Industrial equipment fault early warning method based on big data analysis

Abstract

The invention specifically relates to the technical field of big data processing, and discloses an industrial equipment fault early warning method based on big data analysis, which comprises the steps of S1, dynamic early warning target definition: outputting an early warning target weight; s2, multi-dimensional data acquisition: outputting standardized equipment operation data to a cloud server; s3, performing automatic feature engineering to obtain a high-value fault feature vector; s4, industrial equipment fault diagnosis: carrying out industrial equipment fault diagnosis to obtain a fault diagnosis result; s5, industrial fault dynamic early warning: determining a fault early warning level, and sending early warning information to an early warning response terminal; s6, model iteration and optimization: realizing model self-evolution; by constructing a full-link data efficient processing system, constructing an automatic feature engineering and deep learning end-to-end diagnosis system, and designing a target-threshold-model three-layer iteration system, the fault recognition precision and generalization ability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and more specifically, to a method for early warning of industrial equipment faults based on big data analysis. Background Technology

[0002] With the advancement of industrial automation, industrial equipment is developing towards integration and intelligence, significantly increasing system complexity. In addition, the popularization of IoT technology has led to a surge in the number of equipment sensors, enabling the real-time collection of massive amounts of data such as temperature, vibration, pressure, and current. Equipment failures may cause production interruptions and a decline in product quality. Early intervention through fault warnings can significantly reduce maintenance costs, improve production efficiency, and ensure personnel safety.

[0003] Existing methods for early warning of industrial equipment failures include defining early warning targets, multi-dimensional data collection, feature engineering, fault diagnosis and anomaly identification, and early warning issuance and response. These methods have enabled a shift from passive maintenance to proactive prediction, which is beneficial for optimizing operation and maintenance resources and reducing production losses.

[0004] However, it still has some drawbacks in actual use. First, the data processing efficiency is low. The traditional data processing architecture of the existing industrial equipment fault early warning method is difficult to support the high throughput writing and low latency query of large-scale data. It is easy to cause data congestion, loss or lag, resulting in low data processing and utilization efficiency, and it is difficult to support accurate early warning in complex scenarios.

[0005] Second, the fault identification capability is insufficient. Existing industrial equipment fault early warning methods rely heavily on traditional algorithms for fault diagnosis and anomaly identification. The core logic is feature matching and empirical rules. Faults are often coupled with multiple components, and fault detection is prone to misjudgment and missed judgment. Frequent manual intervention is required, resulting in a weak ability to identify hidden faults.

[0006] Third, there is a lack of dynamic adaptive mechanisms. The existing industrial equipment fault early warning methods are mostly static closed loops throughout the entire process, maintaining fixed logic operation for a long time. They lack data feedback and model iteration mechanisms from big data analysis and rely solely on manual periodic calibration, resulting in insufficient adaptive and iterative optimization capabilities of the system and easy decay of early warning accuracy over the long term. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide an industrial equipment fault early warning method based on big data analysis. By constructing an edge-cloud layered collaborative + distributed storage computing architecture and a full-link efficient data processing system, an automated feature engineering + deep learning end-to-end diagnostic system is built, and a target-threshold-model three-layer iterative system is designed, which effectively solves the problems of low data processing efficiency, insufficient fault identification capability and lack of dynamic adaptive mechanism proposed in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for early warning of industrial equipment faults based on big data analysis, comprising a multimodal sensor terminal, an edge computing terminal, a cloud server, an early warning response terminal, and a model management terminal, the steps of which are as follows:

[0009] S1: Dynamic early warning target definition: Read the basic information of the device, the life cycle stage and historical data through the cloud server, output the weight of the early warning target, and transmit and store it to the model management terminal;

[0010] S2: Multi-dimensional data acquisition: Collect equipment operation data through multi-modal sensor terminals, transmit it to edge computing terminals for preprocessing, and output standardized equipment operation data to cloud servers;

[0011] S3: Automated Feature Engineering: Extracting and filtering features from standardized equipment operation data via a cloud server to obtain high-value fault feature vectors;

[0012] S4: Industrial equipment fault diagnosis: Industrial equipment fault diagnosis is performed through a cloud server based on high-value fault feature vectors and a pre-trained model stored in the model management terminal to obtain fault diagnosis results;

[0013] S5: Dynamic early warning of industrial faults: Calculates dynamic thresholds through a cloud server, determines the fault early warning level based on fault diagnosis results and dynamic thresholds, and sends early warning information to the early warning response terminal;

[0014] S6: Model Iteration and Optimization: Model performance is monitored through the model management terminal, and incremental training calculations are performed on the cloud server based on operation and maintenance feedback data to achieve model self-evolution.

[0015] The technical effects and advantages of this invention are as follows:

[0016] 1. This invention constructs an edge-cloud layered collaborative + distributed storage computing architecture and a full-link efficient data processing system. It processes the raw data collected by multimodal sensor terminals through edge computing terminals to achieve invalid data filtering and effective data simplification. It also supports million-level high-throughput writes and millisecond-level queries through the time-series database of the cloud server, achieving efficient processing and low-latency flow of large-scale data.

[0017] 2. This invention constructs an automated feature engineering + deep learning end-to-end diagnostic system, extracts multi-dimensional features through cloud servers, and performs deep feature mining to form a high-value feature vector composed of basic features and deep features. This effectively captures the static statistical patterns of equipment faults and significantly improves fault identification accuracy and generalization ability.

[0018] 3. This invention designs a three-layer iterative system of target-threshold-model, outputs personalized early warning target weights for devices based on association rules and lifecycle adaptation coefficients, and adapts to device aging and operating condition fluctuations by constructing a dynamic threshold calculation model. Through the model management terminal, it monitors core indicators such as accuracy and false alarm rate in real time, and updates model parameters regularly to achieve system self-evolution and continuous optimization of early warning accuracy. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0020] Figure 2 This is a schematic diagram of the overall structure of the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the steps for obtaining high-value fault feature vectors according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] As attached Figure 1 The method for early warning of industrial equipment faults based on big data analysis includes a multimodal sensor terminal, an edge computing terminal, a cloud server, an early warning response terminal, and a model management terminal.

[0024] In a more specific application of the present invention, the multimodal sensor terminal is used to collect raw data of equipment operation in multiple dimensions, including physical state data, electrical parameter data and environmental impact data, including vibration sensors, temperature sensors, pressure sensors, current / voltage sensors and acoustic sensors. The multimodal sensor terminal is connected to the edge computing terminal through a 5G network / industrial Ethernet.

[0025] Edge computing terminals are used to preprocess raw data locally, load lightweight models, trigger local instant responses, and transmit effective data to cloud servers to achieve cloud-edge collaboration. This includes an edge computing framework, a data preprocessing algorithm library, and a lightweight fault identification model. An edge computing terminal connects to a cloud cluster service node of a cloud server via industrial-grade 5G to achieve data transmission.

[0026] The cloud server is used for distributed storage of massive time-series data and supports efficient retrieval. It performs automated feature engineering and drives fault diagnosis, performs incremental model training and optimization, and realizes system self-evolution. It includes a feature engineering tool library and a fault diagnosis model library. The feature engineering tool library includes FFT spectrum analysis, wavelet transform, CNN / LSTM / Transformer feature extraction modules, and mutual information feature screening algorithms. The fault diagnosis model includes a CNN-LSTM fusion model, a transfer learning module, and an expert rule base. The cloud server connects to the early warning response terminal via Gigabit Ethernet / mobile APP and connects to the model management terminal via Gigabit Ethernet.

[0027] Early warning response terminals are used to issue graded early warning information, display fault details, and manage closed-loop responses. These terminals include industrial control screens and smartphones, and respond to early warnings based on early warning level information, fault diagnosis results, handling suggestions, and dynamic thresholds transmitted from cloud servers.

[0028] The model management terminal is used to manage model versions, monitor model performance in real time, schedule automated training, and deploy and iterate models. Based on the model performance indicators, new training data, and model update requests transmitted from the cloud server, it outputs model optimization instructions, optimized model packages, model performance reports, and version rollback instructions.

[0029] For the connection methods of the aforementioned multimodal sensor terminal, edge computing terminal, cloud server, early warning response terminal, and model management terminal, please refer to [link / reference]. Figure 2 .

[0030] The specific embodiments of the present invention include the following steps:

[0031] S1: Dynamic early warning target definition: Read the basic information of the device, the life cycle stage and historical data through the cloud server, output the weight of the early warning target, and transmit and store it to the model management terminal;

[0032] Furthermore, the basic equipment information includes equipment ID, equipment type, rated parameters, installation location, and commissioning time. The life cycle stage is determined based on the commissioning time and the design service life. Historical data includes historical fault data, historical maintenance data, and historical operating condition data.

[0033] In this embodiment, it is necessary to specify the equipment type, such as asynchronous motor, gearbox, centrifugal compressor, and the rated parameters, including speed, power, and load range. The life cycle determination needs to be based on the commissioning time t and the design service life T. When t < 0.1T, it is the new machine break-in period; when 0.1T ≤ t ≤ 0.8T, it is the stable operation period; when t > 0.8T, it is the aging and decline period. Historical fault data includes fault type, fault occurrence time, triggering conditions, and fault impact. Historical operation and maintenance data includes maintenance records, spare parts replacement lists, and periodic maintenance reports. Historical operating condition data includes historical load curves, environmental temperature and humidity change trends, and production scheduling plans.

[0034] Furthermore, the steps for obtaining the weights of the early warning targets are as follows:

[0035] S1.1: Obtain basic equipment information, lifecycle stage, and historical data; preprocess the historical data; define the dimensions of early warning targets; and form a target set G={G1,G2,……,G...} m}, where m is the number of common fault types of the target industrial equipment;

[0036] In this embodiment, it should be specifically explained that data preprocessing refers to removing erroneous records from fault data, retaining valid fault samples (≥500 groups), and performing data encoding and alignment to ensure that the operating conditions of each group of data are evenly distributed. The early warning target dimensions include fault type dimension, early warning priority dimension, and monitoring parameter dimension. The early warning priority dimension refers to the basic score for priority division based on the degree of fault impact, and the monitoring parameter dimension refers to the core monitoring parameters corresponding to each fault type, such as the vibration peak value, kurtosis, and temperature corresponding to bearing wear.

[0037] S1.2: Mine the correlation rules between operating condition parameters and fault types in historical data, and calculate the correlation degree R between operating condition and fault. i Calculate the lifecycle adaptation coefficient C based on the current lifecycle stage of the device. i And obtain the fault impact weight I i , for R i C i and I i We perform a weighted summation to obtain the target weight W for the early warning system. i ;

[0038] In this embodiment, it is necessary to specifically explain the operating condition-fault correlation degree R. i The calculation requires constructing transaction sets, each including load range, ambient temperature range, and fault type. A minimum support of 5% and a minimum confidence of 30% are set to obtain association rules for each fault G. i Take the mean confidence score of all association rules, normalize it to the interval [0.1, 0.9], and obtain R.i Lifecycle adaptation coefficient C i The calculation requires statistical analysis of the historical occurrence frequency f of each fault within the current lifecycle stage. i , through f i The fit coefficient C was calculated. i =f i / max(f1,f2,……,f m Fault impact weight I i The acquisition of R requires assigning impact weights based on the economic losses and security risks caused by the failure; i Reflecting predictability, C i Reflecting the risks at the current stage, I i It reflects the value of early warning.

[0039] S1.3: The cloud server is based on the warning target weight W i A warning target configuration table is generated and synchronously pushed to the model management terminal. The model management terminal establishes an independent warning target version file for each device.

[0040] In this embodiment, it should be specifically noted that the warning target configuration table includes device ID, lifecycle stage, warning target, and warning target weight W. i The system includes core detection parameters, early warning lead time requirements, and false alarm rate thresholds. The early warning target version file is used to record the weight calculation time, data source, and calculation parameters, and supports version backtracking.

[0041] S2: Multi-dimensional data acquisition: Collect equipment operation data through multi-modal sensor terminals, transmit it to edge computing terminals for preprocessing, and output standardized equipment operation data to cloud servers;

[0042] Furthermore, the collection of equipment operation data requires the edge computing terminal to retrieve the early warning target weight table, obtain the core detection parameters, and send parameter acquisition instructions to the multimodal sensor terminal. The parameter acquisition instructions include the acquisition frequency, accuracy requirements, and transmission cycle. The equipment operation data includes physical status data, electrical parameter data, and environmental impact data.

[0043] In this embodiment, it should be specifically noted that the acquisition frequency in the parameter acquisition command refers to 1kHz for high-fault core parameters, 100Hz for ordinary parameters, and 10Hz for low-weight fault parameters; the accuracy requirements are ±0.1g for vibration sensors, ±0.01°C for temperature sensors, and ±0.1A for current sensors; the transmission cycle refers to high-frequency data being transmitted in 100ms chunks and low-frequency data being transmitted in batches of 1s.

[0044] Further explanation is needed regarding the physical state data, which includes vibration frequency, temperature, equipment load, and rotational speed; the electrical parameter data, which includes three-phase current, voltage, and power; and the environmental impact data, which includes ambient temperature, humidity, and dust concentration. The multimodal sensor terminal collects equipment operation data according to a preset sampling frequency, identifies outliers in the real-time collected data, removes invalid data caused by sensor drift and transient interference, performs linear interpolation to repair the data, ensures the continuity of time-series data, aligns the data collected by different sensors to ensure consistent timestamps across multiple data sources, and removes redundant data to obtain standardized equipment operation data.

[0045] S3: Automated Feature Engineering: Extracting and filtering features from standardized equipment operation data via a cloud server to obtain high-value fault feature vectors;

[0046] Furthermore, such as Figure 3 As shown, the steps for obtaining high-value fault feature vectors are as follows:

[0047] S2.1: The cloud server obtains standardized equipment operation data, retrieves the early warning target weight table to extract the core monitoring parameters corresponding to high-weight faults, sets a 10-second time window with a step size of 5 seconds, and obtains standardized equipment operation data in segments.

[0048] In this embodiment, it is necessary to specifically explain that the core monitoring parameters corresponding to high-weight faults are obtained as the priority basis for feature extraction. The time window needs to be adapted to the time scale of the equipment fault characteristics. The cloud-distributed computing nodes of the cloud server distribute the sharded tasks according to the parameter type and device ID in a load-balanced manner to ensure the efficiency of feature extraction.

[0049] S2.2: Based on the core monitoring parameters corresponding to high-weight faults, the time-domain features, frequency-domain features, time-frequency-domain features, and depth features of standardized equipment operation data are automatically extracted to obtain the original fault feature set, which includes basic features and 64-dimensional depth features.

[0050] In this embodiment, it should be specifically explained that time-domain feature extraction is for continuous parameters in standardized equipment operation data, obtaining the mean, kurtosis, and peak factor; frequency-domain feature extraction is for parameters with strong periodicity such as vibration and acoustics, extracting the spectral peak and frequency centroid from the spectral amplitude; time-frequency domain feature extraction is for non-stationary data during equipment startup / shutdown and load abrupt changes, using db4 wavelet transform to decompose the time-domain data into wavelet coefficients of 8 scales, calculating the energy of each scale, and quantifying the signal complexity through energy entropy; depth feature extraction requires calling the cloud-based CNN-LSTM fusion model to automatically extract spatial features and capture time-dependent features, outputting a 64-dimensional depth feature vector.

[0051] S2.3: Use mutual information method to filter features of the original fault feature set, and use weighted fusion method to fuse features of the original fault feature set to obtain high-value fault feature vectors. Fix the dimension of the high-value fault feature vectors to 10-15 dimensions.

[0052] In this embodiment, it is important to specify that feature selection requires using the fault label as the target variable. The mutual information MI(X,Y) between each original feature and the fault label is calculated, where X is the feature and Y is the label. A mutual information threshold is set, for example, MI≥0.2, determined based on historical performance data from the model management terminal. Core features with mutual information higher than the threshold are selected. For features corresponding to high-weight faults, the threshold is appropriately lowered to ensure that the key features of that fault are not overlooked. Feature fusion requires allocating basic features and deep features using the formula F=0.6F. b +0.4F d The high-value fault feature vectors are obtained by fusion.

[0053] S4: Industrial equipment fault diagnosis: Industrial equipment fault diagnosis is performed through a cloud server based on high-value fault feature vectors and a pre-trained model stored in the model management terminal to obtain fault diagnosis results;

[0054] Furthermore, industrial equipment fault diagnosis requires the cloud server to retrieve a CNN-LSTM fusion pre-trained model, including CNN layers, LSTM layers, and fully connected layers, from the model management terminal. High-value fault feature vectors are input into the model, and the convolutional output feature map y is obtained based on the CNN convolutional feature extraction formula. conv It also uses LSTM gating and cell state update formulas to output the 64-dimensional hidden state h. t Through the formula:

[0055] ,

[0056] The confidence level P(y=k) for the equipment to belong to the k-th type of fault is calculated, where k takes the value 1, 2, ..., C, z. k This represents the linear output of the fully connected layer for the k-th type of fault, where C is the number of fault categories. The fault diagnosis result is generated based on the confidence P(y=k) that the device belongs to the k-th type of fault, including the fault type, severity, and confidence level. The format is Device ID-Time-Fault Type-Severity-Confidence.

[0057] In this embodiment, it should be specifically noted that the CNN convolution feature extraction formula is expressed as follows:

[0058] ,

[0059] Where y convThe output feature map is generated by the convolution, where K is the number of convolution kernels and w k f is the weight of the k-th convolutional kernel. pk The k-th local patch of the feature vector, for example, 3×1 in size, corresponds to 3 consecutive features. b is the bias term, σ(x)=max(0,x), used to automatically capture the spatial feature pattern corresponding to the fault. The LSTM gating and cell state update formula receives the 32-dimensional feature map output by the CNN, filters effective information through the gating mechanism, removes random fluctuation features, captures the temporal evolution law of equipment faults, and outputs a 64-dimensional hidden state h. t It integrates spatial features and temporal dependencies to provide comprehensive input for fault classification; k Through formula z k =W fc ×h t +b fc The calculation yielded W. fc b represents the weights of the fully connected layer. fc This is a bias term.

[0060] It should be further explained that the fault type refers to the fault category corresponding to the maximum value of P(y=k); the severity is determined based on the magnitude of P(y=k) and the feature vector, with P≥0.9 being medium-term and P≥0.95 being critical; and the confidence level refers to the maximum value of P(y=k).

[0061] S5: Dynamic early warning of industrial faults: Calculates dynamic thresholds through a cloud server, determines the fault early warning level based on fault diagnosis results and dynamic thresholds, and sends early warning information to the early warning response terminal;

[0062] Furthermore, the calculation of the dynamic threshold requires the cloud server to retrieve real-time device operation data. Based on the device's cumulative running time t1, the current actual load of the device, and the baseline threshold b1 in the real-time device operation data, the dynamic threshold T(t1,load) under the current operating condition of the device is calculated using the formula T(t1,load)=a1×t1+a2×load+b1, where a1 is the running time weight coefficient and a2 is the load weight coefficient.

[0063] In this embodiment, it should be specifically noted that the equipment running time t1 is after standardization processing, the current actual load of the equipment refers to the actual load / rated load, the running time weight coefficient is obtained by training with historical data, and reflects the impact of equipment aging on the threshold. For example, for a motor, a1=-0.2, the longer it runs, the lower the threshold. The load weight coefficient reflects the impact of the load on the fault threshold. For example, the threshold decreases under high load, a2=-0.3. The baseline threshold b1 is the standard threshold when the equipment is new and has no load. For example, the motor vibration baseline threshold b=5.0g. By replacing the traditional fixed threshold with a dynamic threshold, the problem of false alarms / missed alarms caused by equipment aging and changes in operating conditions is solved.

[0064] Furthermore, the determination of the fault warning level requires obtaining the fault severity score S1, fault confidence score S2, and equipment importance score S3 based on the fault diagnosis results, dynamic thresholds, and equipment importance. The warning weighted score S is calculated using the formula S=0.6×S1+0.3×S2+0.1×S3, and the fault warning level is divided into the first warning level, the second warning level, and the third warning level based on the warning weighted score.

[0065] In this embodiment, it should be specifically noted that the severity scores for faults are 8 points for emergency faults, 4 points for mid-term faults, 2 points for early-stage faults, and 0 points for normal faults, determined based on the diagnostic results; the fault confidence score is calculated using S2=P(y=k)×10, and the equipment importance score is 10 points for core production equipment and 5 points for auxiliary equipment; the specific rules for classifying fault warning levels are as follows:

[0066] When S < 3, it is at the first warning level, requiring SMS notification but no emergency action is needed.

[0067] When 3≤S<6 points, it is at the second warning level, and an APP pop-up and SMS will be sent. It will be dealt with within 24 hours.

[0068] When S≥6, it is at the third warning level, requiring audible and visual alarms, APP, SMS and voice notifications, and handling within 2 hours, triggering shutdown of core equipment.

[0069] It should be further explained that if the fault warning level is the third warning level and the equipment is a core production equipment, the cloud server synchronously sends a shutdown command to the edge computing terminal. The edge computing terminal sends a shutdown signal to the equipment control unit through hardwiring, and at the same time sends a backup command through the industrial bus to ensure reliable execution. After the equipment is shut down, the edge computing terminal sends a shutdown confirmation signal back to the cloud server, and the warning response terminal synchronously updates the work order status to "emergency handling". The maintenance personnel submit the handling results through the APP / work order system, feed back the data to the cloud server, and store it in the model management terminal. The feedback data includes the authenticity of the fault, the handling completion time, the maintenance effect, and abnormal feedback.

[0070] S6: Model Iteration and Optimization: Model performance is monitored through the model management terminal, and incremental training calculations are performed on the cloud server based on operation and maintenance feedback data to achieve model self-evolution.

[0071] Furthermore, the monitoring indicators for model performance monitoring include model accuracy, false positive rate, false negative rate, and dynamic threshold adaptation error. Incremental training calculation refers to updating model parameters through gradient descent and updating dynamic threshold parameters as the model iterates.

[0072] In this embodiment, it should be specifically noted that the triggering condition for model performance monitoring is that when the accuracy is <90%, the false positive rate is >5%, or the adaptation error is >3%, an iteration request is automatically sent to the cloud; the data for incremental training consists of high-value feature vectors and operation and maintenance feedback data; the optimized model has an accuracy ≥93% and a false positive rate ≤4%. If the target is met, a new version of the model is generated, and the optimized model package is pushed to the edge computing terminal via SFTP to replace the old model. The model management terminal records version logs and supports historical version rollback.

[0073] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0074] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning of industrial equipment faults based on big data analysis, characterized in that, Including multimodal sensor terminals, edge computing terminals, cloud servers, early warning and response terminals, and model management terminals, the steps are as follows: S1: Dynamic early warning target definition: Read the basic information of the device, the life cycle stage and historical data through the cloud server, output the weight of the early warning target, and transmit and store it to the model management terminal; S2: Multi-dimensional data acquisition: Collect equipment operation data through multi-modal sensor terminals, transmit it to edge computing terminals for preprocessing, and output standardized equipment operation data to cloud servers; S3: Automated Feature Engineering: Extracting and filtering features from standardized equipment operation data via a cloud server to obtain high-value fault feature vectors; S4: Industrial equipment fault diagnosis: Industrial equipment fault diagnosis is performed through a cloud server based on high-value fault feature vectors and a pre-trained model stored in the model management terminal to obtain fault diagnosis results; S5: Dynamic early warning of industrial faults: Calculates dynamic thresholds through a cloud server, determines the fault early warning level based on fault diagnosis results and dynamic thresholds, and sends early warning information to the early warning response terminal; S6: Model Iteration and Optimization: Model performance is monitored through the model management terminal, and incremental training calculations are performed on the cloud server based on operation and maintenance feedback data to achieve model self-evolution.

2. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The basic equipment information includes equipment ID, equipment type, rated parameters, installation location, and commissioning time. The life cycle stage is determined based on the commissioning time and the designed service life. Historical data includes historical fault data, historical maintenance data, and historical operating condition data.

3. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The steps for obtaining the weight of the early warning target are as follows: S1.1: Obtain basic equipment information, lifecycle stage, and historical data; preprocess the historical data; define the dimensions of early warning targets; and form a target set G={G1,G2,……,G...} m }, where m is the number of common fault types of the target industrial equipment; S1.2: Mine the correlation rules between operating condition parameters and fault types in historical data, and calculate the correlation degree R between operating condition and fault. i Calculate the lifecycle adaptation coefficient C based on the current lifecycle stage of the device. i And obtain the fault impact weight I i , for R i C i and I i We perform a weighted summation to obtain the target weight W for the early warning system. i ; S1.3: The cloud server is based on the warning target weight W i A warning target configuration table is generated and synchronously pushed to the model management terminal. The model management terminal establishes an independent warning target version file for each device.

4. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The acquisition of the equipment operation data requires the edge computing terminal to retrieve the early warning target weight table, obtain the core detection parameters, and send parameter acquisition instructions to the multimodal sensor terminal. The parameter acquisition instructions include the acquisition frequency, accuracy requirements, and transmission cycle. The equipment operation data includes physical status data, electrical parameter data, and environmental impact data.

5. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The steps for obtaining the high-value fault feature vector are as follows: S2.1: The cloud server obtains standardized equipment operation data, retrieves the early warning target weight table to extract the core monitoring parameters corresponding to high-weight faults, sets a 10-second time window with a step size of 5 seconds, and obtains standardized equipment operation data in segments. S2.2: Based on the core monitoring parameters corresponding to high-weight faults, the time-domain features, frequency-domain features, time-frequency-domain features, and depth features of standardized equipment operation data are automatically extracted to obtain the original fault feature set, which includes basic features and 64-dimensional depth features. S2.3: Use mutual information method to filter features of the original fault feature set, and use weighted fusion method to fuse features of the original fault feature set to obtain high-value fault feature vectors. Fix the dimension of the high-value fault feature vectors to 10-15 dimensions.

6. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The industrial equipment fault diagnosis requires the cloud server to retrieve a CNN-LSTM fusion pre-trained model, including CNN layers, LSTM layers, and fully connected layers, from the model management terminal. High-value fault feature vectors are input into the model, and the convolutional output feature map y is obtained based on the CNN convolutional feature extraction formula. conv It also uses LSTM gating and cell state update formulas to output the 64-dimensional hidden state h. t Through the formula: ， The confidence level P(y=k) for the equipment to belong to the k-th type of fault is calculated, where k takes the value 1, 2, ..., C, z. k This represents the linear output of the fully connected layer for the k-th type of fault, where C is the number of fault categories. The fault diagnosis result is generated based on the confidence P(y=k) that the device belongs to the k-th type of fault, including the fault type, severity, and confidence level. The format is Device ID-Time-Fault Type-Severity-Confidence.

7. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The calculation of the dynamic threshold requires the cloud server to retrieve real-time device operation data. Based on the device's cumulative running time t1, the current actual load of the device, and the baseline threshold b1 in the real-time device operation data, the dynamic threshold T(t1,load) under the current operating condition of the device is calculated using the formula T(t1,load)=a1×t1+a2×load+b1, where a1 is the running time weight coefficient and a2 is the load weight coefficient.

8. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The determination of the fault warning level requires obtaining a fault severity score S1, a fault confidence score S2, and an equipment importance score S3 based on the fault diagnosis results, dynamic thresholds, and equipment importance. The warning weighted score S is calculated using the formula S=0.6×S1+0.3×S2+0.1×S3, and the fault warning level is divided into the first warning level, the second warning level, and the third warning level based on the warning weighted score.

9. The industrial equipment fault early warning method based on big data analysis according to claim 1, characterized in that: The monitoring indicators for model performance monitoring include model accuracy, false positive rate, false negative rate, and dynamic threshold adaptation error. Incremental training calculation refers to updating model parameters through gradient descent and updating dynamic threshold parameters as the model iterates.

Images