Industrial big data analysis method and platform based on industrial internet

Abstract

The application discloses an industrial big data analysis method and platform based on an industrial internet, and relates to the technical field of big data analysis. The method comprises the following steps: collecting multi-source industrial data, using a one-dimensional CNN encoder for self-supervised pre-training, learning normal state time sequence features through a time sequence order discrimination task, generating synthetic fault samples by using a generative adversarial network, introducing fault types or working condition parameters to enhance fault data, adopting a two-stage PU learning strategy to identify high-confidence normal samples first, and then training a binary classification model to output fault probability, using Monte Carlo Dropout to quantify prediction uncertainty, formulating maintenance decisions based on double thresholds of fault probability and uncertainty, and deploying the model on the cloud and continuously optimizing it. The platform comprises self-supervised pre-training, fault sample generation, PU learning fault classification, decision-making and deployment modules. The application improves fault detection accuracy and reliability, adapts to new fault modes, and realizes efficient predictive maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, and more particularly, to an industrial big data analysis method and platform based on an industrial internet. BACKGROUND

[0002] In the field of industrial internet and predictive maintenance, fault detection methods based on big data have become a research hotspot. However, the existing technical solutions still face significant challenges in actual deployment. First, industrial field fault samples are usually extremely scarce and uneven in type, which leads to overfitting of models based on supervised learning and poor generalization ability. Second, obtaining a large amount of high-quality and fully labeled training data is costly, and data under normal operation of the equipment is easy to obtain but lacks effective labels and is difficult to be fully utilized. In addition, most methods only output fault probability, lack of evaluation of model prediction confidence, and thus have high decision-making risk and are prone to false positives or false negatives; in view of the above problems, the present application provides a solution. SUMMARY

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an industrial big data analysis method and platform based on an industrial internet to solve the problems proposed in the background art.

[0004] To achieve the above object, the present application provides the following technical scheme:

[0005] The industrial big data analysis method based on the industrial internet comprises the following steps:

[0006] Step S1, collecting multi-source industrial data, using a one-dimensional CNN encoder for self-supervised pre-training to learn time sequence features under normal state, training the encoder through a time sequence order discrimination task to extract general feature representation;

[0007] Step S2, using a generative adversarial network to generate synthetic fault samples, introducing fault types or working condition parameters, merging the synthetic samples with real fault samples, and screening reasonable samples;

[0008] Step S3, using the pre-trained encoder and the enhanced fault data, adopting a two-stage PU learning strategy, identifying high-confidence normal samples first, then training a classifier, training a binary classification model, and outputting fault probability;

[0009] Step S4, using Monte Carlo t quantification to predict uncertainty, formulating maintenance decisions based on fault probability and uncertainty double thresholds, deploying the model on the cloud, and continuously optimizing the model and adapting to new fault modes.

[0010] In a preferred embodiment, step S1 comprises the following contents:

[0011] The collection of multi-source industrial data includes real-time data collected by various sensors installed on equipment, the transmission of sensor data to the platform through industrial gateways, and the division of the cleaned and aligned time series data into several sample segments according to the input requirements of the predictive maintenance model. The time series data is divided into several sample segments using a sliding time window method.

[0012] The self-supervised pre-training uses a one-dimensional convolutional neural network as a feature extraction encoder, which includes multiple convolutional layers, each followed by a ReLU activation and pooling. The encoder is trained through a temporal order discrimination task, which includes dividing each unlabeled sample window into two sub-sequences, constructing positive and negative samples for the anchor segment, outputting a discrimination probability using the encoder and temporal relationship discrimination head, and training through a binary cross-entropy loss function.

[0013] In a preferred embodiment, step S2 includes the following:

[0014] The use of a generative adversarial network to generate synthetic fault samples includes a GAN model architecture composed of a generator and a discriminator. The generator accepts a random noise vector as input and outputs a synthetic signal. The discriminator is a binary classification network that inputs real or synthetic signals and outputs a discrimination probability. The generator and discriminator are trained alternately using a binary cross-entropy loss function.

[0015] The generative adversarial network introduces fault types or operating condition parameters as conditional information to make the generator produce specific fault patterns. After merging the synthetic samples with the real fault samples, unreasonable samples are removed through clustering or screening to form an enhanced fault sample set.

[0016] In a preferred embodiment, step S3 includes the following:

[0017] The two-stage PU learning strategy includes:

[0018] In the first stage, the augmented fault sample set and all unlabeled samples are used to train a temporary classifier to identify high-confidence normal samples as a reliable negative sample set from the unlabeled samples.

[0019] In the second stage, the fault sample set and the reliable negative sample set are used to train the final fault classification model. The encoder obtained from the self-supervised pre-training is used for feature extraction, and a fully connected network is added to form a classifier. The fault classification model outputs a fault probability through a Sigmoid activation.

[0020] Identifying high-confidence normal samples includes inferring the fault probability of all unlabeled samples through a temporary classifier, selecting samples with a predicted probability below a pre-set threshold as a reliable negative sample set, and training the final fault classification model using a binary cross-entropy loss function.

[0021] In a preferred embodiment, step S4 comprises the following:

[0022] Quantifying the prediction uncertainty using Monte Carlo Dropout includes keeping the Dropout layer in the neural network open, repeatedly performing model prediction multiple times for the same input, calculating the mean and standard deviation of the prediction results, the mean as the failure probability estimate, and the standard deviation as the prediction uncertainty;

[0023] Making maintenance decisions based on double thresholds of failure probability and uncertainty includes setting failure probability threshold and uncertainty threshold, classifying prediction results into different levels according to the combination of failure probability estimate and prediction uncertainty, and triggering corresponding maintenance actions, while the model is deployed on the cloud or edge node, continuously optimizing the model and adapting to new failure modes through feedback loops.

[0024] The industrial big data analysis platform based on industrial internet includes a self-supervised pre-training module, a failure sample generation module, a PU learning failure classification module, and a decision and deployment module, and the modules are connected by signals;

[0025] The self-supervised pre-training module: collects multi-source industrial data and uses a one-dimensional CNN encoder for self-supervised pre-training to learn time series features in normal state, trains the encoder through a time sequence discrimination task, and extracts general feature representation;

[0026] The failure sample generation module: uses a generative adversarial network to generate synthetic failure samples, introduces failure types or working condition parameters, merges synthetic samples with real failure samples, and selects reasonable samples;

[0027] The PU learning failure classification module: uses the pre-trained encoder and enhanced failure data, adopts a two-stage PU learning strategy, first identifies high-confidence normal samples, then trains the classifier, trains a binary classification model, and outputs the failure probability;

[0028] The decision and deployment module: uses Monte Carlo to quantify the prediction uncertainty, makes maintenance decisions based on double thresholds of failure probability and uncertainty, and deploys the model on the cloud to continuously optimize the model and adapt to new failure modes.

[0029] The technical effects and advantages of the industrial big data analysis method based on the industrial internet are as follows: normal operation features of equipment are automatically learned from a large amount of unlabeled data through self-supervised pre-training, and the dependence on artificial labeled data is significantly reduced; a generative adversarial network is used to synthesize diversified fault samples, effectively solving the problems of scarcity and class imbalance of fault samples in an industrial scene; a two-stage PU learning strategy is combined, reliable normal samples in unlabeled data are fully utilized, and the generalization ability and robustness of the fault classification model are improved; Monte Carlo Dropout is further introduced to quantify the prediction uncertainty, and a decision is made based on the double thresholds of fault probability and uncertainty, so that the maintenance warning is more reliable, false positives are reduced, and false negatives are avoided, thereby providing a complete and feasible technical solution for realizing precise and adaptive industrial predictive maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A structure diagram of the industrial big data analysis method based on the industrial internet is shown.

[0031] Figure 2 A module diagram of the industrial big data analysis system based on the industrial internet is shown. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0033] Embodiment: Please refer to Figure 1 The industrial big data analysis method based on the industrial internet is disclosed, including the following steps:

[0034] Step S1, collect multi-source industrial data, use a one-dimensional CNN encoder for self-supervised pre-training, learn time sequence features in a normal state, train the encoder through a time sequence order discrimination task, and extract general feature representation;

[0035] Step S2, use a generative adversarial network to generate synthetic fault samples, introduce fault types or working condition parameters, merge the synthetic samples with real fault samples, and screen reasonable samples;

[0036] Step S3, use the pre-trained encoder and the enhanced fault data, adopt a two-stage PU learning strategy, first identify high-confidence normal samples, then train a classifier, train a binary classification model, and output a fault probability;

[0037] Step S4: Use Monte Carlo t-quantization to predict uncertainty, and make maintenance decisions based on the dual thresholds of failure probability and uncertainty. Deploy the model in the cloud and continuously optimize the model to adapt to new failure modes.

[0038] In step S1, multi-source industrial data is collected and a one-dimensional CNN encoder is used for self-supervised pre-training to learn the temporal features under normal conditions. The encoder is then trained through a temporal order discrimination task to extract general feature representations. Specific details include:

[0039] Data collected in real time by various sensors installed on the equipment, such as vibration acceleration, temperature, pressure, current, voltage, and rotational speed, is transmitted to the platform via an industrial gateway or PLC. Different sensors have different sampling frequencies; for example, the vibration signal sampling frequency... It can be set to 5-20kHz to capture high-frequency characteristics of mechanical faults. The sampling rate for slow variables such as temperature and voltage can be 1-10Hz. The selection of the sampling frequency must ensure that the target fault characteristics are not distorted. An appropriate value should be selected according to the equipment speed and the frequency of fault impact characteristics.

[0040] Operating condition and control data: Equipment operating condition parameters and control signals from the industrial control system, such as equipment load, speed settings, valve opening, production formula, and log events, are uploaded in real time via fieldbus or Ethernet. Operating condition data is typically recorded at second or minute intervals to provide environmental and operational context.

[0041] Maintenance and fault records come from maintenance work orders, maintenance records, and fault logs in the maintenance management system. They include the time of each equipment failure, a description of the failure type, the parts replaced, and the content and date of regular maintenance. Maintenance records are generally entered manually and are not continuous in time but contain important tag information to mark the time period and failure category of failures in historical data.

[0042] Based on the input requirements of the predictive maintenance model, the cleaned and aligned time-series data is divided into several sample segments, and features are extracted from them. A sliding time window method is preferred for segmenting the time-series data: window length... The window length is selected based on the equipment's operating cycle and fault evolution characteristics. For example, a 1-second window can be selected for high-speed rotating machinery, while a 1-minute window can be selected for slow-process equipment. A window length that is too short may not be able to cover fault symptoms, while a window length that is too long will increase the computational load and incorporate too many state changes. Empirically, It should cover at least one equipment operation cycle or fault characteristic occurrence cycle. The time series should be... Divide the length into segments, and set the step size of the sliding window. Take half the window length, i.e., 50% overlap sampling, for example, a 1-second window generates a new sample every 0.5 seconds, in order to balance the number of samples and the correlation between adjacent samples;

[0043] Self-supervised learning is performed on unlabeled mass normal operation data, and an encoding model capable of extracting general features is trained to improve the effect of fault detection. The goal of self-supervised pre-training is to enable the model to learn the internal structure and time relationship of the data in the normal state of the device without any manual labels. The specific implementation is as follows:

[0044] A one-dimensional convolutional neural network is used as a feature extraction encoder , which includes, for example, 3 convolutional layers, each followed by a ReLU activation and pooling, and the convolution kernel size is, for example, 16, 32, 64, and the kernel width is, for example, 5, to extract time sequence features at different levels of abstraction;

[0045] The output of the encoder is a feature vector of length d, such as d=128, representing a compact representation of the input signal, and the specific structure and parameters of the network can be adjusted according to the complexity of the data, such as increasing the depth of the convolutional layer or using a larger kernel to capture long-term dependencies, and when the data has a periodic pattern, a layer of LSTM can also be introduced to remember long sequence features;

[0046] According to the characteristics of industrial time series data, the application preferably selects a self-supervised task of time sequence discrimination, specifically: each unlabeled sample window is further divided into two sub-sequences, each of which can be For a given anchor segment , a positive sample is constructed as the second half of the anchor segment, and a negative sample is constructed as a random segment from a different time period;

[0047] The anchor segment, positive segment and negative segment are input into an encoder with shared parameters to obtain corresponding feature representations , , , then a time sequence relationship discrimination head is introduced, which accepts or as input and outputs a discrimination probability p for identifying whether the two segments are continuous in time;

[0048] The goal of self-supervised training is to enable the model to discriminate between positive sample pairs and negative sample pairs: that is, to discriminate between positive pairs from consecutive segments of the same sequence as 1, and to discriminate between unrelated negative pairs as 0. For this purpose, a binary cross-entropy loss function can be used for training to force the output of to approach 1, and the output of to approach 0. Through this form of contrastive learning, the encoder is forced to extract features that can represent the time sequence continuity and dynamic patterns;

[0049] In self-supervised pre-training, a large number of unlabeled sample pairs are fed into the model training in a certain batch. The batch size is set to 256, for example, and the Adam optimizer is used to update the parameters iteratively with a learning rate of 0.001 . The number of training rounds can be set according to the convergence condition, for example, the training is stopped when it is observed that the self-supervised task loss no longer decreases on the validation set;

[0050] Typically, the validation loss is calculated after 100 epochs of training, and if the change is less than 1e-4, it is considered to be converged. The evaluation indicator of the self-supervised task can be the discrimination accuracy or the contrast loss value. After training, the parameters of the encoder will be learned to represent the normal mode well without labels;

[0051] After completing the self-supervised pre-training, we obtain a trained feature extraction network, i.e., the encoder , whose output feature representation can sufficiently reflect the running state features of the input time series data. At this time, the model has not yet utilized any fault label information, but has established a baseline behavior model on normal data.

[0052] In step S2, a generative adversarial network is used to generate synthetic fault samples, introduce fault types or working condition parameters, merge the synthetic samples with the real fault samples, and select reasonable samples. The specific content includes:

[0053] A generative model is introduced to expand the fault data set, and a generative adversarial network (GAN) is used to generate simulated fault signals, thereby increasing the number of fault samples for training;

[0054] GAN model architecture: GAN consists of two neural networks, generator G and discriminator D. The generator G accepts a random noise vector z with a dimension of, for example as input, and outputs a synthetic signal with the same length as the real fault signal;

[0055] The discriminator D is a binary classification network that inputs a signal, real x or synthetic signal , and outputs the probability of discriminating whether it is real fault data. The generator can use a deconvolutional network structure similar to the encoder to gradually upsample the low-dimensional noise into a time series;

[0056] The discriminator can use a convolutional network to extract features from the input signal and output true or false discrimination through full connection. To improve the generation quality, conditional information can be introduced in GAN: for example, conditional GAN takes fault types or certain working condition parameters as additional input, so that the generator learns to generate fault patterns of a specific category. If the actual scene fault pattern is single or the sample is extremely small, a standard GAN can also be used to generate general fault samples;

[0057] Training strategy: GAN training adopts the way of generator and discriminator game. The loss function includes the discrimination loss of the discriminator and the adversarial loss of the generator , the discriminator tries to judge the real fault signal as true and the synthetic signal as false as possible;

[0058] The generator tries to deceive the discriminator to make it output the signal as true. The specific loss can adopt the classical binary cross-entropy form: , During training, D and G are updated alternately, for example, first train the discriminator for one step using a batch of real and current generated samples, and then train the generator for one step;

[0059] In order to stabilize the training, improved techniques such as WGAN-GP gradient penalty or gradually increasing the difficulty of generation can be used to avoid mode collapse. The number of epochs needs to be relatively large to ensure that the generator learns the fault distribution sufficiently, for example, it can be trained for more than 1000 epochs, or until the classification accuracy of the discriminator on the validation set approaches 50%, indicating that the generated samples are difficult to distinguish from the real samples;

[0060] After training, use the generator G to generate the required number of synthetic fault samples. The choice of synthetic number can make the total number of fault samples and normal samples relatively balanced or at least the same order of magnitude, for example, if there are only real fault samples, and there are thousands of unlabeled normal samples, you can generate synthetic fault signals, which expand the total number of fault samples by 5 times to 120, the multiple can be adjusted according to needs, but should not be too large, otherwise the synthetic data will occupy too high a proportion and introduce bias.

[0061] It is preferred that the generated fault samples cover the characteristic range of the real fault in the statistical distribution, and a certain diversity can be introduced, for example, multiple samples can be generated by sampling the noise input z multiple times, and then typical and high-quality synthetic samples can be selected through clustering or screening;

[0062] The degree of agreement of the amplitude range and spectral shape of the synthetic samples with the real samples, for example, the synthetic samples that produce signals exceeding the sensor range or have normal operation characteristics in the spectrum are removed;

[0063] The final enhanced fault sample set is merged with the original real fault samples , which are labeled as fault class and used for subsequent model training;

[0064] This generative data augmentation significantly alleviates the problem of imbalance between positive and negative samples and can effectively reduce classification decision bias during model training. Existing cases have shown that data generated by GANs can significantly improve the performance of fault diagnosis models.

[0065] In step S3, using the pre-trained encoder and the enhanced fault data, a two-stage PU learning strategy is employed. First, high-confidence normal samples are identified, then the classifier is trained to train a binary classification model, which outputs the fault probability. Specific details include:

[0066] The fault classification model training based on PU learning utilizes the pre-trained model and augmented fault data to train a fault detection model for predictive maintenance. Since normal samples are not explicitly labeled and may contain unobserved potential faults, this step employs the PU learning strategy to fully utilize unlabeled data and train a robust fault classification model.

[0067] The fault classification model is essentially a binary classification model, outputting the probability that a given input sequence belongs to a fault or a normal state. The optimal approach is to use the self-supervised pre-trained encoder. Transferring the feature extraction layer used in the classification model;

[0068] When training a classification model, you can choose to freeze the encoder. The parameters are kept unchanged to prevent overfitting in small-sample supervised training; alternatively, they can be fine-tuned to further adapt to the fault identification task. After the feature z output by the encoder, one or more fully connected networks are added to form a classifier. Ultimately, the failure probability is output through Sigmoid activation. For example, a hidden layer can be added after the 128-dimensional features before the 1-dimensional output, with the initial classification threshold set at 0.5. Determined as a malfunction. It is judged as normal, but the threshold can be adjusted according to actual needs in the case of unbalanced data.

[0069] PU learning two-stage strategy: Due to the lack of clearly labeled normal samples, directly treating all unlabeled data as normal to train the classifier may lead to bias. Therefore, this embodiment adopts the classic two-stage PU learning algorithm:

[0070] A set of highly reliable normal samples is selected from the unlabeled sample set U as a temporary negative sample set. The selection criteria can be achieved by training a preliminary classifier or using statistical methods. We can initially assume all unlabeled samples are normal, and then train a temporary classifier using the expanded fault sample set P and a large number of unlabeled samples U. Training can use standard supervised binary classification methods, but attention should be paid to the problem of severe imbalance between positive and negative samples.

[0071] To address this, the weights of positive samples can be increased or unlabeled samples can be downsampled to balance the proportions. For example, an equal number of unlabeled samples as positive samples can be randomly selected for training in each iteration, or a weighting factor can be assigned to positive samples in the loss function. ,like This is approximately the ratio of unlabeled to positive examples to balance the impact;

[0072] After several epochs of training, we obtain ,use To infer the failure probability of all unlabeled samples U, select the... Samples deemed normal and with extremely high confidence levels were added. Specifically, it can be based on the predicted probability. Sort unlabeled samples from low to high, and take the portion closest to 0, for example, select... The samples are considered reliable and normal, i.e., the failure probability is less than 10%, or the top N samples with the lowest predicted probability are selected as reliable and normal samples. The threshold of 0.1 is chosen conservatively to ensure that the selected samples are almost certainly not faulty, thus guaranteeing... Purity;

[0073] We obtain a batch of negative class samples with high confidence, and their number is denoted as . This number may be much smaller than the total number of unlabeled samples, but it is sufficient for the next step of training.

[0074] The positive sample set P includes real and generated faults, as well as identified reliable negative samples. The data is used as training data to train the final fault classification model. This can release the encoder. The model is frozen and then jointly trained to fine-tune the feature extractor for the fault detection task. Although the number of positive and negative samples remains unequal during training... Perhaps not as good as Compared to treating all unlabeled samples as negative, this strategy reduces the negative contamination rate, decreases classification bias, and improves model stability.

[0075] The loss function uses binary cross-entropy. ;in, For real labels, fault = 1, normal = 0. To predict probabilities for the model, the Adam optimizer is used, and the learning rate can be set. Training, for example, for 50 epochs or until the accuracy, recall and other metrics on the validation set no longer improve;

[0076] After the model training is completed, the performance is evaluated on the validation dataset, and if it meets the expectations, it can enter the deployment stage; if the performance is insufficient, the number of generated samples can be adjusted, the reliable negative sample threshold can be reselected, or other unlabeled sample utilization strategies can be introduced;

[0077] The final model trained by PU learning makes full use of all positive examples and unlabeled data. Unlike traditional supervised learning, which only uses P and manually labeled negative samples, the PU strategy reduces the risk of mislabeling normal data and improves the model's generalization ability with the help of unlabeled data information;

[0078] This model can output the probability of a given piece of equipment data belonging to a fault state It should be noted that if in the application it can be certain that some normal data is not faulty, it can also be directly labeled as negative samples to participate in training, further simplifying the PU learning process.

[0079] In step S4, the Monte Carlo t quantization prediction uncertainty is used, the maintenance decision is made based on the fault probability and uncertainty double threshold, the model is deployed in the cloud, the model is continuously optimized and adapted to new fault modes, and the specific content includes:

[0080] The Monte Carlo Dropout method is used to quantify the model prediction uncertainty, and the Dropout layer in the neural network is kept open, that is, even in prediction, some neurons are randomly inactivated according to the dropout rate set during training. For the same input data, the model prediction is repeated n times, for example, n = 100, and each time a slightly different output result is obtained due to random inactivation , then the mean and standard deviation of these repeated prediction results are calculated, the mean can be used as the final fault probability estimate, and the standard deviation reflects the uncertainty of the model prediction: if is large, it means that the model output is unstable for this input, i.e., the prediction lacks confidence; on the contrary small indicates that the model is relatively consistent and confident in the prediction. The dropout rate is generally consistent with the training, such as the dropout setting of 0.5 during model training, which is followed during inference. Similarly, the number of repeated predictions n is selected to balance the accuracy and overhead, such as 100 times to obtain stable mean and variance estimates, and the computational time is within an acceptable range. For applications with higher real-time requirements, n can be reduced or a model ensemble can be used to evaluate uncertainty, i.e. training 5 models with the same structure but different initial random seeds. For the same input, take 5 prediction values to calculate the variance. This ensemble method is equivalent to Dropout sampling, which can also provide uncertainty measurement;

[0081] Failure probability based on model output and its uncertainty , the predictive maintenance decision rule is made using a two-threshold judgment, i.e. setting failure probability threshold and uncertainty threshold Two parameters, the prediction results are divided into different levels:

[0082] When significantly higher than the failure decision threshold and at the same time the uncertainty is lower than the confidence threshold, it is determined as a high-confidence failure alarm, for example , , if the model predicts that the current failure probability of a device , then and , it can be considered that the model is very confident in warning of the occurrence of failure, and a formal alarm is generated for this result, which informs the operator to take maintenance action through the industrial internet platform. Because of the high confidence, maintenance can be directly dispatched to avoid downtime caused by potential failure;

[0083] When is higher than but is higher than , it is determined as a suspicious failure. That is, the model tends to predict failure but lacks confidence, for example and , it will not directly trigger emergency shutdown and other operations, but mark this state as an alarm that needs attention. On the platform, it can prompt the operation and maintenance personnel that the device has abnormal signs but the model is uncertain, and further observation or manual inspection and verification is needed. The data monitoring period of the device can be shortened to obtain more information, or on-site personnel can be arranged to check whether there is a potential failure, through this two-stage processing, it avoids rash action when the model is uncertain, and reduces unnecessary downtime and maintenance;

[0084] When is lower than and is lower than , i.e. the model is confident that it is normal, it is determined that the device is running normally, for example , , then the system continues to monitor normally and no measures need to be taken, and only the result is recorded for trend tracking;

[0085] When is lower than but is higher than , it is determined as an unknown state. That is, the model roughly judges no failure but lacks confidence, such as , ; such cases can correspond to new working conditions or abnormal noises that the model has not seen before. The system can take a warning strategy, such as continuing to monitor at a higher frequency, and if high uncertainty persists for multiple times, it can prompt the need to check the sensor status or model applicability. Such a state does not trigger maintenance, but is archived as a potential risk signal;

[0086] The above decision logic only uses the combination of two parameters and , and the two thresholds , can be adjusted according to specific applications: for example, for high reliability requirements, the can be reduced to 0.4 to increase sensitivity, and a lower can be set to only alert at high confidence; for scenarios where false positives are not allowed, the can be increased or a gray area can be added to handle the model's prediction results in a hierarchical manner, improving the credibility and interpretability of the system's automatic decision-making;

[0087] The fault detection model can be deployed in a containerized form to the platform's cloud or edge nodes. When choosing the deployment location, the real-time requirements and network bandwidth can be considered. For devices with large data volumes and real-time response requirements, the model can be deployed on the edge side close to the device, such as on an industrial gateway or local server, to perform data collection and prediction at millisecond intervals, reducing cloud transmission delay;

[0088] For models that do not require strict real-time and have high computational complexity, they can be deployed in the cloud for centralized processing of data streams from multiple devices. Regardless of the cloud or edge, the deployed model instances need to be integrated with the data access module to subscribe to the data topic of the corresponding device sensors and implement online inference, such as invoking the deployed model interface to calculate the fault probability and uncertainty for each new 1-second vibration signal window buffered by the edge gateway;

[0089] The model inference itself is mostly matrix multiplication operations due to optimization, which can fully utilize GPU or vectorization acceleration. Processing a window prediction on a typical industrial PC only takes a few tens of milliseconds, fully meeting the detection requirements of seconds or even sub-seconds;

[0090] The platform should provide a unified interface to deliver the model output results to the upper-layer application modules. For example, by publishing the fault probability and confidence to the specified topic through MQTT, the alarm service can subscribe and process it. During deployment, model version management and configuration parameter consistency should be considered. The best parameters verified during the training phase should be used, and decision thresholds such as , should be set according to the needs of the field and managed uniformly in the platform configuration file, making it easy to adjust in the future without rewriting the code;

[0091] When the model determines that an alarm needs to be triggered, for example, a high-confidence fault condition is met, the application service of the platform generates an alarm event, including device identification, time, fault type or location, fault probability, confidence, and other information. This alarm can be notified to relevant personnel and systems in various ways, such as highlighted display on the monitoring dashboard interface of the operation and maintenance personnel, sent to the on-duty engineer through SMS or email, automatically generated in the maintenance management system as a to-be-processed work order, etc.

[0092] When the model issues a fault alarm, the device may be immediately repaired by manual or automatic means, regardless of whether an actual fault is ultimately found. The results should be recorded and fed back, such as if a component fault is found after inspection, the alarm is confirmed as a true positive, and the maintenance personnel record the fault details in the system and associate them with the corresponding sensor data window at the time, forming a new labeled fault sample. Conversely, if no abnormalities are found after inspection, it may be a false alarm, and the data window should be marked as normal and archived for model improvement. In the context of PU learning, this normal feedback is very valuable because it gives some previously unlabeled data a clear negative label, which can be used to expand the reliable negative sample set in the next round of training;

[0093] Some devices will be regularly repaired and replaced with components according to a preventive maintenance plan. If a potential fault is found during a preventive maintenance, the related data should be labeled as a potential fault or degradation state, even if the model did not issue an alarm. This type of data can be used for early fault pattern learning of the model;

[0094] Over time, industrial equipment may develop new fault patterns, or changes in operating conditions may cause the data distribution to drift. To ensure that the model remains effective, a model update process can be triggered periodically when a certain number of new labeled data is collected, such as when the cumulative number of new fault samples reaches 10 or when 6 months have passed;

[0095] During retraining, the latest complete data set is used, including old and newly accumulated labeled fault samples, original unlabeled and newly collected unlabeled data. The model is retrained in steps S1-S3. Due to the increase in the data set and the increase in the number of fault patterns covered, the performance of the model is expected to be further improved. After the model is updated, it is confirmed to be better than the old model through A / B testing or offline verification, and then seamlessly replaces the deployed model version, allowing the model to adapt to equipment aging and environmental changes;

[0096] During continuous learning, the complexity and overfitting risk of the model should also be monitored, and if necessary, regularization or model structure simplification can be used, such as using a smaller network to distill a student model for deployment based on pre-training, to ensure that the new model is stable and reliable.

[0097] Referring to Figure 2 As shown in the drawings, the application discloses an industrial big data analysis platform based on an industrial internet, comprising a self-supervised pre-training module, a fault sample generation module, a PU learning fault classification module and a decision and deployment module, and the modules are signal connected;

[0098] The self-supervised pre-training module: collects multi-source industrial data, uses a one-dimensional CNN encoder for self-supervised pre-training, learns time sequence features in a normal state, trains the encoder through a time sequence order discrimination task, and extracts general feature representation.

[0099] The fault sample generation module: uses a generative adversarial network to generate synthetic fault samples, introduces fault types or working condition parameters, merges the synthetic samples and real fault samples, and screens reasonable samples.

[0100] The PU learning fault classification module: uses the pre-trained encoder and enhanced fault data, adopts a two-stage PU learning strategy, first identifies high-confidence normal samples, then trains a classifier, trains a binary classification model, and outputs a fault probability.

[0101] The decision and deployment module: uses Monte Carlo quantification to predict uncertainty, formulates a maintenance decision based on a fault probability and uncertainty double threshold, deploys the model on a cloud, and continuously optimizes the model and adapts to new fault modes.

[0102] The above formulas are all dimensionless values, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the nearest real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0103] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized in the form of a computer program product wholly or partially.

[0104] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application of the technical solution and the constraints of the application. A person skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0105] In addition, the functional modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically independently, or two or more modules can be integrated in one module.

[0106] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0107] Finally, the above merely provides the preferred embodiments of the present application, but is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An industrial big data analysis method based on an industrial internet, characterized by, Comprising the steps of; Step S1, collecting multi-source industrial data using a one-dimensional CNN encoder for self-supervised pre-training, learning time series features under normal conditions, training the encoder through a time sequence discrimination task, and extracting general feature representations; Learning The timing features in the normal state specifically include: dividing each unlabeled sample window into two sub-sequences in front and back, each with a length of For a given anchor segment , a positive sample is constructed for the first half of the segment, and a negative sample is constructed for the second half of the segment that immediately follows in time; and Input the anchor segment, positive segment and negative segment into the encoder of shared parameters respectively , to obtain corresponding feature representations , , , introduce a time sequence relationship discrimination head , accept or as input, and output a discrimination probability p for identifying whether the two segments are continuous in time; The goal of self-supervised training is to enable the model to distinguish between positive and negative sample pairs: i.e. to assign a score of 1 to positive pairs from consecutive segments of the same sequence and a score of 0 to negative pairs of unrelated sequences, using a binary cross-entropy loss function to train the model to approximate to 1 and to 0. Step S2, generating synthetic fault samples using a generative adversarial network, introducing fault types or working condition parameters, merging synthetic samples with real fault samples, and screening reasonable samples; Step S3, using the pre-trained encoder and enhanced fault data, adopting a two-stage PU learning strategy, identifying high-confidence normal samples first, then training the classifier, training a binary classification model, and outputting fault probability; Step S4, using Monte Carlo t quantization to predict uncertainty, making maintenance decisions based on fault probability and uncertainty double thresholds, deploying the model on the cloud, and continuously optimizing the model and adapting to new fault modes. 2.The industrial internet-based big data analysis method of claim 1, wherein, Collecting multi-source industrial data includes real-time data collected by various sensors installed on equipment, transmitting sensor data to the platform through industrial gateways, and dividing the cleaned and aligned time series data into several sample segments according to the input requirements of the predictive maintenance model, and using a sliding time window to cut the time series data. 3.The industrial internet-based big data analysis method of claim 2, wherein, Self-supervised pre-training uses a one-dimensional convolutional neural network as a feature extraction encoder, which contains multiple convolutional layers, each followed by a ReLU activation and pooling, and trains the encoder through a time sequence discrimination task, which includes dividing each unlabeled sample window into two sub-sequences, constructing positive and negative samples for the anchor segment, using the encoder and time sequence relationship discriminator to output discrimination probability, and training through a binary cross-entropy loss function. 4.The industrial internet-based big data analysis method of claim 1, wherein, Generating synthetic fault samples using a generative adversarial network includes using a GAN model architecture composed of a generator and a discriminator, the generator accepts random noise vectors as input and outputs synthetic signals, the discriminator is a binary classification network that inputs real or synthetic signals and outputs discrimination probability, and the generator and discriminator are trained alternately, using a binary cross-entropy loss function. 5.The industrial internet-based big data analysis method of claim 4, wherein, The generative adversarial network introduces fault types or working condition parameters as conditional information, allowing the generator to produce different categories of fault modes, and after merging the synthetic samples with the real fault samples, removes unreasonable samples through clustering or screening to form an enhanced fault sample set. 6.The industrial internet-based big data analysis method according to claim 1, wherein, Two-stage PU learning strategy includes: The first stage, using the expanded fault sample set and all unlabeled samples to train a temporary classifier, identifying high-confidence normal samples from unlabeled samples as a reliable negative sample set; The second stage, using the fault sample set and the reliable negative sample set to train the final fault classification model, where the encoder obtained from self-supervised pre-training is transferred for feature extraction, a fully connected network is added to form a classifier, and a Sigmoid activation is used to output fault probability. 7.The industrial internet-based big data analysis method of claim 6, wherein, Identifying high-confidence normal samples includes using a temporary classifier to infer the fault probability of all unlabeled samples, selecting samples with predicted probability below a preset threshold as a reliable negative sample set, and using a binary cross-entropy loss function to train the final fault classification model. 8.The industrial internet-based big data analysis method of claim 1, wherein, The Monte Carlo Dropout quantifies the prediction uncertainty, which includes keeping the Dropout layer in the neural network open, repeating the model prediction multiple times for the same input, calculating the mean and standard deviation of the prediction results, the mean as the failure probability estimate, and the standard deviation as the prediction uncertainty. 9.The industrial internet-based big data analysis method of claim 8, wherein, The maintenance decision is made based on the double thresholds of failure probability and uncertainty, which includes setting the failure probability threshold and the uncertainty threshold, dividing the prediction results into different levels according to the combination of the failure probability estimate and the prediction uncertainty, and triggering the corresponding maintenance action. At the same time, the model is deployed on the cloud or edge node, and the model is continuously optimized and adapted to new failure modes through a feedback loop.

10. An industrial big data analysis platform based on industrial internet, used to implement the industrial big data analysis method based on industrial internet in any one of claims 1-9, characterized in that ； Self-supervised pre-training module: Collect multi-source industrial data and use one-dimensional CNN encoder for self-supervised pre-training to learn time series features in normal state. The encoder is trained through a time sequence discrimination task to extract general feature representation. Learning The timing features in the normal state specifically include: dividing each unlabeled sample window into two sub-sequences in front and back, each with a length of For a given anchor segment , a positive sample is constructed for the first half of the segment, and a negative sample is constructed for the second half of the segment that immediately follows in time; and random segments from different time periods. Inputting the anchor segment, the positive segment and the negative segment into an encoder of shared parameters respectively , to obtain corresponding feature representations , , , introducing a time relationship discrimination head , accepting or as input, and outputting a discrimination probability p for identifying whether the two segments are continuous in time; The goal of self-supervised training is to enable the model to distinguish between positive and negative pairs of samples: i.e. to assign a value of 1 to positive pairs from consecutive segments of the same sequence and a value of 0 to negative pairs of unrelated samples, using a binary cross-entropy loss function to train the model to approximate to 1 and to 0; Fault sample generation module: Use a generative adversarial network to generate synthetic fault samples, introduce fault types or working condition parameters, merge synthetic samples with real fault samples, and select reasonable samples. PU learning fault classification module: Use the pre-trained encoder and enhanced fault data, adopt a two-stage PU learning strategy, first identify high-confidence normal samples, then train the classifier, train a binary classification model, and output the failure probability. Decision and deployment module: Use Monte Carlo to quantify the prediction uncertainty, make maintenance decisions based on the double thresholds of failure probability and uncertainty, deploy the model on the cloud, and continuously optimize the model and adapt to new failure modes.

Images