes from stability to shock or from order to disorder, representing the mutation characteristics of the signal structure.

[0144] To compare the protective effects of adaptive segmentation and traditional fixed windows on feature continuity, the purple curve in the figure is a composite signal containing steady-state vibration, transient impact, modulated waveform and random noise, the cyan curve is the change in autocorrelation coefficient, the red dotted line is the dynamic judgment threshold, and the orange vertical line marks the detected segmentation boundary. The experiment shows that in the transition area 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 modulated signal and random noise, the traditional fixed window will cause feature truncation, while the dynamic segmentation automatically adjusts the window length according to the change in local correlation, so that the feature consistency is maintained within each segment. The adaptive division based on the intrinsic characteristics of the signal solves the problem that sudden features are easily segmented in mechanical fault diagnosis.

[0145] The feature vectors extracted by the traditional method and the method of the present invention are compared through visualization after feature dimensionality reduction, and the separability of the two feature extraction methods in the feature space is demonstrated. The five typical faults in the data set are "normal", "unbalanced", "fault", "loose" and "resonance". 300 sample points are selected for each category and projected into two-dimensional space.

[0146] The traditional method uses deep neural networks based on gradient descent and error back propagation for feature extraction.

[0147] The experiment uses the t-SNE algorithm to map high-dimensional features to a two-dimensional plane and observe their distribution characteristics. The main calculation process is as follows:

[0148] ① Feature extraction was performed on two groups of 1500 two-dimensional sample data (5 categories × 300 items);

[0149] ②Use the sklearn.manifold.TSNE module for nonlinear dimensionality reduction, setting the parameters n_components=2 and perplexity=30;

[0150] ③ Perform dimensionality reduction mapping on the feature data of the traditional method and the method of the present invention respectively, compressing them into two-dimensional space for intuitive visualization;

[0151] ④ Use different colors to render the coordinate values of the two-dimensional data points, and mark each type of fault with color to reflect the distribution pattern and classification potential.

[0152] In the feature distribution after feature extraction by traditional methods, the distribution boundary of each type of samples is not obvious, the mean is 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.

[0153] The method of the present invention has a compact distribution of samples in each category of feature data, a far distance between means of each category, and a small covariance, indicating that after the present invention utilizes the prototype contrast learning mechanism and gated feature selection, the extracted features have strong intra-class compactness and inter-class separability.

[0154] like Figure 2 and Figure 3 As shown in Figure 2, the classification potential of different methods is compared by visualizing the feature space distribution, and the manifold learning algorithm is used to project high-dimensional features onto a two-dimensional plane. Figure 2 It shows that the features extracted by traditional methods are distributed in an overlapping manner, especially the bearing fault and looseness anomaly categories have a large number of overlapping areas. Figure 3 The features extracted by the method of the present invention show a clear star-shaped radial distribution. The five categories each form a tight cluster with clear boundaries. The colors of the points in the figure represent different fault types, and the distribution shape reflects the model's ability to distinguish. This proves that the multi-granularity prototype comparison strategy establishes a more discriminative metric relationship in the feature space by constructing a learnable category prototype vector, thereby significantly improving the inter-class separation.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A distributed underground well data processing method, characterized by: The following steps are involved: S1, collects vibration data of the air conditioning units in the airport’s distributed wells through acceleration sensors; S2, based on the wavelet decomposition layer number and threshold strategy with dynamic adjustment of signal-to-noise ratio, combined with the time segment clustering algorithm, denoises the collected vibration data of the air-conditioning unit; The wavelet decomposition layer number and threshold value strategy for dynamic adjustment of the signal-to-noise ratio in S2 are as follows: perform wavelet decomposition on the original vibration signal, dynamically adjust the decomposition layer number and threshold value according to the signal-to-noise ratio of the signal, calculate the signal-to-noise ratio estimate for each layer of wavelet coefficients, perform threshold processing on the wavelet coefficients using a dynamic threshold function, and reconstruct the processed wavelet coefficients to obtain the denoised vibration signal, which can be 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, For the Layer wavelet coefficients, is the original vibration signal, is the dynamic threshold function, For the Layer wavelet coefficient signal-to-noise ratio estimation; For rounding operation, is a logarithmic function, is the standard deviation function, It is to find the standard deviation of the original vibration signal; for norm; is an exponential function with a natural constant as its base, is the slope parameter, is the signal-to-noise ratio threshold, =0.1; S3, adopts the sliding window dynamic partitioning strategy, combines the autocorrelation coefficient within the window to determine the segmentation boundary, and performs segmentation feature encoding on the vibration data of the air-conditioning unit; S4, a deep neural network model based on feedback rewards is trained and optimized through a feedback reward mechanism to improve the performance of the model in complex fault diagnosis tasks; S5, using the trained feedback-reward based deep neural network model to perform fault diagnosis and classification on the vibration data of the air-conditioning unit.

2. The distributed well data processing method according to claim 1, characterized in that: The dynamic partitioning strategy of the sliding window in S3 is as follows: the denoised vibration signal is scanned using a sliding window, and the local autocorrelation coefficient of each window position is calculated. When the local autocorrelation coefficient is greater than 1.5 times the global average autocorrelation coefficient, the position is used as the segmentation boundary to divide the signal into multiple segments. The signal in each segment is feature encoded using a preset bidirectional LSTM, and the forward and reverse LSTM outputs are spliced to obtain the feature vector of each segment, which is expressed as: ， In the above formula, is the segmented vibration signal set, is the total number of dynamic segments, is a positive integer, For the The starting time point of each segment, For the The end time point of each segment, For arrive The denoised vibration signal fragment, For location The local autocorrelation coefficient of is the boundary determination threshold, is the global average autocorrelation coefficient; The window radius for calculating the local autocorrelation coefficient, is a positive integer, For location Left side The denoised signal value of the sampling point, For location Right side The denoised signal value of each sampling point; is the total duration of the vibration signal, is a positive integer, For the The local autocorrelation coefficient at each time point.

3. The distributed well data processing method according to claim 1, characterized in that: 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 the pre-classification confidence, calculate the gradient expectation and reward product in the pre-training phase, so that the initial weights are biased towards the direction of high-confidence features; S402, perform adaptive gated feature selection: Adopt 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; S403. Perform forward propagation of data: Use a residual propagation mechanism guided by spatiotemporal attention to dynamically construct skip connection paths based on gated feature selection results. S404. Perform category prediction: Use a multi-granularity prototype comparison classification strategy to construct a learnable class prototype vector in the feature space, and determine the final category through the multi-scale matching score between the gated features and the prototype vector; S405. Calculate the loss function: Use layered reward-weighted cross entropy loss to dynamically adjust the loss weights of each category, and strengthen the focus on difficult-to-classify faults; S406. Perform dynamic sparse connection optimization of deep neural networks: A connection pruning strategy based on reward gradients modifies the gradient direction through cumulative rewards to enhance the ability to capture persistent failure modes. S407, calculating the parameter update amount: using a 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 that the model training is completed.

4. The distributed well data processing method according to claim 3, characterized in that: The weight initialization strategy for the pre-classification confidence in S401 is as follows: in the pre-training stage, for each layer of the deep neural network, the gradient of its pre-classification cross entropy loss with respect to the weight is calculated, the gradient is multiplied by the pre-classification confidence reward factor to obtain the gradient expectation, the gradient expectation is multiplied by the initial learning rate, and then added to the initial weight of the He normal distribution of the layer to obtain the initial weight of the layer.

5. The distributed well data processing method according to claim 3, characterized in that: The method for performing adaptive gated feature selection in S402 is as follows: for each segmented feature vector, the activation value of its dynamic feature selection gate is calculated, the activation value is mapped to the interval [0, 1] by the Sigmoid activation function, and the output of the feature selection gate is obtained. The output is used to weight the feature vector element by element, retaining important features, and at the same time, the unselected features are weightedly fused with the selected features at the previous moment to obtain the final feature selection result.

6. The distributed well data processing method according to claim 3, characterized in that: The residual propagation mechanism guided by spatiotemporal attention in S403 is as follows: the spatiotemporal attention weights are calculated for the output features of the previous layer through a weight matrix, which is multiplied element-by-element with a conventional linear transformation, and the output features of the previous layer are nonlinearly projected and summed through multiple residual paths. The attention weighted results and the residual path outputs are linearly interpolated according to the attention weights, and a jump connection path is dynamically constructed.

7. The distributed well data processing method according to claim 3, characterized in that: The multi-granularity prototype comparison classification strategy in S404 is as follows: for each segment feature, calculate the exponential normalized score of its negative L2 distance with all category prototype vectors, weight the scores of each segment by the gating coefficient, use the fully connected layer to calculate the classification score of the global average feature, and finally proportionally fuse the prototype matching score with the traditional classification score, and select the category with the largest score, which is expressed as: ， Where, For the Multi-scale prototype matching score of the class, is the total number of dynamic segments, is a positive integer, For the The gating coefficient of each segment, is the scale factor, For the The final retained feature vector obtained by the adaptive gated feature selection mechanism for each vibration signal segment is: For the class trainable prototype vector, for norm, is the total number of fault categories, For the class trainable prototype vector, is the category index currently being calculated, It is a temporary index variable when traversing all categories. is an exponential function with a natural constant as its base; is the predicted probability distribution output from the pre-classification stage, Indicates taking the category index that maximizes the value in 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.

8. The distributed well data processing method according to claim 3, characterized in that: The calculation method of the loss function in S405 is as follows: calculate the accuracy of each category on the validation set, and the F1 score 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 a weighted loss value, and sum the weighted loss values of all categories to obtain the final total loss function value.

9. The distributed underground well data processing method according to claim 3, characterized in that: The connection pruning strategy of the reward gradient in S406 is as follows: the gradient value of each connection is calculated and multiplied by the corresponding weight value to obtain an indicator to measure the importance of the connection. The gradient direction is corrected according to the cumulative reward to enhance the ability to capture persistent failure modes. The threshold of each connection is calculated. Based on the calculated threshold and the connection importance indicator, a mask matrix is generated to set unimportant connections to zero, thereby achieving dynamic sparse connection optimization, which is expressed as: ， Where, For the layer mask matrix, is a positive integer, is the indicator function, is the symbol of partial derivative, is the weighted cross entropy loss value, For the Tier neurons to the The connection weights of neurons, For the Dynamic pruning threshold at iterations; is the first adjustment parameter, is the median calculation function, is the element-wise product, For the layer weight matrix, is the second adjustment parameter, is the recent average reward; The deep neural network parameter update rule of the reward accumulation amount in S407 is as follows: calculate the gradient of the loss function with respect to the weight, add it to the gradient direction after the sign function processing, multiply it by the learning rate and the reward gain coefficient, and combine it with the decay weighted accumulation of historical rewards to obtain the parameter update amount; The reward perception update strategy of momentum acceleration in S408 is as follows: the historical update direction is recorded through the momentum vector, the old momentum is decayed by the momentum decay rate at each update, and the current update amount is superimposed, and the momentum vector is applied to the weight update.

Images