An edge cloud cooperative device multi-working condition abnormal mode identification method

Abstract

The application provides a kind of edge cloud collaborative equipment multi-condition abnormal mode identification method, comprising the following steps: complex manufacturing system equipment sensor deployment and multi-modal processing data acquisition: equipment time series data preprocessing: the multi-modal processing data collected from each sensor is cleaned, standardized and feature engineered on the edge side;Edge side model building and updating: on the edge side, the improved OPTICS algorithm is used to cluster the preprocessed equipment time series data, and the other data in the entire cluster obtained by clustering is identified using label data to update the edge side database;Upload the equipment time series data preprocessed on the edge side in step two to the cloud, and deploy a transfer learning model on the cloud;Precise identification of abnormal patterns in complex manufacturing systems is achieved using edge-cloud collaboration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of complex manufacturing system equipment monitoring and fault diagnosis, and particularly relates to a device multi-condition abnormal pattern recognition method based on an edge-cloud collaborative architecture, which realizes efficient and accurate recognition of abnormal patterns in multiple conditions during device operation by fusing an improved OPTICS algorithm (Ordering Points to identify the clustering structure) and a deep adversarial neural network (DANN), so as to ensure stable operation of a complex manufacturing system and improve production efficiency. BACKGROUND

[0002] With the continuous upgrading of global manufacturing and the improvement of national industrialization level, manufacturing has become an important pillar of economic development in various countries. In modern complex manufacturing systems, the stable operation of equipment is of vital importance to production efficiency, product quality and enterprise economic benefits. With the continuous development of manufacturing technology, the structure and operating conditions of equipment are becoming increasingly complex, and traditional equipment abnormal pattern recognition methods face many challenges.

[0003] Traditional methods based on single sensor data or simple threshold judgment are difficult to comprehensively and accurately capture abnormal information of equipment under multiple conditions. Due to the characteristics of multi-modal, time series and complex data distribution of complex manufacturing system equipment operation data, the analysis of a single data source often has limitations, and cannot comprehensively consider the influence of multiple key factors on the equipment state, resulting in low accuracy of abnormal pattern recognition [1] .

[0004] In terms of data processing, traditional methods lack effective data preprocessing and feature extraction methods, making it difficult to extract valuable information for abnormal identification from massive raw data, and when facing data of different magnitudes and dimensions, they cannot perform effective normalization processing, affecting the accuracy of subsequent model training and analysis.

[0005] Existing clustering algorithms such as the OPTICS algorithm have difficulties in parameter selection, high sensitivity to data, and poor performance in handling density-changing data distribution when applied to equipment abnormal pattern recognition, resulting in unsatisfactory clustering results and inaccurate classification of equipment abnormal pattern categories [2] . In addition, traditional methods lack effective transfer learning ability for different condition data, and when faced with new condition data, they need to collect a large amount of data and train the model again, which is low in computational efficiency and difficult to meet the needs of real-time monitoring and rapid fault diagnosis of intelligent manufacturing systems [3] .

[0006] At the same time, the data architecture of modern complex manufacturing systems is developing towards a cloud-edge-end mode, forming an "end-side collection, edge-side aggregation, cloud processing" data uplink and a "cloud training, edge inference, end execution" data downlink [4] The changes in data structure also pose new requirements for the deployment of algorithms, which need to fully leverage the advantages of cloud-edge-end architecture.

[0007] In summary, there is a need for a new device multi-condition abnormal pattern recognition method to overcome the shortcomings of traditional methods, improve the accuracy and computational efficiency of complex manufacturing system device abnormal pattern recognition, and ensure the stability and reliability of the production process.

[0008] Related Literature

[0009] [1]LONG M, CAO Y, WANG J, et al. Learning Transferable Features with Deep Adaptation Networks [A / OL]. arXiv, 2015 [2024-11-25]. http: / / arxiv.org / abs / 1502.02791. DOI:10.48550 / arXiv.1502.02791.

[0010] [2]RAJABI F, MCARTHUR J J. Applying OPTICS with and without PCA for fault detection of fan coil units using building automation system data [J / OL]. Energy and Buildings, 2024, 317:114368. DOI:10.1016 / j.enbuild.2024.114368.

[0011] [3]Shan F, Wang C, Zeng J, et al. Inconsistency fault degree analysis method of power battery monomer based on OPTICS clustering and voltage anomaly index [J]. Science in China Series E: Technological Sciences, 2024, 54(8):1611-1624.

[0012] [4]ECC releases "Edge Computing and Cloud Computing Cooperation White Paper 2.0", "5G Era Industrial Internet Edge Computing Network White Paper", and "Industrial Internet Edge Computing Node White Paper 1.0" [J]. Automation Review, 2021, 38(2):3-3. SUMMARY

[0013] The application provides a method for identifying abnormal patterns of equipment in multiple working conditions of a complex manufacturing system.

[0014] A method for identifying abnormal patterns of equipment in multiple working conditions in an edge-cloud collaborative manner, comprising the following steps:

[0015] Step 1: Sensor deployment of equipment in a complex manufacturing system and acquisition of multi-modal processing data

[0016] Step 2: Preprocessing of equipment time series data: cleaning, standardizing and feature engineering of multi-modal processing data collected from sensors on the edge side

[0017] Step 3: Building and updating of edge-side model: on the edge side, an improved OPTICS algorithm is used to cluster the preprocessed equipment time series data, and other data in the entire cluster identified by the label data is used to update the edge-side database; wherein the improved OPTICS algorithm is as follows: the local density calculation method is improved, the local density is calculated based on the reciprocal of the average distance of k-nearest neighbors of the data point; the density threshold is adaptively adjusted, the density threshold is determined according to the mean and standard deviation of the local density of the data point; the cluster center is determined based on the density peak value, the data point with local density greater than the density threshold is taken as the cluster center; the merging operation of the cluster, the distance matrix between the cluster centers is calculated, when the cluster center distance is less than the preset distance threshold and the local density difference is within a certain range, the cluster merging operation is performed;

[0018] Step 4: uploading the equipment time series data preprocessed on the edge side to the cloud, and deploying a transfer learning model on the cloud

[0019] Step 5: identifying abnormal patterns of a complex manufacturing system by using edge-cloud collaboration

[0020] Further, the method of step 2 is as follows:

[0021] 1) Data cleaning, including

[0022] Missing value processing: Newton interpolation method is used

[0023] Abnormal value monitoring and processing: box plot technology and 3σ principle are used to identify abnormal values

[0024] Deletion of duplicate data

[0025] 2) Data standardization: Min-Max standardization is adopted;

[0026] 3) Feature engineering: the method of feature construction and multi-dimensional feature fusion is adopted, and the automatic encoder is used to fuse multiple time series signals containing device state features to extract fault signal features.

[0027] Further, in step three, the density threshold ρ th The calculation formula is as follows:

[0028]

[0029] In the formula, is the mean and σ ρ is the standard deviation.

[0030] Further, in step three, the distance threshold is 0.1 of the average near-neighbor distance; the local density difference is within a certain range, which means that the local density difference is within the range of 0.1 of the standard deviation of the local density of the data points.

[0031] Further, in step three, the method of cluster merging operation is: for two cluster centers, comparing the data point density of the two cluster centers, merging the cluster center with smaller data point density into another cluster center, and modifying the cluster label of the related data points to the same label as the other cluster center.

[0032] Further, in step four, the pre-trained network RestNeXt is used to extract features from the preprocessed device time series data.

[0033] Further, the method of step five is:

[0034] 1) Joint training using cloud-stored historical operating condition data to identify the state of the operating state data;

[0035] 2) Identify the mode of the operating state data using the DANN-based transfer recognition architecture: use the ResNeXt pre-trained network to further extract features from the historical operating condition data and the operating state data; initialize the DANN network, use the further extracted feature data, and use the contrast loss joint training method to train the DANN network, while using the back propagation method to update the parameters;

[0036] 4) Use the trained DANN network to identify the mode of the operating state data and evaluate the identified results;

[0037] 5) After identifying sufficient labeled data, the labels are issued to the edge-side database for updating the edge-side database.

[0038] The patent is based on the improved OPTICS algorithm and DANN algorithm using edge cloud collaborative data distribution architecture, which effectively realizes complex manufacturing system equipment multi-working condition abnormal mode recognition with real-time, high efficiency and accuracy, etc. Advantages, providing an efficient solution for complex manufacturing system equipment abnormal mode recognition. This method not only can improve production efficiency, reduce maintenance cost, but also can enhance the safety and stability of production line, inject new vitality and power for the development of intelligent manufacturing industry. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 Migration learning technology route

[0040] Figure 2 Clustering algorithm route

[0041] Figure 3 DANN network structure

[0042] Figure 4 ResNeXt network structure data DETAILED DESCRIPTION

[0043] The present application is realized by the following technical scheme, the migration learning process is as shown in Figure 1 The specific implementation steps are as follows:

[0044] Step 1: Complex manufacturing system equipment sensor deployment and multi-modal processing data acquisition

[0045] 1) According to the specific structure of the complex manufacturing system equipment and the operation characteristics of the key parts, select and arrange multiple types of sensors. Deploy the following types of sensors (including but not limited to the following categories) at each key node:

[0046] Vibration sensor, installed on the mechanical moving parts. Vibration sensors can detect small vibrations and irregular movements during equipment operation, which is of great significance for early warning of mechanical failures such as bearing damage and gear wear;

[0047] Temperature sensor, fixed near heat generating components such as motors and transformers. These sensors can monitor temperature changes during equipment operation;

[0048] Current sensor, installed on the main power line, used to monitor the current state of the cable production equipment. Abnormal fluctuations in current are often a precursor to electrical faults such as short circuits or aging of electrical components;

[0049] Sound sensor, deployed around noisy equipment, to identify mechanical failures by analyzing changes in sound spectrum.

[0050] 2) Data collection and integration

[0051] After the sensor deployment is complete, all sensors are connected to a centralized data collection system, which is responsible for collecting data from each sensor and forwarding it to the main data processing center on the edge side.

[0052] 3) Communication Technology

[0053] Industrial Ethernet or wireless communication technology is used to ensure real-time data transmission. High data transmission rate and strong anti-interference ability are used to ensure the reliability and real-time performance of data transmission.

[0054] 4) Data Visualization

[0055] Advanced data visualization techniques are implemented, and real-time data dashboards are used to provide operators with real-time production line status displays. These visualization tools not only help operators quickly identify problem areas, but also monitor production processes in real time to optimize operational decisions.

[0056] Step 2: Preprocessing of equipment time series data

[0057] Data preprocessing is a crucial step in identifying abnormal patterns in complex manufacturing systems. To ensure the accuracy and effectiveness of subsequent model training, raw data collected from sensors is thoroughly cleaned, standardized, and feature engineered on the edge side. The specific steps are as follows:

[0058] 1) Data cleaning. The main purpose of data cleaning is to exclude invalid, erroneous or incomplete data that may interfere with model training and prediction accuracy. This process mainly includes: missing value processing, since this patent collects time series data, Newton interpolation method is used to estimate missing values. The steps of using Newton interpolation method for calculation are as follows: first, find all order difference quotient formulas of known n points (x1, y1), (x2, y2)…(x n ,y n )

[0059]

[0060] Solve the above difference quotient formulas to establish the following interpolation polynomial f(x).

[0061] f(x)=f(x1) + (x-x1)f[x2,x1] + (x-x1)(x-x2)f[x3,x2,x1]+(x-x1)(x-x2)(x-x3)f[x4,x3,x2,x1]+…+(x-x1)(x-x2)…(x-x n-1 )f[x n ,x n-1 ,…x2,x1]+(x-x1)(x-x2)…(x-xn f[x n ,x n-1 ,…x1,x]=P(x)+R(x)

[0062] where:

[0063] P(x)=f(x1) + (x-x1)f[x2,x1] + (x-x1)(x-x2)f[x3,x2,x1]+(x-x1)(x-x2)(x-x3)f[x4,x3,x2,x1]+…+(x-x1)(x-x2)…(x-xn -1 f[x n ,x n-1 ,…,x2,x1]

[0064] R(x)=(x-x1)(x-x2)…(x-x n f[x n ,x n-1 ,…x1,x]

[0065] P(x) is the Newton interpolation approximation function, and R(x) is the error function. The approximate value of the missing value f(x) can be obtained by substituting the point x corresponding to the missing value into the interpolation polynomial.

[0066] Abnormal value monitoring and processing, sensor failure, operation error or external environmental disturbance may cause particularly obvious abnormal values, and the box plot (IQR, interquartile range) technology and 3σ principle (i.e. data values exceeding the average value ± 3 times the standard deviation range are considered abnormal) are used to identify abnormal values. Once the abnormal value is detected, the abnormal value is regarded as a missing value and the interpolation method is used for supplement.

[0067] 2) Data standardization. Data standardization is to eliminate the influence of different magnitudes and dimensions, so that the data is in the same order of magnitude, and is convenient for comprehensive analysis. The patent adopts Min-Max standardization to scale the data to a specified maximum and minimum value (usually 1 and 0) in proportion, and the formula is:

[0068]

[0069] 3) Feature engineering. The patent uses feature construction and multi-dimensional feature fusion method to extract the most useful information from the original data for model prediction. The patent performs STFT transformation on the data obtained by the vibration sensor, constructs frequency domain features, effectively reveals the running state of the mechanical equipment, and then fuses multiple time series signals containing device state features, such as multiple vibration features, through an autoencoder to extract more comprehensive fault signal features.

[0070] Through the above detailed data preprocessing steps, the quality of the data and the performance of the subsequent model can be greatly improved, laying a solid data foundation for accurate identification of the running state of the equipment. These processing procedures not only accurately process the raw data collected from the complex production environment, but also extract information that is crucial for anomaly identification through scientific methods, providing reliable support for subsequent identification of equipment anomaly patterns. At the same time, when training the cloud model later, uploading the preprocessed data to the cloud can reduce the amount of data transmission and save computing resources.

[0071] Step three: building of edge side model. In the edge side, the improved OPTICS algorithm is used to identify unknown data using labeled data. OPTICS will cluster the preprocessed data, and different types of data will be assigned to the same cluster. If there is a labeled data in the same cluster, the other unknown data in the cluster should also be of the same label type. In this way, the identification of unknown patterns is achieved. The improvement of the OPTICS algorithm mainly has the following four aspects:

[0072] 1) Improvement of local density calculation method

[0073] In the OPTICS algorithm, the density calculation of data points is often based on a fixed domain radius and a fixed number of neighbors, and the density is measured by counting the number of data points in the neighborhood. For example, when using a density calculation method based on domain radius, for data point x i , its density estimation is:

[0074]

[0075] where N ∈ (x i ) represents the domain with x i as the center and a radius of ∈, and χ(x j ) is an indicator function that takes the value 1 when x j is in the domain, otherwise 0.

[0076] The traditional local density calculation method is difficult to capture complex distributed data, which can lead to inaccurate division of different density regions. Therefore, the reciprocal of the average distance of the k-nearest neighbors of the data point is used to calculate the local density. Let the data point set be X = {x1, x2, …, x n}, and for data point x i , its k-nearest neighbor distance set is D k (x i ) = {d i1 , d i2 , …, d ik}, then the local density calculation formula is as follows:

[0077]

[0078] Compared with the traditional fixed neighborhood radius based method, the improved calculation method can better reflect the closeness of data points in their local area with surrounding data points, and can more flexibly capture the actual density of data points for data with uneven or complex distribution of data density. At the same time, since the method dynamically measures the density according to the near neighbor distance of the data itself, it avoids the problem of inaccurate division of different density regions caused by pre-setting a fixed neighborhood radius.

[0079] 2) Adaptive adjustment of density threshold

[0080] The traditional OPTICS algorithm usually needs to manually set a fixed value according to experience or preliminary observation of the data when determining the density threshold, which is difficult to grasp the appropriate threshold when facing complex and diverse data, and the manual setting method does not meet the requirements of intelligent manufacturing for automatic identification.

[0081] Based on the above reasons, the patent introduces an adaptive adjustment mechanism for the density threshold. First, calculate the mean ρ and standard deviation σ th of the density of all data points The formula of the density threshold ρ

[0082]

[0083] Through this adaptive method based on the statistical characteristics of the data density itself, the algorithm can dynamically determine a reasonable threshold according to the actual density distribution of the input data, which is used to distinguish different potential clusters. This makes the clustering algorithm better adapt to the density changes without too much human intervention, and more accurately identify clusters with different density characteristics.

[0084] 3) Cluster center determination based on density peak value

[0085] OPTICS algorithm does not explicitly determine the cluster center in a specific density peak value based way in the clustering process, but gradually reveals the clustering structure by constructing an ordered queue and subsequent analysis of reachable relationships, and the cluster center is relatively implicit in the structure formed by the entire clustering process.

[0086] Therefore, the patent explicitly takes the data points with local density greater than the determined density threshold ρ th as the cluster center. Let the data point index set be I = {1, 2,..., n}, then the index set C of the cluster center can be determined as follows:

[0087] C = {i ∈ I | ρ(x i )> ρth}

[0088] This way of determining cluster centers based on density peaks makes the algorithm more targeted to focus on those points in the data distribution that have relatively high density and are in the core position of the cluster. These cluster centers can provide a reasonable reference for subsequent cluster merging and other operations, helping to more clearly divide and integrate different clusters, especially when there are multiple density different but interrelated clustering structures in the data, it can better comb out the core area of each cluster

[0089] 4) Cluster merging operation

[0090] After clustering, the traditional OPTICS algorithm will generate clustering results according to the rules based on density reachability and ordered queue construction, but for some data distribution with fuzzy boundary, density gradient and other conditions, it may lead to generate too many relatively fine clusters, which may be should be merged as a whole in the actual data meaning, and the algorithm itself lacks effective automatic merging mechanism.

[0091] Based on the above problems, the cluster merging operation is introduced. First, the distance matrix D cc between cluster centers is calculated and (where c i ,c j ∈C), the distance is denoted as The distance matrix element is expressed as It can be obtained by calculating the Euclidean distance through the 'cdist' function in python. Then for any two cluster centers and (i≠j), if the following two conditions are met, the cluster merging operation is performed:

[0092]

[0093] The first condition indicates that the distance between the two cluster centers is relatively small, less than a certain proportion of the average near neighbor distance (here 0.1), which means they are relatively close in space position; The second condition indicates that the density difference between the two cluster centers is within a certain range (measured by a certain multiple of the density standard deviation), which means they are relatively similar in local density characteristics.

[0094] When the above conditions are met, the cluster merging is performed according to the density of the two cluster centers, if then the clustering label of the data points related to the cluster center (the near neighbor relationship index recorded by 'indices' is used to determine) is modified to the cluster center The same label, and vice versa. Through such cluster merging operations, the clustering results can be reasonably integrated, avoiding over-fining, so that the final clustering results are more consistent with the actual internal structure and distribution pattern of the data.

[0095] Through the above improvements, the adaptability of the algorithm in the intelligent manufacturing system equipment abnormal mode recognition task is enhanced, and the automation degree and accuracy of the algorithm are improved. The overall technical route of using the improved OPTICS algorithm to identify the equipment abnormal mode is as shown in Figure 2 .

[0096] Step four: deploy the transfer learning model in the cloud. The specific steps are as follows:

[0097] 1) Use the pre-trained network RestNeXt to extract features from the preprocessed data. The network structure information used is as shown in Figure 4 .

[0098] 2) Initialize the DANN network. The DANN network is as shown in Figure 3 , which is mainly composed of three parts: feature extractor (Feature Extractor), label predictor (Label Predictor), and domain discriminator (Domain Discriminator).

[0099] The feature extractor has the characteristic of inter-domain sharing. Let the input data be x, and the feature extractor be represented by function G f , and the extracted features be f=G f (x). Usually, the feature extractor is a neural network composed of multiple convolutional layers or fully connected layers. For example, for a simple fully neural network as a feature extractor, assuming that the input x is a d-dimensional vector, the first fully connected layer has n1 neurons, the weight matrix is W1, and the bias vector is b1, then the output of this layer h1=σ(W1x+b1), where σ is the activation function (for example, ReLU function σ(z)=max(0,z)). After multiple layers of such calculations, the final features f are obtained.

[0100] The label predictor is used to predict the label of the data according to the extracted features. Let the label predictor be G y , and the predicted label probability distribution be y=G y (f). For the complex manufacturing system equipment abnormal mode recognition task, we can regard it as a multi-classification problem, and use the softmax function to predict the label probability. Assuming that the dimension of the feature f is m, the number of classification categories is C, the weight matrix of the last layer of the label predictor is W y , and the bias vector is b y , then:

[0101] y = softmax(W y f + b y )

[0102]

[0103] z = W y f + b y

[0104] where each element y i of y represents the probability that the data belongs to the i-th class.

[0105] The role of the domain discriminator is to distinguish whether the feature is from the source domain or the target domain. Let the domain discriminator be G d , which takes the extracted feature f as input and outputs a scalar representing the probability that the input feature belongs to the source domain, d = G d (f). For example, for a simple fully connected layer as the domain discriminator, assuming the dimension of the input feature f is n, the weight matrix is Wd, and the bias vector is bd, the sigmoid function can be used to output the probability:

[0106] d = σ s (W d f + b d )

[0107] where:

[0108] DANN has label prediction loss, domain discrimination loss, and overall loss function when training.

[0109] For source domain data x s , the label y s is known, and cross-entropy loss is used to measure the accuracy of label prediction.

[0110] Let the label prediction loss be Ly, then

[0111]

[0112] where y s,i is the i-th element of the true label y s , G y (G f (x s )) i is the i-th element of the predicted label probability distribution G y (G f (x s )).

[0113] The goal of domain discrimination loss is to enable the domain discriminator to distinguish the source and target domains as much as possible, while allowing the features learned by the feature extractor to confuse the domain discriminator. Let the source domain data be represented as x. s The target domain data is represented as x t The probability that the domain discriminator outputs data from the source domain is:

[0114] d s =G d (G f (x s ))

[0115] The probability of outputting data in the target domain is

[0116] d t =G d (G f (x t ))

[0117] The performance of the domain discriminator is measured using binary cross-entropy loss, where the domain discrimination loss is:

[0118]

[0119] Where d x When x = x s When is d s When x = x t When is d t The ultimate overall training objective is the total loss function.

[0120] L = L y -λL d

[0121] Here, λ is a balancing parameter used to balance the importance of label prediction loss and domain discrimination loss. During training, this process is achieved through a Gradient Reversal Layer (GRL) to minimize the overall loss L.

[0122] 3) Train the model using a joint training method with contrastive loss. Let x1 and x2 be the representations of two samples in the feature space, then the distance between them is measured using Euclidean distance (where n is the feature dimension).

[0123]

[0124] Define the contrastive loss L. For a pair of samples (x1, x2) and their labels y (y = 1 indicates that the samples are similar, and y = 0 indicates that the samples are dissimilar), the contrastive loss formula is:

[0125]

[0126] In the calculation, the stable loss is obtained by averaging the contrast loss of the source domain data and the target domain data to realize the optimization of the model.

[0127] 4) Update the model parameters using back propagation.

[0128] Step five: use edge cloud cooperation to realize accurate identification of abnormal patterns of complex manufacturing systems, such as Figure 1 , the specific steps are as follows:

[0129] 1) The edge side sensor collects data and pre-processes the data.

[0130] 2) Upload the pre-processed data on the edge side to the cloud, and use the historical operating condition data stored in the cloud for joint training to identify the state of the operating condition data.

[0131] 3) Use the DANN-based transfer recognition architecture to identify the pattern of the operating condition data. First, use the ResNeXt pre-training network to further extract features from the historical operating condition data and the operating condition data. Then initialize the DANN network, use the extracted feature data, and use the contrast loss joint training method to train the network, and use the back propagation method to update the parameters.

[0132] 4) Use the trained DANN network to identify the pattern of the operating condition data and evaluate the identified results.

[0133] 5) After identifying enough labeled data, the labels can be issued to the edge side database, and the improved OPTICS clustering algorithm mentioned in step three is used for synchronous identification. Use the labeled data to identify and cluster other data in the entire cluster to improve the identification efficiency and accuracy, and efficiently use the computing resources.

Claims

1. An edge cloud collaborative device multi-condition abnormal mode identification method, comprising the following steps: Step one, complex manufacturing system device sensor deployment and multi-modal processing data collection: Step two, preprocessing of device time series data: cleaning, standardizing and feature engineering of the multi-modal processing data collected from the sensors on the edge side, the feature engineering being: using a feature construction and multi-dimensional feature fusion method, fusing multiple time series signals containing device state features through an autoencoder to extract fault signal features; Step three, building and updating of edge side model: on the edge side, using an improved OPTICS algorithm to cluster the preprocessed device time series data, using label data to identify other data in the entire cluster obtained by clustering to update the edge side database; wherein the improved OPTICS algorithm is: improving the local density calculation method, calculating the local density based on the reciprocal of the average distance of the k-nearest neighbors of the data points; adaptive adjustment of the density threshold, determining the density threshold according to the mean and standard deviation of the local density of the data points; cluster center determination based on density peaks, taking data points with local density greater than the density threshold as cluster centers; cluster merging operation, calculating the distance matrix between cluster centers, and performing cluster merging operation when the cluster center distance is less than the preset distance threshold and the local density difference is within a certain range; Step four, uploading the device time series data preprocessed on the edge side to the cloud, deploying a transfer learning model on the cloud, and using a pre-trained network RestNeXt to extract features from the preprocessed device time series data; Step five, using edge cloud collaboration to realize identification of abnormal patterns of complex manufacturing systems, the method being as follows: 1) Joint training using historical operating condition data stored in the cloud to identify the state of the operating condition data; 2) Using a DANN-based transfer identification architecture to identify the mode of the operating condition data: using a ResNeXt pre-trained network to further extract features from the historical operating condition data and the operating condition data; initializing the DANN network, using the further extracted feature data, and using a contrastive loss joint training method to train the DANN network, while updating the parameters using the backpropagation method; 4) Using the trained DANN network to identify the mode of the operating condition data and evaluating the identified results; 5) After sufficient label data is identified, the labels are issued to the edge side database for updating of the edge side database.

2. The edge cloud collaboration device multi-condition abnormal pattern recognition method according to claim 1, characterized in that, The method of step one is as follows: 1) According to the device structure and key part operation characteristics, arrange multiple types of sensors, including vibration sensors, temperature sensors, current sensors and sound sensors; vibration sensors are installed on mechanical moving parts; temperature sensors are fixed near heat generating components including motors and transformers; current sensors are installed on main power lines to monitor the current state of cable production equipment; sound sensors are used to identify mechanical failures by analyzing changes in sound spectrum; 2) Sensor data collection and integration: Collect data from each sensor, obtain multi-modal machining data, and forward the multi-modal machining data to the main data processing center on the edge side.

3. The edge cloud collaboration device multi-condition abnormal pattern recognition method according to claim 1, characterized in that, The method of step two is, 1) Data cleaning, including Missing value processing: Newton interpolation method is used; Abnormal value monitoring and processing: box plot technology and 3σ principle are used to identify abnormal values; Repeat data deletion; 2) Data standardization: Min-Max standardization is used; 3) Feature engineering.

4. The edge cloud collaboration device multi-condition abnormal pattern recognition method according to claim 1, characterized in that, In step three, the density threshold p th is calculated as follows: wherein is the mean and σ the standard deviation of the local density of data points ρ is the standard deviation.

5. The edge cloud collaboration device multi-condition abnormal pattern recognition method according to claim 1, characterized in that, In step three, the distance threshold is 0.1 of the average near-neighbor distance; the local density difference is within a certain range, that is, the local density difference is within the range of 0.1 of the standard deviation of the local density of the data points.

6. The edge cloud collaboration device multi-condition abnormal pattern recognition method according to claim 1, characterized in that, In step three, the method for performing cluster merging operation is: for two cluster centers, comparing the data point densities of the two cluster centers, merging the cluster center with smaller data point density into another cluster center, and modifying the cluster labels of the data points related to it to the same label as the other cluster center.

Images