Intelligent abnormal value detection and processing method based on deep learning product quality data

Abstract

The invention provides an intelligent abnormal value detection and processing method based on deep learning product quality data. The method comprises the following steps: collecting multi-dimensional product quality data from a production line; an encoder module is constructed, and an encoder adopts a multi-layer neural network structure and is used for compressing and mapping input high-dimensional quality data into low-dimensional hidden layer feature representation; constructing a decoder module, receiving low-dimensional representation features output by an encoder, and reconstructing original data through a reverse neural network structure; training the automatic encoder by adopting an unsupervised learning mode, and learning a distribution mode and feature representation of normal data; and based on the trained automatic encoder, performing reconstruction error calculation on the quality data input in real time, judging an abnormal value through a preset threshold value, and triggering an alarm. Through the steps, the problem that high-precision detection of product quality data in production cannot be achieved in the prior art is solved, the alarm function is achieved, and therefore the quality problem of products in the future is avoided.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, specifically to an intelligent outlier detection and processing method based on deep learning product quality data. Background Technology

[0002] Poor product quality and insufficient durability not only affect the market competitiveness of manufacturers but can also lead to serious economic losses and damage to brand reputation. Therefore, how to effectively manage and handle intelligent outliers in quality data has become a pressing issue for manufacturers. Traditional data collection methods rely too heavily on manual labor, resulting in low efficiency, low data coverage, and a lack of timeliness. Current anomaly detection methods are largely based on experience-based judgment, lacking scientific rigor and a holistic, systematic approach; they remain superficial and cannot accurately identify quality data anomalies. Furthermore, traditional quality control processes rely heavily on manual operation, lacking intelligent and automated support. For alarms related to product anomalies, current technologies mostly rely on manual judgment after an alarm is triggered, resulting in insufficient timeliness and a lack of responsiveness.

[0003] To address the reliance on manual intervention in existing water quality anomaly detection methods, Chinese invention patent CN119416119A discloses a method and system for detecting water quality anomalies based on deep learning-generated errors. This approach acquires water quality monitoring data using multiple sensors such as temperature and pH, and categorizes the data into different datasets. A two-level verification method is then employed: a first-level verification is performed; if the first level fails, a second-level verification is performed to further detect anomalies. If the data passes verification, a deep learning recurrent neural network (RNN) model is used for analysis, calculating the generated error, and determining whether the water quality is abnormal based on the error. In practical applications, this model and method have improved the accuracy and reliability of water quality monitoring.

[0004] The above-mentioned scheme uses a deep learning-based method for monitoring water quality anomalies. It collects water quality data using multiple sensors and calculates errors by establishing an RNN deep learning model to determine water quality anomalies. However, relying solely on Class I and Class II validation for data filtering may result in some boundary anomalies not being identified in a timely manner. Furthermore, using RNNs for deep learning introduces a long-term dependency issue, affecting the accuracy of water quality trend prediction. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing technologies cannot achieve high-precision detection of product quality data during production and to implement an alarm function, thereby avoiding future product quality problems. This invention provides an intelligent outlier detection and processing method based on deep learning product quality data.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent outlier detection and processing method based on deep learning product quality data, the innovation of which lies in including the following steps:

[0007] S1. Data Acquisition: Collect multi-dimensional product quality data from the production line;

[0008] S2. Data Processing: Construct an encoder module. The encoder adopts a multi-layer neural network structure to compress and map the input high-dimensional quality data into low-dimensional hidden layer feature representations.

[0009] S3. Build a deep learning model: Construct a decoder module to receive the low-dimensional representation features output by the encoder and reconstruct the original data through a reverse neural network structure;

[0010] S4. Model Training: The autoencoder is trained using unsupervised learning to learn the distribution patterns and feature representations of normal data;

[0011] S5. Real-time outlier detection: Based on the trained autoencoder, the system performs reconstruction error calculation on the real-time input quality data, identifies outliers based on preset thresholds, and triggers alarms.

[0012] S6. Outlier Handling: When an immediate data anomaly is detected, the system will immediately issue an alarm signal to notify the operator to take emergency measures. At the same time, in order to avoid further losses, the system will automatically stop the operation of the relevant equipment. When a trend-based data anomaly is detected, the system will start an adaptive control algorithm and use the PID algorithm to optimize the control parameters online.

[0013] Furthermore, step 1 specifically involves: deploying various sensors and testing equipment on the production line to detect product data, collecting production quality data in real time during the production process, and clearly indicating the source and type of the data.

[0014] Furthermore, step 2 specifically involves: performing overall preprocessing on the collected data, cleaning the data, and removing duplicate data, missing values, and obviously erroneous data values; then performing data standardization using the Z-Score method.

[0015] First, calculate the mean μ:

[0016] Represents all data points x in the dataset i The summation is performed; after obtaining the mean, the standard deviation σ is calculated.

[0017]

[0018] Finally, calculate the standardized value z. i

[0019]

[0020] z i It refers to the data points after Z-Score standardization, representing x. i The degree of deviation from the mean μ.

[0021] Furthermore, step 3 specifically involves: selecting an autoencoder to build the model based on the characteristics of the product quality data and the requirements for outlier detection; the goal of the autoencoder is to map the input data x to a low-dimensional latent space representation z through the encoder, and then reconstruct the latent space representation z into output data through the decoder. Make the reconstructed output As close as possible to the original input x;

[0022] Encoder: Transforms input data X∈R n Compressed to a low-dimensional latent space Z∈R m The encoder contains L layers of neural networks, with the input to the l-th layer being h. (l-1) The output is h (l) , where h (0) =x; each layer uses linear transformation and nonlinear activation; the linear transformation is the transformation of the input h (l-1) Through the weight matrix W (l) and bias vector b (l) Implement and obtain intermediate result a (l) The formula is expressed as:

[0023] a (l) =W (l) h (l-1) +b (l)

[0024] The result of the transformation a (l) By applying the nonlinear activation function σ(·), the output h of the l-th layer is obtained. (l) The formula is as follows:

[0025] h (l) =σ(a (l) )

[0026] After L layers of transformation, the encoder output (latent space) is represented as: Z = h (L)

[0027] Decoder: Reconstructs the latent space representation z into an output with the same dimension as the original input data. The decoder has a K-layer neural network, where the input to the k-th layer is g. (k-1) The output is g(k) , where g (0) =Z; the transformation process is similar to that of the encoder; expressed by the formula:

[0028] d (k) =V (k) g (k-1) +c (k)

[0029] The result of the transformation d (k) By applying the nonlinear activation function ρ(·), the output g of the k-th layer is obtained. (k) ; expressed as a formula:

[0030] g (k) =ρ(d (k) )

[0031] After K layers of transformation, the decoder output is the reconstructed result:

[0032] Loss function: The mean squared error (MSE) loss is used to measure the difference between the original input x and the reconstructed output. The differences between them; the loss function for a single sample is:

[0033]

[0034] For a training set {x1, x2, ..., xn} containing N samples... N The overall loss function is:

[0035]

[0036] Where θ represents all trainable parameters in the autoencoder; expressed as θ={W (1) b (1) , ···, W (L) ,b (L) V (1) c (1) ,···,V (K) ,c (K)}

[0037] Furthermore, step 4 specifically involves minimizing the overall loss function J(θ) using an optimization algorithm during model training; this is achieved using the Adam optimization algorithm.

[0038] The algorithm uses first-moment estimates (mean) and second-moment estimates (variance) to dynamically adjust the learning rate for each parameter;

[0039] First-order moment estimation update formula:

[0040]

[0041] Second-order moment estimation update formula:

[0042]

[0043] Bias corrections should be applied to the first-order and second-order moment estimates to reduce bias in the initial stage; the formula is:

[0044]

[0045] Finally, the parameter θ is updated based on the corrected first-order moment estimate and second-order moment estimate, and the update step size is controlled by the learning rate α; expressed as:

[0046]

[0047] By iterating through the above steps, the parameters θ of the autoencoder are adjusted so that the model learns the normal patterns and features of the data.

[0048] Furthermore, step 5 specifically involves using streaming inference or multi-dimensional correlation analysis to quickly evaluate the new data and determine whether it is an anomaly.

[0049] Furthermore, the streaming inference is specifically as follows:

[0050] (1) Data input: For real-time data streams {x1, x2, ..., x...} t}, where x t Let x represent the input data vector received at time t. t =[x t1 ,x t2 ,···,x tn ], where n is the feature dimension of the data;

[0051] (2) Model calculation reconstruction error: The trained autoencoder deep learning model is used to calculate the input data x. t Reconstruct the data to obtain the reconstructed data. The reconstruction error is represented by the mean squared error (MSE):

[0052]

[0053] E t Let xti represent the reconstruction error at time t, and xti be the i-th eigenvalue of the input data vector. It is the i-th eigenvalue of the reconstructed data vector;

[0054] (3) Comparison with dynamic threshold: The dynamic threshold is calculated using an exponentially weighted moving average (EWMA):

[0055] θ t =αE t-1+(1-α)θ t-1

[0056] α is the smoothing coefficient, with a value range of (0,1), E t-1 θ represents the reconstruction error from the previous time step. t-1 It is the threshold of the previous time step; when the reconstruction error E is at this moment t With dynamic threshold θ t Compare, if E t >θ t If so, the current data point is considered an outlier;

[0057] (4) Output anomaly probability: Use logical functions to convert reconstruction error into anomaly probability to accurately assess the degree of anomaly;

[0058]

[0059] P t Let t represent the anomaly probability at time t, and k be an adjustment parameter that controls the steepness of the probability curve.

[0060] Furthermore, the multi-dimensional correlation analysis is detailed as follows:

[0061] Multidimensional correlation analysis combines time series correlation and spatial correlation to fully explore the potential information in the data, thereby improving the ability to identify complex anomalies.

[0062] (1) Time series correlation analysis: For time series data of a single variable (such as the temperature of an equipment) {y1, y2, ..., y...} t The autocorrelation function (ACF) is used to measure the correlation between different time points; it is defined as:

[0063]

[0064] ρ(k) represents the autocorrelation coefficient with a lag of k periods, and T is the length of the time series. It is the mean of the time series; a threshold range for the autocorrelation coefficient is set. If the autocorrelation coefficient of a certain lag k period exceeds this range, it is considered that the correlation of the time series has changed abnormally.

[0065] (2) Spatial correlation analysis: For data from multiple workstations on the same production line, assuming there are m workstations in total, and each workstation has n features; at time t, the data matrix is ​​X t =[x t1 ,x t2 ,···,x tm ] T , where x tiLet be the feature vector of the i-th workstation; the Pearson correlation coefficient is used to measure the correlation between different workstations; for workstations i and j, the Pearson correlation coefficient is defined as:

[0066]

[0067] x tik It is the k-th feature value of the i-th workstation at time t. It is the mean of the feature vector of the i-th workstation; similarly, a threshold range for the correlation coefficient can be set. If the correlation coefficient between a pair of workstations exceeds this range, it is considered that the spatial correlation has become abnormal.

[0068] (3) Comprehensive judgment of anomalies: The analysis results of time series correlation and spatial correlation are combined by weighted summation to obtain the comprehensive anomaly score St;

[0069]

[0070] in, It is an anomaly score in time series correlation. It is the anomaly score of spatial correlation, where w1 and w2 are weighting coefficients, and w1+w2=1;

[0071] Based on the comprehensive anomaly score S t It is compared with a preset threshold to determine whether there is a complex anomaly.

[0072] The beneficial effects of this invention are:

[0073] 1. High precision:

[0074] This invention employs a deep learning model that can automatically learn complex patterns and features in data, resulting in higher accuracy and adaptability in outlier detection. It can effectively identify outliers that are difficult to detect using traditional methods.

[0075] 2. Improve production efficiency and product quality:

[0076] This invention enables timely and accurate detection and processing of outliers in product quality data, which helps to identify problems in the production process in a timely manner, take corresponding measures for adjustment and improvement, thereby improving product quality, reducing production costs, and increasing production efficiency.

[0077] 3. Intelligent decision-making optimization:

[0078] This invention avoids the data loss problem caused by simply deleting or replacing outliers by classifying and processing them according to different strategies. It can handle outliers more reasonably and improve the integrity of data and the accuracy of analysis results.

[0079] 4. Versatility:

[0080] The method of this invention can be applied to different types of product quality data, supports access to multiple data sources (such as PLC and machine vision sensors), has strong versatility and adaptability, and can meet the needs of multiple industries such as aerospace manufacturing. Attached Figure Description

[0081] Figure 1 This is a flowchart of the present invention;

[0082] Figure 2 This is a flowchart illustrating the application process of the present invention. Detailed Implementation

[0083] The invention will now be further described with reference to the accompanying drawings.

[0084] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0085] Example 1

[0086] See Figure 1 and Figure 2 The specific steps of this invention are as follows:

[0087] 1. Data Collection

[0088] Various sensors and testing equipment are deployed on the production line to collect production quality data in real time. The source and type of the data are clearly labeled. This mainly includes, but is not limited to, product dimensions, temperature, quality, and performance indicators.

[0089] 2. Data Processing

[0090] The collected data first undergoes overall preprocessing, including data cleaning to remove duplicate data, missing values, and obviously erroneous data. Then, data standardization is performed using the Z-Score method.

[0091] First, calculate the mean μ:

[0092]

[0093] Represents all data points x in the dataset i The summation of the total values. After obtaining the mean, calculate the standard deviation σ:

[0094]

[0095] Finally, calculate the standardized value z. i

[0096]

[0097] z i It refers to the data points after Z-Score standardization, representing x. i The degree of deviation from the mean μ.

[0098] 3. Build a deep learning model

[0099] Based on the characteristics of product quality data and the need for outlier detection, an autoencoder was chosen for model building. The goal of an autoencoder is to map the input data x to a low-dimensional latent space representation z using an encoder, and then reconstruct the latent space representation z into output data using a decoder. Make the reconstructed output Get as close as possible to the original input x.

[0100] Encoder: Transforms input data X∈R n Compressed to a low-dimensional latent space Z∈R m The encoder consists of L layers of neural networks, with the input to the l-th layer being h. (l-1) The output is h (l) , where h (0) =x. Each layer uses linear transformation and nonlinear activation. The linear transformation is the transformation of the input h. (l-1) Through the weight matrix W (l) and bias vector b (l) Implement and obtain intermediate result a (l) The formula is expressed as:

[0101] a (l) =W (l) h (l-1) +b (l)

[0102] The result of the transformation a (l) By applying the nonlinear activation function σ(·), the output h of the l-th layer is obtained. (l) Expressed as a formula:

[0103] h (l) =σ(a (l) )

[0104] After L layers of transformation, the encoder output (latent space) is represented as: Z = h (L)

[0105] Decoder: Reconstructs the latent space representation z into an output with the same dimension as the original input data. The decoder has a K-layer neural network, where the input to the k-th layer is g.(k-1) The output is g (k) , where g (0) =Z. The transformation process is similar to that of the encoder. It can be expressed by the formula:

[0106] d (k) =V (k) g (k-1) +c (k)

[0107] The result of the transformation d (k) By applying the nonlinear activation function ρ(·), the output g of the k-th layer is obtained. (k) Expressed as a formula:

[0108] g (k) =ρ(d (k) )

[0109] After K layers of transformation, the decoder output is the reconstructed result:

[0110] Loss function: The mean squared error (MSE) loss is used to measure the difference between the original input x and the reconstructed output. The differences between them. The loss function for a single sample is:

[0111]

[0112] For a training set {x1, x2, ..., xn} containing N samples... N The overall loss function is:

[0113]

[0114] Where θ represents all trainable parameters in the autoencoder. This is represented as θ = {W} (1) b (1) , ···, W (L) ,b (L) V (1) c (1) ,···,V (K) ,c (K)}

[0115] 4. Model Training

[0116] The overall loss function J(θ) is minimized using an optimization algorithm during model training. This is achieved using the Adam optimization algorithm.

[0117] The algorithm uses first-moment estimates (mean) and second-moment estimates (variance) to dynamically adjust the learning rate for each parameter.

[0118] First-order moment estimation update formula:

[0119]

[0120] Second-order moment estimation update formula:

[0121]

[0122] Bias corrections should be applied to the first-order and second-order moment estimates to reduce initial bias. The formula is:

[0123]

[0124] Finally, the parameter θ is updated based on the corrected first-order moment estimate and second-order moment estimate, with the update step size controlled by the learning rate α. This is expressed as:

[0125]

[0126] By iterating through the above steps, the parameters θ of the autoencoder are adjusted so that the model learns the normal patterns and features of the data.

[0127] 5. Real-time outlier detection

[0128] Use streaming inference to quickly evaluate new data and determine whether it is anomaly.

[0129] (1) Data input: For real-time data streams {x1, x2, ..., x...} t}, where x t Let x represent the input data vector received at time t. t =[x t1 ,x t2 ,···,x tn ], where n is the feature dimension of the data.

[0130] (2) Model calculation reconstruction error: The trained autoencoder deep learning model is used to calculate the input data x. t Reconstruct the data to obtain the reconstructed data. The reconstruction error is represented by the mean squared error (MSE):

[0131]

[0132] E t Let xti represent the reconstruction error at time t, and xti be the i-th eigenvalue of the input data vector. It is the i-th eigenvalue of the reconstructed data vector.

[0133] (3) Comparison with dynamic threshold: The dynamic threshold is calculated using an exponentially weighted moving average (EWMA):

[0134] θ t =αEt-1 +(1-α)θ t-1

[0135] α is the smoothing coefficient, with a value range of (0,1), E t-1 θ represents the reconstruction error from the previous time step. t-1 It is the threshold from the previous time step. When the reconstruction error E is at this moment... t With dynamic threshold θ t Compare, if E t >θ t If the current data point is considered an outlier, then it is considered an outlier.

[0136] (4) Output anomaly probability: Use a logic function to convert the reconstruction error into anomaly probability to accurately assess the degree of anomaly.

[0137]

[0138] P t Let t represent the anomaly probability at time t, and k be an adjustment parameter that controls the steepness of the probability curve.

[0139] 6. Outlier Handling

[0140] When performing anomaly detection, deep learning models output an anomaly probability value for each data point. This probability value reflects the degree to which the data point deviates from the normal pattern.

[0141] Anomalies are categorized into immediate anomalies and trend-based anomalies. Immediate anomalies typically manifest as sudden equipment failures, characterized by their abruptness and severity, immediately and severely impacting production and potentially leading to product quality defects and other serious accidents. Trend-based anomalies are instances where parameters slowly deviate from their normal range, resulting in declining product quality and reduced equipment performance over time.

[0142] When an immediate data anomaly is detected, the system will immediately issue an alarm signal to notify operators to take emergency measures. Simultaneously, to prevent further losses, the system will automatically shut down the relevant equipment.

[0143] When an anomaly is detected as a trend-type data anomaly, the system will activate an adaptive control algorithm, using a PID algorithm to optimize the control parameters online.

[0144] By utilizing a feedback mechanism, the results of anomaly handling are fed back to the Manufacturing Execution System (MES), forming a closed-loop control of "detection-processing-optimization." The system can then understand the effectiveness of anomaly handling in real time, providing a basis for subsequent production decisions.

[0145] Example 2

[0146] When performing outlier detection, multi-dimensional correlation analysis can be used instead of streaming inference.

[0147] Multidimensional correlation analysis combines time series correlation and spatial correlation to fully explore the potential information in the data, thereby improving the ability to identify complex anomalies.

[0148] Time series correlation analysis: For time series data {y1, y2, ..., y...} of a single variable (such as the temperature of a device) t The autocorrelation function (ACF) is used to measure the correlation between different time points. It is defined as:

[0149]

[0150] ρ(k) represents the autocorrelation coefficient with a lag of k periods, T is the length of the time series, and y is the mean of the time series. A threshold range for the autocorrelation coefficient is set; if the autocorrelation coefficient with a lag of k periods exceeds this range, the correlation of the time series is considered to have changed abnormally.

[0151] Spatial correlation analysis: For data from multiple workstations on the same production line, suppose there are m workstations in total, and each workstation has n features. At time t, the data matrix is ​​X. t =[x t1 ,x t2 ,···,x tm ] T , where x ti Let be the feature vector of the i-th workstation. The Pearson correlation coefficient is used to measure the correlation between different workstations. For workstations i and j, the Pearson correlation coefficient is defined as:

[0152]

[0153] x tik It is the k-th feature value of the i-th workstation at time t. It is the mean of the feature vector of the i-th workstation. Similarly, a threshold range for the correlation coefficient can be set. If the correlation coefficient between a pair of workstations exceeds this range, the spatial correlation is considered to be abnormal.

[0154] (3) Comprehensive judgment of anomalies: The analysis results of time series correlation and spatial correlation are combined by weighted summation to obtain the comprehensive anomaly score St.

[0155]

[0156] in, It is an anomaly score in time series correlation. It is the anomaly score of spatial correlation, where w1 and w2 are weighting coefficients, and w1+w2=1.

[0157] Based on the comprehensive anomaly score S t It is compared with a preset threshold to determine whether there is a complex anomaly.

[0158] In summary, the present invention has the following advantages:

[0159] (1) Multi-dimensional data acquisition and preprocessing: Supports access to multiple data sources (such as PLC and sensor data), covering multi-dimensional quality data such as product size, temperature, and performance. Z-Score standardization is used to eliminate dimensional differences and improve the quality of model input data.

[0160] (2) Autoencoder model construction and training: A multi-layer neural network structure is adopted for the encoder and decoder. The feature distribution of normal data is automatically extracted through unsupervised learning to achieve low-dimensional mapping and reconstruction of high-dimensional quality data. The mean squared error (MSE) loss function and Adam optimization algorithm are combined to dynamically adjust the model parameters and improve the anomaly detection accuracy.

[0161] (3) Dynamic threshold and anomaly probability calculation: An exponentially weighted moving average (EWMA) is introduced to dynamically update the threshold, adapting to changes in the real-time data stream. The reconstruction error is converted into anomaly probability through a logical function to quantify the degree of anomaly and avoid misjudgment.

[0162] (4) Anomaly Classification and Handling: Anomalies are classified into immediate (sudden failures) and trend-based (slow parameter deviations), and immediate shutdown alarms and adaptive PID control optimization measures are taken respectively. Combined with a feedback mechanism, the processing results are fed back to the Manufacturing Execution System (MES) to form a closed-loop control of "detection-processing-optimization".

[0163] The core of this invention lies in intelligent outlier detection and processing based on deep learning-based product quality data. Specific protection points include, but are not limited to, multi-dimensional data fusion and preprocessing procedures, the construction and training of deep learning autoencoder models, anomaly classification and closed-loop control strategies, and multi-dimensional correlation analysis in alternative solutions.

[0164] The above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent outlier detection and processing method based on deep learning product quality data, characterized in that, The method comprises the following steps: S1, data acquisition: collecting multi-dimensional product quality data from the production line; S2, data processing: constructing an encoder module, the encoder adopts a multi-layer neural network structure for compressing and mapping the input high-dimensional quality data into low-dimensional hidden layer feature representation; S3, building a deep learning model: constructing a decoder module to receive the low-dimensional representation features output by the encoder and reconstructing the original data through a reverse neural network structure; S4, model training: training the autoencoder in an unsupervised learning manner to learn the distribution pattern and feature representation of normal data; S5, real-time anomaly detection: based on the trained autoencoder, the reconstruction error of the real-time input quality data is calculated, and the abnormal value is determined by a preset threshold and an alarm is triggered; S6, abnormal value processing: when detecting an immediate data anomaly, the system will immediately issue an alarm signal to notify the operator to take emergency measures; at the same time, in order to avoid further loss, the system will automatically stop the operation of related equipment; when detecting a trend data anomaly, the system will start an adaptive control algorithm to optimize the control parameters online using a PID algorithm. 2.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 1, characterized in that, The step 1 is specifically: arranging various sensors and detection equipment on the production line to collect real-time production quality data during the production process, and marking the source and type of the data. 3.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 1, characterized in that, The step 2 is specifically: first, performing overall preprocessing on the collected data to clean the data, remove duplicate data, missing values, and obviously incorrect data values; then, performing data standardization using the Z-Score method; First, calculate the mean mu: representing the sum of all data points x i the total value of the sum; after calculating the mean, the standard deviation σ is calculated: Finally the normalized value z is calculated i z i is the data point after Z-Score standardization, indicating x i the degree of deviation from the mean μ. 4.The method of claim 1, wherein, The step 3 is specifically: according to the characteristics of product quality data and the requirement of abnormal value detection, a self-encoder is selected to build a model; the target of the self-encoder is to map the input data x to a low-dimensional latent space representation z through an encoder, and then reconstruct the latent space representation z to output data through a decoder So that the reconstructed output Is as close as possible to the original input x; Encoder: compress input data X ∈ R n to a low-dimensional latent space Z ∈ R m ; the encoder contains L layers of neural networks, the input of the l-th layer is h (l-1) , and the output is h (l) , where h (0) = x; each layer is implemented by a linear transformation and a nonlinear activation; the linear transformation is realized by a weight matrix W (l-1) and a bias vector b (l) on the input h (l) , and an intermediate result a (l) is obtained; the formula is represented as: a (l) = W (l) h (l-1) + b (l) The result a of the transformation (l) Applying a non-linear activation function σ(·), the output h of the l-th layer is obtained (l) ; which is expressed in formula as: h (l) = σ(a (l) ) After the transformation through L layers, the output of the encoder (latent space) is represented as: Z = h (L) Decoder: reconstructs latent space representation z into output of same dimension as original input data The decoder has K layers of neural networks, the input of the k-th layer is g (k-1) and the output is g (k) where g (0) = z; the transformation process is similar to the encoder; mathematically expressed as: d (k) = V (k) g (k-1) + c (k) The result d of the transformation (k) Applying a non-linear activation function p(·) to obtain the output g of the kth layer (k) ; expressed in formula: g (k) = p(d (k) ) After the transformation through K layers, the output of the decoder is the reconstruction result as follows: Loss function: Mean Squared Error (MSE) loss is used to measure the difference between the original input x and the reconstructed output ; the loss function for a single sample is: For a training set {x1, x2, ..., xn} containing N samples... N The overall loss function is: where θ denotes all trainable parameters in the encoder; and is denoted as θ = {W (1) , b (1) ,..., W (L) , b (L) , V (1) , c (1) ,..., V (K) , c (K)}. 5.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 1, characterized in that, The step 4 is specifically: minimizing the overall loss function J(θ) through the optimization algorithm in model training; and achieving through the Adam optimization algorithm; This algorithm uses first-order moment estimation (mean) and second-order moment estimation (variance) to dynamically adjust the learning rate of each parameter; The first-order moment estimation update formula is: The second-order moment estimation update formula is: The first-order moment estimation and the second-order moment estimation are corrected to reduce the bias in the initial stage; the formula is: Finally, the parameters θ are updated according to the corrected first-order moment estimation and second-order moment estimation, and the step size of the update is controlled by the learning rate α; it is represented as: Through the above steps, the parameters θ of the autoencoder are adjusted, and the model learns the normal pattern and features of the data. 6.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 1, characterized in that, The step 5 is specifically: using stream inference or multi-dimensional correlation analysis to quickly evaluate new data and determine whether it is abnormal. 7.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 6, characterized in that, The stream inference is specifically as follows: (1) Data input: for real-time data stream as {x1, x2, …, xn}, where x t represents the input data vector received at time t, x t = [x t , x t1 , …, x t2 ], n is the feature dimension of data; tn ​ (2) Model reconstruction error: using the trained autoencoder deep learning model to reconstruct the input data x t , and get the reconstructed data The reconstruction error is expressed by mean square error (MSE): E t denotes the reconstruction error at time t, xti is the i-th eigenvalue of the input data vector, is the i-th eigenvalue of the reconstructed data vector; (3) Comparison with dynamic threshold: using exponential weighted moving average (EWMA) to calculate the dynamic threshold: θ t = aE t-1 + (1 - a) θ t-1 a is a smoothing coefficient, and a value range is (0, 1), E t-1 is a reconstruction error of the last moment, and θ t-1 is a threshold value of the last moment; when the reconstruction error E t of the current moment is compared with the dynamic threshold value θ t , if E t > θ t , the current data point is considered as an abnormal point; (4) Output anomaly probability: using a logic function to convert the reconstruction error into an anomaly probability to accurately evaluate the degree of abnormality; P t represents the abnormal probability at time t, and k is an adjustment parameter for controlling the steepness of the probability curve. 8.The intelligent outlier detection and processing method based on deep learning product quality data according to claim 6, characterized in that, The multi-dimensional correlation analysis is specifically as follows: Multi-dimensional correlation analysis fully excavates the potential information in the data by combining time series correlation and spatial correlation, thereby improving the recognition ability of complex anomalies; (1) Time series correlation analysis: For the time series data {y1, y2, …, y t} of a single variable (e.g., temperature of a device), the autocorrelation function (ACF) is used to measure the correlation between different time points; defined as: ρ(k) denotes autocorrelation coefficient of lag k, T is the length of time series, is the mean of time series; a threshold range of autocorrelation coefficient is set, if autocorrelation coefficient of lag k is beyond the range, it is considered that the correlation of time series has abnormal change; (2) Spatial correlation analysis: For multiple station data of the same production line, there are m stations in total, each station has n characteristics; at time t, the data matrix is X t =[x t1 ,x t2 ,…,x tm ] T , where x ti is the characteristic vector of the i-th station; the Pearson correlation coefficient is used to measure the correlation between different stations; for stations i and j, the Pearson correlation coefficient is defined as: x tik is the kth feature value of the ith station at time t, is the mean of the ith station feature vector; similarly, a threshold range of the correlation coefficient can be set, and if the correlation coefficient between some station pairs exceeds the range, it is considered that the spatial correlation has been abnormal. (3) Comprehensive judgment of abnormality: the analysis results of time series correlation and spatial correlation are comprehensively obtained by weighted summation, and a comprehensive abnormality score S is obtained t ; wherein, is the anomaly score of the time series correlation, is the anomaly score of the spatial correlation, w1 and w2 are weight coefficients, w1 + w2 = 1; According to the comprehensive abnormality score S t The preset threshold is compared to determine whether there is a complex abnormality.

Images