Operation data management system and method based on artificial intelligence

Abstract

The invention discloses an operation data management system and method based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the following steps: 1, collecting multi-source data in real time through an intelligent sensor and industrial equipment, achieving time synchronization, and ensuring data alignment; step 2, cleaning, labeling and standardizing the collected data, extracting feature vectors of the data, and generating multi-modal feature vectors through dynamic weighted fusion of an attention mechanism; step 3, calculating a reconstruction error based on an auto-encoder to identify an abnormal state, quantifying the contribution degree of each feature to the abnormity in combination with an SHAP algorithm, and positioning key influence factors; and step 4, dynamically adjusting the detection threshold by adopting reinforcement learning to reduce the false alarm and missing alarm rate, and evaluating the prediction confidence by quantifying the model error and the data error to realize the dynamic adjustment of the early warning priority. According to the invention, the condition that abnormal data and false alarm of the abnormal data are difficult to effectively warn in the prior art can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and specifically to an operation data management system and method based on artificial intelligence. Background Art

[0002] In the industrial field, with the increasing complexity and intelligence of equipment, operation data management has become crucial. The rapid development of intelligent sensor technology has made it possible to collect multi-source data such as real-time and accurate equipment operation status, environmental parameters, and load status. At the same time, artificial intelligence algorithms such as deep learning and reinforcement learning have shown powerful capabilities in data processing and analysis.

[0003] However, there are still many deficiencies in existing operation data management solutions. On the one hand, the data collection sources are relatively single, resulting in difficult accurate correlation during data analysis and reducing the reliability of analysis results. On the other hand, when traditional data preprocessing methods handle complex data, the data standardization method is easily affected by outliers, affecting the accuracy of subsequent models. In addition, when existing technologies make wrong predictions and judgments on abnormal data, it is difficult to provide timely, accurate, and clearly prioritized warning information for operation and maintenance personnel, seriously affecting equipment maintenance efficiency and production stability. Summary of the Invention

[0004] The purpose of the present invention is to provide an operation data management system and method based on artificial intelligence to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An operation data management method based on artificial intelligence, the method includes the following steps:

[0006] Step 1: Real-time collect multi-source data through intelligent sensors and industrial equipment, achieve time synchronization, and ensure data alignment;

[0007] Step 2: Clean, label, and standardize the collected data, extract each data feature vector, and generate a multi-modal feature vector through dynamic weighted fusion by the attention mechanism;

[0008] Step 3: Identify abnormal states based on the reconstruction error calculated by the autoencoder, combine the SHAP algorithm to quantify the contribution of each feature to the abnormality, and locate the key influencing factors;

[0009] Step 4: Dynamically adjust the detection threshold by reinforcement learning to reduce the false alarm and missed alarm rates, evaluate the prediction confidence by quantifying the model error and data error, and achieve dynamic adjustment of the warning priority;

[0010] A method for managing operation data based on artificial intelligence according to claim 1, characterized in that: in step 1, during the operation of the system, intelligent sensors and data acquisition devices are installed to record operation status data at fixed time intervals:

[0011] Device operation status acquisition: expressed as [R1, R2,..., Ru]; where u is a positive integer representing the number of types of device operation status, and R1 to Ru respectively represent the 1st to u-th types of device operation status;

[0012] Sources of device operation status data: temperature sensors, pressure sensors, current voltage meters, etc.; the collected data such as temperature, power, rotational speed, current, etc.;

[0013] Environmental parameter acquisition: expressed as [E1, E2,..., Ev]; where v is a positive integer representing the number of types of environmental parameters, and E1 to Ev respectively represent the 1st to v-th types of environmental parameters;

[0014] Environmental parameters are used to monitor the impact of the external environment on device operation; data sources: environmental sensors; the collected data such as humidity, light, atmospheric pressure;

[0015] Load status acquisition: expressed as [L1, L2,..., Lw]; where w is a positive integer representing the number of types of load status, and L1 to Lw respectively represent the 1st to w-th types of load status;

[0016] Sources of load status data: task management system, controller data;

[0017] Visual information acquisition: expressed as [V1, V2,..., Vz]; where z is a positive integer representing the number of types of device appearances, and V1 to Vz respectively represent the 1st to z-th types of device appearances;

[0018] Sources of visual information data: industrial cameras; the collected data such as cracks, damages, dust accumulation, etc.;

[0019] Audio information acquisition: expressed as [A1, A2,..., Ap]; where p is a positive integer representing the number of types of device audio, and A1 to Ap respectively represent the 1st to p-th types of device audio;

[0020] Sources of audio information data: industrial microphones; the collected data such as abnormal noises, vibration noises, etc.;

[0021] Set data acquisition time points, expressed as [T1, T2,..., Tm]; where m is a positive integer representing the number of device data acquisition time points, and T1 to Tm respectively represent the 1st to m-th data acquisition time points; the time intervals between adjacent moments are equal;

[0022] Time synchronization: Use the NTP protocol to align the data in time to ensure that multi-modal data is analyzed under the same timestamp.

[0023] In step 2, preprocess the data:

[0024] Data annotation: Manually annotate the anomalies in the data;

[0025] Data cleaning: Remove duplicate data and fill in missing values with the mean;

[0026] Invalid data elimination: Set the range of each parameter and filter invalid data based on physical rules; when sensor data conflicts, filter invalid data based on state consistency detection; remove invalid data caused by equipment failures or transmissions;

[0027] Data standardization: Standardize the data using median-MAD;

[0028] For numerical data (equipment operating status, environmental parameters, load status): Build an LSTM time series model: Take multi-dimensional time series numerical data as input and use the long short-term memory network (LSTM) to model the data; LSTM captures long dependencies in the time series through its memory units, learns rules and dynamic change patterns; perform feature extraction through a multi-layer LSTM structure, and the hidden state captures the temporal changes of the equipment state; output a numerical feature vector and predicted values for future moments; the numerical feature vector is represented as X1;

[0029] For visual information data (equipment appearance): Take visual information data as input, extract image features through CNN (Convolutional Neural Network), and output a visual information feature vector; the visual information feature vector is represented as X2;

[0030] For audio information data (equipment sound): Use the short-time Fourier transform to convert the audio into a time-frequency graph, calculate the MFCC Mel-frequency cepstral coefficients, and extract audio features; Transformer performs temporal modeling: Learn long-term dependencies based on the self-attention mechanism to identify abnormal sound patterns; output an audio data feature vector; the audio data feature vector is represented as X3;

[0031] Adopt an attention module trained based on historical data to learn the importance of each feature vector through one or more neural network layers: Non-linearly combine each feature vector with a learned query vector, and then output an attention score after passing through an activation function; Use the softmax operation to convert the original attention score into a normalized weight to ensure that the sum of all weights is 1;

[0032] Train the attention module using historical data and corresponding anomaly detection results to maximize the detection accuracy; update the parameters of the attention module through backpropagation and gradient descent algorithms; input the latest feature vectors, and calculate the real-time weights α1, α2, and α3 according to the characteristics of the current data; perform weighted summation to obtain the fused feature: X = α1·X1 + α2·X2 + α3·X3;

[0033] In step 3, use the multi-modal fused feature X as the input; train the autoencoder using a large amount of historical data; the encoder compresses the input data into a low-dimensional feature representation, and then the decoder reconstructs the original input; use the mean square error (MSE) as the loss function to minimize the reconstruction error;

[0034] Calculate the Euclidean distance between the current input and the reconstructed output as the reconstruction error ε: set the threshold θ; when the real-time calculated reconstruction error ε exceeds the set threshold θ, determine that the current data is abnormal and record it;

[0035] For the samples determined to be abnormal and their corresponding multi-modal fused features X, use the SHAP algorithm (such as KernelSHAP or DeepSHAP) to calculate the SHAP values of each input feature Compare the magnitudes of the SHAP values of each feature and select the top N factors that have the most influence on anomaly detection.

[0036] In step 4, false positive optimization analysis: Definition of false positive rate F1: F1 = number of false positives / total number of actually normal samples; where, "number of false positives" refers to the number of samples that are wrongly determined to be abnormal when the system is in a normal state; Definition of false negative rate F2: F2 = number of false negatives / total number of actually abnormal samples; where, "number of false negatives" refers to the number of samples that actually have anomalies but are wrongly determined to be normal;

[0037] Construct a reinforcement learning intelligent agent to adjust the anomaly detection threshold θ according to the current state information in each decision-making cycle; the state information includes the false positive rate F1, false negative rate F2, and model reconstruction error distribution within the current time period;

[0038] Design the reward function F = -λ1·F1 - λ2·F2; where, λ1 and λ2 are weight parameters set according to business requirements; maximize the reward function to enable the intelligent agent to automatically select the optimal threshold adjustment strategy;

[0039] Use historical data to pre-train the intelligent agent to master the preliminary threshold adjustment strategy; monitor the false positive and false negative situations in real time, and the intelligent agent makes adjustments based on the current state and conducts online learning and optimization through continuous feedback (including false positive feedback detected automatically by the system and manually labeled);

[0040] Feedback means that the operation and maintenance personnel manually annotate the alarm results (judged as real anomalies or false alarms), and the annotation information will be input into the reinforcement learning module as an additional supervision signal to help the agent adapt to the actual application scenario faster;

[0041] Error source quantification: Model prediction error σ1: In the training stage of the autoencoder, through the statistics of historical reconstruction errors, the inherent mean square error of the model is obtained as an uncertainty index; Input data error σ2: Combining the accuracy of the sensor and the errors generated by mean filling and standardization in the data preprocessing process, the overall uncertainty of the input data is calculated;

[0042] Total uncertainty calculation: σ = (σ1 2 + σ2 2 ) 1 / 2 ;

[0043] Set a threshold, and adjust the confidence of the current anomaly detection result according to the σ value: When σ is higher than the threshold, it indicates that the uncertainty of the current prediction result is relatively large, and the early warning is postponed or the early warning priority is reduced; When σ is not higher than the threshold, the alarm is triggered immediately or the early warning priority is increased.

[0044] An operation data management system based on artificial intelligence, which includes a data acquisition module, a preprocessing module, an anomaly detection module and an optimization module;

[0045] The data acquisition module is used to collect multi-source data in real time through intelligent sensors and industrial equipment, achieve time synchronization, and ensure data alignment;

[0046] The preprocessing module is used to clean, annotate and standardize the collected data, extract the feature vectors of each data, and generate multi-modal feature vectors through dynamic weighted fusion by the attention mechanism;

[0047] The anomaly detection module is used to identify the abnormal state based on the reconstruction error calculated by the autoencoder, quantify the contribution of each feature to the anomaly by combining the SHAP algorithm, and locate the key influencing factors;

[0048] The optimization module is used to dynamically adjust the detection threshold by reinforcement learning to reduce the false alarm and missed alarm rates, evaluate the prediction confidence by quantifying the model error and data error, and realize the dynamic adjustment of the early warning priority.

[0049] The data acquisition module includes a data acquisition unit and a time synchronization unit;

[0050] The data acquisition unit is used to collect multi-source data such as the operating state of the device, environmental parameters, load state visual information and audio information in real time;

[0051] The time synchronization unit is used to align the timestamps of multi-source data using the NTP protocol to ensure the temporal consistency of multi-modal data.

[0052] The preprocessing module includes a data cleaning unit, a feature extraction unit, and a fusion weighting unit;

[0053] The data cleaning unit is used to eliminate duplicate and invalid data, filter outliers based on physical rules, and fill in missing data with the mean;

[0054] The feature extraction unit is used to process numerical, image, and audio data in parallel, and extract feature vectors using LSTM, CNN, and Transformer respectively;

[0055] The fusion weighting unit is used to dynamically calculate feature weights through an attention mechanism and generate a weighted fusion multi-modal feature vector.

[0056] The anomaly detection module includes an autoencoder unit, a threshold determination unit, and a factor localization unit;

[0057] The autoencoder unit is used to learn the normal mode through an encoding-decoding structure and calculate the reconstruction error of the input data;

[0058] The threshold determination unit is used to trigger an anomaly alarm judgment by comparing the real-time reconstruction error with the dynamic threshold;

[0059] The factor localization unit is used to analyze the feature contribution degree using the SHAP algorithm and locate the key parameters that cause anomalies.

[0060] The optimization module includes a reinforcement learning unit, an error quantification unit, and a confidence adjustment unit;

[0061] The reinforcement learning unit is used to construct a reward function based on the false alarm / miss rate and dynamically optimize the anomaly determination threshold;

[0062] The error quantification unit is used to calculate the model prediction error and the data acquisition error, and evaluate the total uncertainty index;

[0063] The confidence adjustment unit is used to dynamically adjust the early warning priority according to the uncertainty index and achieve credibility-level alarm.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a reinforcement learning intelligent agent, the present invention dynamically adjusts the anomaly detection threshold according to the current false alarm rate, missed alarm rate, model reconstruction error distribution and other state information. By designing a reasonable reward function and maximizing the reward, the intelligent agent can automatically learn the optimal threshold adjustment strategy, effectively adapt to the dynamic changes of the device operation state, significantly reduce the false alarm rate and missed alarm rate, and improve the practicability and stability of the anomaly detection system; The present invention quantifies the error to evaluate the prediction confidence: evaluates the prediction confidence by quantifying the model error and data error, and dynamically adjusts the warning priority accordingly. When the total uncertainty is higher than the threshold, the warning is postponed or the warning priority is reduced to avoid false warnings caused by high uncertainty; When the total uncertainty is not higher than the threshold, the alarm is triggered immediately or the warning priority is increased. This way makes the warning information more targeted and reliable, helps the operation and maintenance personnel to reasonably arrange the maintenance work, and improves the utilization efficiency of operation and maintenance resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of the steps of a method for managing operation data based on artificial intelligence according to the present invention;

[0066] Figure 2 It is a schematic diagram of the process of a system for managing operation data based on artificial intelligence according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] Embodiment: As Figure 1 - Figure 2 shown, the present invention provides a technical solution, a method for managing operation data based on artificial intelligence, and the method includes the following steps:

[0069] Step 1: Real-time collect multi-source data through intelligent sensors and industrial devices, achieve time synchronization, and ensure data alignment;

[0070] Step 2: Clean, label and standardize the collected data, extract the feature vectors of each data, and generate a multi-modal feature vector through dynamic weighted fusion by the attention mechanism;

[0071] Step 3: Identify the abnormal state based on the reconstruction error calculated by the autoencoder, combine the SHAP algorithm to quantify the contribution of each feature to the abnormality, and locate the key influencing factors;

[0072] Step 4: Dynamically adjust the detection threshold using reinforcement learning to reduce the false alarm and missed detection rates, evaluate the prediction confidence by quantifying the model error and data error, and achieve dynamic adjustment of the early warning priority;

[0073] A method for managing operation data based on artificial intelligence according to claim 1, characterized in that: in step 1, during the operation of the system, intelligent sensors and data acquisition devices are installed to record the operation status data at fixed time intervals:

[0074] Device operation status acquisition: expressed as [R1, R2,..., Ru]; where u is a positive integer representing the number of types of device operation status, and R1 to Ru respectively represent the 1st to u-th types of device operation status;

[0075] Sources of device operation status data: temperature sensors, pressure sensors, current and voltage meters, etc.; the collected data such as temperature, power, rotation speed, current, etc.;

[0076] Environmental parameter acquisition: expressed as [E1, E2,..., Ev]; where v is a positive integer representing the number of types of environmental parameters, and E1 to Ev respectively represent the 1st to v-th types of environmental parameters;

[0077] Environmental parameters are used to monitor the impact of the external environment on the device operation; data sources: environmental sensors; the collected data such as humidity, light, atmospheric pressure;

[0078] Load status acquisition: expressed as [L1, L2,..., Lw]; where w is a positive integer representing the number of types of load status, and L1 to Lw respectively represent the 1st to w-th types of load status;

[0079] Sources of load status data: task management system, controller data;

[0080] Visual information acquisition: expressed as [V1, V2,..., Vz]; where z is a positive integer representing the number of types of device appearances, and V1 to Vz respectively represent the 1st to z-th types of device appearances;

[0081] Sources of visual information data: industrial cameras; the collected data such as cracks, damages, dust accumulation, etc.;

[0082] Audio information acquisition: expressed as [A1, A2,..., Ap]; where p is a positive integer representing the number of types of device audio, and A1 to Ap respectively represent the 1st to p-th types of device audio;

[0083] Sources of audio information data: industrial microphones; the collected data such as abnormal noises, vibration noises, etc.;

[0084] Set the data collection time points, denoted as [T1, T2,..., Tm]; where m is a positive integer representing the number of device data collection time points, and T1 to Tm respectively represent the 1st to mth data collection time points; the time intervals between adjacent moments are equal;

[0085] Time synchronization: Use the NTP protocol to align the time of the data to ensure that multi-modal data is analyzed under the same time stamp.

[0086] In step 2, preprocess the data:

[0087] Data annotation: Manually annotate the anomalies existing in the data;

[0088] Data cleaning: Remove duplicate data and fill in missing values with the mean;

[0089] Invalid data elimination: Set the ranges of each parameter, and filter invalid data based on physical rules; when there are conflicts in sensor data, filter invalid data based on state consistency detection; remove invalid data caused by device failures or transmissions;

[0090] Data standardization: Standardize the data using median-MAD;

[0091] For numerical data (device operating status, environmental parameters, load status): Construct an LSTM time series model: Take multi-dimensional time series numerical data as input, and use the long short-term memory network (LSTM) to model the data; LSTM captures long-term dependencies in the time series through its memory units, learns the rules and dynamic change patterns; perform feature extraction through a multi-layer LSTM structure, and the hidden state captures the temporal changes of the device state; output a numerical feature vector and the predicted value at the future moment; the numerical feature vector is denoted as X1;

[0092] For visual information data (device appearance): Take visual information data as input, extract image features through a CNN (convolutional neural network), and output a visual information feature vector; the visual information feature vector is denoted as X2;

[0093] For audio information data (device sound): Use the short-time Fourier transform to convert the audio into a time-frequency diagram, calculate the MFCC mel-frequency cepstral coefficients, and extract audio features; Transformer performs temporal modeling: Learn long-term dependencies based on the self-attention mechanism to identify abnormal sound patterns; output an audio data feature vector; the audio data feature vector is denoted as X3;

[0094] An attention module trained based on historical data is adopted to learn the importance of each feature vector through one or more neural network layers: non-linearly combine each feature vector with a learned query vector, and then output an attention score through an activation function; use the softmax operation to convert the original attention score into a normalized weight to ensure that the sum of all weights is 1;

[0095] Use historical data and corresponding anomaly detection results to train the attention module to maximize the detection accuracy; update the parameters of the attention module through backpropagation and gradient descent algorithms; input the latest feature vectors of each, and calculate real-time weights α1, α2, and α3 according to the characteristics of the current data; obtain the fused feature by weighted summation: X = α1·X1 + α2·X2 + α3·X3;

[0096] In step 3, use the multi-modal fused feature X as the input; train an autoencoder with a large amount of historical data; the encoder compresses the input data into a low-dimensional feature representation, and then the decoder reconstructs the original input; use the mean square error (MSE) as the loss function to minimize the reconstruction error;

[0097] Calculate the Euclidean distance between the current input and the reconstructed output as the reconstruction error ε: set a threshold θ; when the real-time calculated reconstruction error ε exceeds the set threshold θ, determine that the current data is abnormal and record it;

[0098] For the samples determined to be abnormal and their corresponding multi-modal fused features X, use the SHAP algorithm (such as KernelSHAP or DeepSHAP) to calculate the SHAP value of each input feature Compare the magnitudes of the SHAP values of each feature and screen out the top N factors that have the most influence on anomaly detection.

[0099] In step 4, misjudgment optimization analysis: Definition of false alarm rate F1: F1 = number of false alarms / total number of actually normal samples; where, "number of false alarms" refers to the number of samples that are wrongly determined to be abnormal when the system is in a normal state; Definition of miss rate F2: F2 = number of missed alarms / total number of actually abnormal samples; where, "number of missed alarms" refers to the number of samples that actually have anomalies but are wrongly determined to be normal;

[0100] Construct a reinforcement learning intelligent agent to adjust the anomaly detection threshold θ according to the current state information in each decision cycle; the state information includes the false alarm rate F1, miss rate F2, and model reconstruction error distribution within the current time period;

[0101] Design a reward function F = -λ1·F1 - λ2·F2; where, λ1 and λ2 are weight parameters set according to business requirements; maximize the reward function to enable the intelligent agent to automatically select the optimal threshold adjustment strategy;

[0102] Pre-train the intelligent agent using historical data to master the initial threshold adjustment strategy; monitor false alarms and missed alarms in real time, and the intelligent agent makes adjustments based on the current state and conducts online learning and optimization through continuous feedback (including false alarm feedback detected automatically by the system and manually labeled);

[0103] Feedback means that the operation and maintenance personnel manually label the alarm results (judged as real anomalies or false alarms), and the labeled information will be input into the reinforcement learning module as an additional supervision signal to help the agent adapt to the actual application scenario faster;

[0104] Quantification of error sources: Model prediction error σ1: In the training stage of the autoencoder, obtain the inherent mean square error of the model as an uncertainty index through the statistics of historical reconstruction errors; Input data error σ2: Combine the accuracy of the sensor and the errors generated by mean filling and standardization in the data preprocessing process to calculate the overall uncertainty of the input data;

[0105] Total uncertainty calculation: σ=(σ1 2 +σ2 2 ) 1 / 2 ;

[0106] Set the threshold, and adjust the confidence level of the current anomaly detection result according to the σ value: When σ is higher than the threshold, it indicates that the uncertainty of the current prediction result is relatively large, and the early warning is postponed or the early warning priority is reduced; When σ is not higher than the threshold, the alarm is triggered immediately or the early warning priority is increased.

[0107] An operation data management system based on artificial intelligence, which includes a data acquisition module, a preprocessing module, an anomaly detection module, and an optimization module;

[0108] The data acquisition module is used to collect multi-source data in real time through intelligent sensors and industrial equipment, achieve time synchronization, and ensure data alignment;

[0109] The preprocessing module is used to clean, label, and standardize the collected data, extract the feature vectors of each data, and generate multi-modal feature vectors through dynamic weighted fusion by the attention mechanism;

[0110] The anomaly detection module is used to identify the abnormal state based on the reconstruction error calculated by the autoencoder, quantify the contribution of each feature to the anomaly in combination with the SHAP algorithm, and locate the key influencing factors;

[0111] The optimization module is used to dynamically adjust the detection threshold by reinforcement learning to reduce the false alarm and missed alarm rates, evaluate the prediction confidence by quantifying the model error and data error, and realize the dynamic adjustment of the early warning priority.

[0112] The data acquisition module includes a data acquisition unit and a time synchronization unit;

[0113] The data acquisition unit is used to collect multi-source data such as the running state of the device, environmental parameters, visual information of the load state, and audio information in real time;

[0114] The time synchronization unit is used to align the timestamps of multi-source data using the NTP protocol to ensure the temporal consistency of multi-modal data.

[0115] The preprocessing module includes a data cleaning unit, a feature extraction unit, and a fusion weighting unit;

[0116] The data cleaning unit is used to eliminate duplicate and invalid data, filter outliers based on physical rules, and fill in missing data with the mean;

[0117] The feature extraction unit is used to process numerical, image, and audio data in parallel, and extract feature vectors using LSTM, CNN, and Transformer respectively;

[0118] The fusion weighting unit is used to dynamically calculate feature weights through an attention mechanism to generate a weighted fusion multi-modal feature vector.

[0119] The anomaly detection module includes an auto-encoding unit, a threshold determination unit, and a factor localization unit;

[0120] The auto-encoding unit is used to learn the normal mode through an encoding-decoding structure and calculate the reconstruction error of the input data;

[0121] The threshold determination unit is used to trigger an anomaly alarm judgment by comparing the real-time reconstruction error with the dynamic threshold;

[0122] The factor localization unit is used to analyze the feature contribution degree using the SHAP algorithm and locate the key parameters that cause anomalies.

[0123] The optimization module includes a reinforcement learning unit, an error quantification unit, and a confidence adjustment unit;

[0124] The reinforcement learning unit is used to construct a reward function based on the false alarm / missed alarm rate and dynamically optimize the anomaly determination threshold;

[0125] The error quantification unit is used to calculate the model prediction error and the data acquisition error and evaluate the total uncertainty index;

[0126] The confidence adjustment unit is used to dynamically adjust the early warning priority according to the uncertainty index to achieve credibility-level alarm.

[0127] In this embodiment, a wind farm deploys 100 wind turbines, and it is necessary to monitor the device status in real time to prevent failures. The system realizes multi-modal data management based on the above AI method, and takes vibration anomaly as an example to illustrate the implementation process.

[0128] Data collection (Step 1): Equipment operating status: temperature sensor (R1), vibration sensor (R2), current and voltage meter (R3); environmental parameters: wind speed sensor (E1), humidity sensor (E2); load status: power generation (L1), pitch angle (L2); visual information: tower / blade camera (V1); audio information: nacelle microphone (A1).

[0129] Time synchronization: Collect data every 10 seconds, align timestamps using the NTP protocol to ensure consistent timestamps for multi-modal data.

[0130] Data preprocessing and feature fusion (Step 2):

[0131] Data annotation: Manually annotate vibration anomaly samples (such as bearing faults) in historical data.

[0132] Invalid data elimination: Records with vibration values exceeding [-10g, 10g] (physical rule); when there is a conflict between temperature sensor and vibration sensor data (such as high temperature but normal vibration), mark as invalid.

[0133] Missing value filling: Fill missing current data with the mean of the first 3 valid values.

[0134] Normalization: Use median-MAD normalization for numerical data (such as temperature, vibration).

[0135] Feature extraction:

[0136] Numerical data (LSTM modeling): Input: Temperature, vibration, current, wind speed, power generation in the past 1 hour (at 10-second intervals, a total of 360 time steps); Output: Hidden layer state vector X1 (128 dimensions).

[0137] Visual data (CNN): Input: Blade image with a resolution of 224×224; Output: Feature vector X2 (2048 dimensions) extracted by ResNet50.

[0138] Audio data (MFCC + Transformer): Input: 1-second audio segment (sampling rate 44.1kHz); Processing: Short-time Fourier transform → MFCC features (40 dimensions) → Transformer time series modeling; Output: Feature vector X3 (128 dimensions).

[0139] Attention mechanism fusion:

[0140] Training stage: Use historical data to train the attention module, with the goal of maximizing the anomaly detection accuracy; learn weights α1 (numerical), α2 (visual), α3 (audio).

[0141] Real-time calculation: At the current moment, each eigenvector: X1, X2, X3; Attention score: q = W·[X1; X2; X3] + b (W is the weight matrix, b is the bias); Normalized weight: α1 = 0.6,

[0142] α2 = 0.3, α3 = 0.1; Fusion feature: X = 0.6·X1 + 0.3·X2 + 0.1·X3;

[0143] Anomaly detection and localization (Step 3):

[0144] Autoencoder training: Input: The fusion feature X of historical normal operations (a total of 100,000 pieces); Structure: Encoder (3 fully connected layers, compressed to 64 dimensions) → Decoder (reconstruct the original X); Loss function: MSE, trained until convergence.

[0145] Real-time detection: Input at the current moment: X = [vibration feature, temperature feature,...] (the fused vector); Reconstruction error: ε = ||X - reconstructed X'||2; Threshold θ: Set based on the 99th percentile of the reconstruction error of historical normal data (e.g., θ = 0.8); Anomaly determination: When ε = 1.2 > θ, trigger an anomaly alarm.

[0146] Key factor localization (SHAP analysis): Input: The fusion feature X of the abnormal sample; Output: SHAP value of the vibration feature φ = 0.7, SHAP value of the current feature φ = 0.2, SHAP values of other features φ < 0.1; Conclusion: Vibration anomaly is the main cause, and the bearing status needs to be checked.

[0147] Dynamic threshold adjustment and confidence evaluation (Step 4):

[0148] Reinforcement learning to adjust the threshold: State information: False alarm rate F1 = 0.5% in the past 24 hours, miss rate F2 = 2%, mean reconstruction error = 0.6; Reward function: F = -0.5·F1 - 0.5·F2 → Current reward F = -1.25; Policy optimization: The intelligent agent increases the threshold θ from 0.8 to 0.9 to reduce false alarms.

[0149] Confidence evaluation: Model error σ1: Standard deviation of historical reconstruction error = 0.3; Data error σ2: Vibration sensor accuracy ±0.1g, preprocessing error = 0.05g → σ2 = √(0.1² + 0.05²) = 0.11;

[0150] Total uncertainty σ = √(0.3 2 + 0.11²) = 0.32; Threshold setting: σ threshold = 0.3. Current σ = 0.32 > 0.3 → Reduce the early warning priority and trigger a yellow alarm.

[0151] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. An operation data management method based on artificial intelligence, characterized in that: The method includes the following steps: Step 1: Collect multi-source data in real time through intelligent sensors and industrial devices, achieve time synchronization, and ensure data alignment; Step 2: Clean, label, and standardize the collected data, extract the feature vectors of each data, and generate multi-modal feature vectors through dynamic weighted fusion by the attention mechanism; Step 3: Identify abnormal states based on the reconstruction error calculated by the autoencoder, quantify the contribution of each feature to the abnormality in combination with the SHAP algorithm, and locate the key influencing factors; Step 4: Dynamically adjust the detection threshold using reinforcement learning to reduce the false alarm and missed detection rates, evaluate the prediction confidence by quantifying the model error and data error, and achieve dynamic adjustment of the warning priority.

2. The operation data management method based on artificial intelligence according to claim 1, wherein: In Step 1, during the operation of the system, install intelligent sensors and data acquisition devices to record the operation status data at fixed time intervals: Device operation status acquisition: expressed as [R1, R2,..., Ru]; where u is a positive integer representing the number of device operation status types, and R1 to Ru respectively represent the 1st to u-th device operation statuses; Environmental parameter acquisition: expressed as [E1, E2,..., Ev]; where v is a positive integer representing the number of environmental parameter types, and E1 to Ev respectively represent the 1st to v-th environmental parameters; Load status acquisition: expressed as [L1, L2,..., Lw]; where w is a positive integer representing the number of load status types, and L1 to Lw respectively represent the 1st to w-th load statuses; Visual information acquisition: expressed as [V1, V2,..., Vz]; where z is a positive integer representing the number of device appearance types, and V1 to Vz respectively represent the 1st to z-th device appearances; Audio information acquisition: expressed as [A1, A2,..., Ap]; where p is a positive integer representing the number of device audio types, and A1 to Ap respectively represent the 1st to p-th device audios; Set the data acquisition time points, expressed as [T1, T2,..., Tm]; where m is a positive integer representing the number of device data acquisition time points, and T1 to Tm respectively represent the 1st to m-th data acquisition time points; the time intervals between adjacent moments are equal; Time synchronization: Use the NTP protocol to align the data in time to ensure that the multi-modal data is analyzed under the same time stamp.

3. The operation data management method based on artificial intelligence according to claim 2, characterized in that: In Step 2, preprocess the data: Data annotation: Manually annotate the abnormalities existing in the data; Data cleaning: Remove duplicate data and fill in the missing values with the mean; Invalid data elimination: Set the ranges of each parameter, filter invalid data based on physical rules; when there are conflicts in sensor data, filter invalid data based on state consistency detection; remove invalid data caused by device failures or transmissions; Data standardization: Standardize the data using the median-MAD; For numerical data, including device operating status, environmental parameters, and load status: Construct an LSTM time series model: Use multi-dimensional time series numerical data as input and utilize the Long Short-Term Memory (LSTM) network to model the data; The LSTM captures long-term dependencies in the time series through its memory cells, learns patterns and dynamic change patterns; Feature extraction is performed through a multi-layer LSTM structure, and the hidden state captures the temporal changes in the device state; Output a numerical feature vector and predicted values for future time points; The numerical feature vector is denoted as X1; For visual information data: Use visual information data as input, extract image features through a Convolutional Neural Network (CNN), and output a visual information feature vector; The visual information feature vector is denoted as X2; For audio information data: Use the Short-Time Fourier Transform to convert the audio into a time-frequency diagram, calculate the Mel Frequency Cepstral Coefficients (MFCC), and extract audio features; The Transformer performs temporal modeling: Based on the self-attention mechanism, learn long-term dependencies and identify abnormal sound patterns; Output an audio data feature vector; The audio data feature vector is denoted as X3; Adopt an attention module trained based on historical data to learn the importance of each feature vector through one or more neural network layers: Non-linearly combine each feature vector with a learned query vector, and then output an attention score through an activation function; Use the softmax operation to convert the original attention score into a normalized weight, ensuring that the sum of all weights is 1; Utilize historical data and corresponding anomaly detection results to train the attention module to maximize the detection accuracy; Update the parameters of the attention module through backpropagation and gradient descent algorithms; Input the latest feature vectors of each type, calculate the real-time weights α1, α2, and α3 according to the characteristics of the current data; Obtain the fused feature through weighted summation: X = α1·X1 + α2·X2 + α3·X3.

4. The operation data management method based on artificial intelligence according to claim 3, wherein: In step 3, use the multi-modal fused feature X as input; Train an autoencoder using a large amount of historical data; The encoder compresses the input data into a low-dimensional feature representation, and then the decoder reconstructs the original input; Use the Mean Squared Error (MSE) as the loss function to minimize the reconstruction error; Calculate the Euclidean distance between the current input and the reconstructed output as the reconstruction error ε: Set a threshold θ; When the real-time calculated reconstruction error ε exceeds the set threshold θ, determine that the current data is abnormal and record it; For the samples determined to be abnormal and their corresponding multi-modal fused features X, use the SHAP algorithm to calculate the SHAP values of each input feature; Compare the magnitudes of the SHAP values of each feature and select the top N factors that have the most influence on anomaly detection.

5. The operation data management method based on artificial intelligence according to claim 4, characterized in that: In step 4, define the false alarm rate F1: F1 = number of false alarms / total number of actually normal samples; where, "number of false alarms" refers to the number of samples that are wrongly determined to be abnormal when the system is in a normal state; Define the miss rate F2: F2 = number of missed detections / total number of actually abnormal samples; where, "number of missed detections" refers to the number of samples that actually have anomalies but are wrongly determined to be normal; Build a reinforcement learning intelligent agent to adjust the anomaly detection threshold θ according to the current state information in each decision cycle; the state information includes the false alarm rate F1, the missed alarm rate F2, and the model reconstruction error distribution in the current time period; Design reward function F = -λ1·F1-λ2·F2; where λ1 and λ2 are weight parameters set according to business requirements; maximize the reward function so that the intelligent agent automatically selects the optimal threshold adjustment strategy; Using historical data, the intelligent agent is pre-trained to master the preliminary threshold adjustment strategy; false alarms and missed alarms are monitored in real time, and the intelligent agent makes adjustments based on the current status, and conducts online learning and optimization through continuous feedback; feedback refers to the manual labeling of alarm results by operation and maintenance personnel, and the labeling information will be input into the reinforcement learning module as an additional supervision signal; Error source quantification: Model prediction error σ1: During the autoencoder training phase, the inherent mean square error of the model is obtained as an uncertainty index through the statistics of historical reconstruction errors; Input data error σ2: Combining the accuracy of the sensor and the errors generated by mean filling and standardization during data preprocessing, the overall uncertainty of the input data is calculated; Total uncertainty calculation: σ = (σ1 2 + σ2 2 ) 1 / 2 ; Set a threshold and adjust the confidence of the current anomaly detection result according to the σ value: when σ is higher than the threshold, it indicates that the uncertainty of the current prediction result is large, and the warning is postponed or the warning priority is lowered; when σ is not higher than the threshold, the alarm is triggered immediately or the warning priority is increased.

6. An operation data management system based on artificial intelligence, which is applied to an operation data management method based on artificial intelligence according to any one of claims 1-5, and is characterized in that: The system includes a data acquisition module, a preprocessing module, an anomaly detection module and an optimization module; The data acquisition module is used to collect multi-source data in real time through intelligent sensors and industrial equipment, achieve time synchronization, and ensure data alignment; The preprocessing module is used to clean, annotate and standardize the collected data, extract the feature vectors of each data, and generate multimodal feature vectors through dynamic weighted fusion of the attention mechanism; The anomaly detection module is used to identify abnormal states based on the reconstruction error calculated by the autoencoder, quantify the contribution of each feature to the anomaly in combination with the SHAP algorithm, and locate key influencing factors; The optimization module is used to dynamically adjust the detection threshold using reinforcement learning to reduce the false alarm and missed alarm rate, and to evaluate the prediction confidence by quantifying the model error and data error to achieve dynamic adjustment of the warning priority.

7. An operation data management system based on artificial intelligence according to claim 6, characterized in that: The data acquisition module includes a data acquisition unit and a time synchronization unit; The data acquisition unit is used to collect multi-source data of equipment operation status, environmental parameters, load status visual information and audio information in real time; The time synchronization unit is used to align the timestamps of multi-source data using the NTP protocol to ensure the timing consistency of multi-modal data.

8. An operation data management system based on artificial intelligence according to claim 7, characterized in that: The preprocessing module includes a data cleaning unit, a feature extraction unit and a fusion weighting unit; The data cleaning unit is used to remove duplicate and invalid data, filter outliers based on physical rules, and fill missing data with the mean value; The feature extraction unit is used to process numerical, image and audio data in parallel, and extract feature vectors using LSTM, CNN and Transformer respectively; The fusion weighting unit is used to dynamically calculate feature weights through an attention mechanism to generate a weighted fused multimodal feature vector.

9. An operation data management system based on artificial intelligence according to claim 8, characterized in that: The anomaly detection module includes a self-encoding unit, a threshold determination unit and a factor location unit; The autoencoder unit is used to learn the normal mode through the encoding-decoding structure and calculate the reconstruction error of the input data; The threshold determination unit is used to trigger abnormal alarm determination based on the comparison between the real-time reconstruction error and the dynamic threshold; The factor positioning unit is used to apply the SHAP algorithm to analyze the feature contribution degree and locate the key parameters that cause anomalies.

10. An operation data management system based on artificial intelligence according to claim 9, characterized in that: The optimization module includes a reinforcement learning unit, an error quantification unit, and a confidence adjustment unit; The reinforcement learning unit is used to construct a reward function based on the false alarm / miss rate and dynamically optimize the anomaly determination threshold; The error quantification unit is used to calculate the model prediction error and data acquisition error and evaluate the total uncertainty index; The confidence adjustment unit is used to dynamically adjust the early warning priority according to the uncertainty index to achieve credibility graded alarm.

Images