An industrial device predictive maintenance system and method based on MCP protocol and LLM

Abstract

The application discloses an industrial equipment predictive maintenance system and method based on an MCP protocol and an LLM, wherein the method comprises the following steps: S1, collecting multi-modal operation data of industrial equipment through a sensor module, and transmitting the operation data to an MCP server through an MCP protocol; S2, preprocessing the collected operation data in the MCP server to improve data quality; S3, extracting time domain features and frequency domain features from the preprocessed data; S4, transmitting the extracted features to an LLM model in the cloud through the MCP protocol; S5, based on the inference ability of the LLM model, performing fault mode identification and diagnosis, analyzing the causes of the fault, and predicting the remaining service life of the industrial equipment or key components in combination with historical data; S6, generating maintenance suggestions and plans according to the fault diagnosis results, cause analysis and life prediction; and S7, outputting the diagnosis results, cause analysis, life prediction and maintenance suggestions to the user.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing, specifically relating to a predictive maintenance system and method for industrial equipment based on the MCP protocol and LLM. Background Technology

[0002] In modern industrial production, the stable operation of key equipment is crucial to production efficiency and product quality. Traditional equipment maintenance mainly employs the following methods: 1. Periodic maintenance: Equipment is inspected and maintained at fixed intervals. This method lacks specificity and is prone to wasting maintenance resources or causing insufficient maintenance. 2. Responsive repair: Repairing equipment after it malfunctions can lead to production interruptions and significant economic losses. 3. Threshold-based monitoring systems: These systems determine equipment status by setting fixed thresholds, but they cannot adapt to changes in equipment operating conditions and have a high false alarm rate.

[0003] In addition, another approach is to use a vibration analysis system. Traditional vibration analysis systems rely on signal processing techniques such as Fourier transform and wavelet transform for vibration analysis. While they can identify some fault characteristics, they have the following limitations: 1. Data silos: Traditional vibration analysis systems typically focus only on a single vibration signal, failing to integrate multi-source data for comprehensive analysis. Data sharing between different devices and systems is difficult, creating information silos. 2. Limited diagnostic capabilities: Traditional vibration analysis systems are insufficient in extracting early, subtle fault characteristics, making it difficult to detect potential equipment faults. 3. Insufficient real-time performance: Traditional systems have long processing speeds and response times, failing to meet the real-time requirements of modern industry. 4. Weak model generalization ability: Traditional mechanistic model-based methods struggle to adapt to changes in different operating conditions and equipment types, exhibiting limited generalization ability. 5. Insufficient maintenance decision support: Traditional systems typically only provide simple alarm functions, lacking intelligent support for maintenance decisions. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial equipment predictive maintenance system based on the MCP protocol and LLM, in order to solve the technical problems existing in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A predictive maintenance method for industrial equipment based on the MCP (Model Context Protocol) and LLM (Large Language Model) includes the following steps:

[0007] S1. Collect multimodal operating data of industrial equipment through sensor modules and transmit the operating data to the MCP server through the MCP protocol;

[0008] S2. On the MCP server, the collected operational data is preprocessed to improve data quality;

[0009] S3. Extract time-domain and frequency-domain features from the preprocessed data;

[0010] S4. Transmit the extracted features to the LLM model in the cloud via the MCP protocol;

[0011] S5. Based on the reasoning capability of the LLM model, it performs fault mode identification and diagnosis, analyzes the causes of faults, and predicts the remaining service life of industrial equipment or key components by combining historical data.

[0012] S6. Based on the fault diagnosis results, cause analysis and life prediction, generate maintenance suggestions and plans;

[0013] S7. Output the diagnostic results, cause analysis, life prediction and maintenance recommendations to the user.

[0014] Furthermore, in step S1, when transmitting runtime data to the MCP server via the MCP protocol, an industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol is employed, specifically:

[0015] Considering the varying demands of industrial equipment operation on real-time data transmission due to vibration intensity and load changes, a timeout period T is constructed to achieve dynamic timeout scheduling. timeout The coupling model with the device's dynamic health index (DHI) and data priority (P) is as follows:

[0016] ;

[0017] Among them, T base Based on timeout, DHI max The theoretical upper limit of the dynamic health index is given by α, where α is the operating condition influence coefficient, β(P) is the priority weighting function, and γ is the weighting factor. EMI As the electromagnetic interference compensation factor, we have:

[0018] ;

[0019] V EMI V represents the actual interference voltage. EMI,th The interference threshold and electromagnetic interference compensation factor γ EMI Used to compensate for transmission delays caused by electromagnetic interference in industrial environments;

[0020] Furthermore, the Dynamic Health Index (DHI) is calculated using the following formula:

[0021] ;

[0022] Where q is the number of health characteristics, and Sj Let j be the normalized health value of the j-th feature. For feature weights, The trend influence coefficient. The rate of change of characteristic health values, To maintain the feedback correction factor.

[0023] Furthermore, in step S1, the industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol further includes:

[0024] To address sudden noise interference in data transmission during industrial scenarios, a dual verification factor C based on CRC32 and data characteristics is employed for interference-resistant integrity verification, as shown in the following formula:

[0025] ;

[0026] Among them, Data i This represents the runtime data for the i-th frame, Check. i This represents the CRC32 checksum of the i-th frame of data, where n is the number of data frames within the sliding window, and γ data This represents the importance coefficient of the data.

[0027] The verification rules are as follows: when C < 0.01, the data is considered to be transmitted without distortion; when 0.01 ≤ C < 0.05, a partial retransmission is triggered; and when C ≥ 0.05, a full retransmission within the window is triggered.

[0028] Furthermore, in step S1, the multimodal operating data includes at least vibration, temperature, and current data;

[0029] In step S2, the preprocessing of the collected running data specifically includes denoising and filtering the running data, and fusing the multimodal running data using a time-weighted fusion algorithm.

[0030] Furthermore, in step S2, when fusing the multimodal runtime data using a time-weighted fusion algorithm, specifically:

[0031] The fused feature vector F is calculated by performing time-series alignment and propagation delay compensation on the multimodal run count using the following formula. fusion :

[0032] ;

[0033] Where m is the number of modes. Let X be the weight of the k-th mode. k This is the original data matrix for the k-th mode. The modal number time-series decay factor, The intermodal time delay compensation factor is represented by PCA, which stands for Principal Component Analysis.

[0034] Furthermore, the weights of each mode are adjusted using the following formula. :

[0035] ;

[0036] ;

[0037] Among them, E k Let p be the information entropy of the k-th mode. ki Let y represent the normalized proportion of the i-th feature in the k-th mode, and y represent the number of features in the k-th mode. The contribution of the fault to the k-th mode.

[0038] Furthermore, in step S5, the remaining useful life (RUL) is predicted using the following formula:

[0039] ;

[0040] Among them, LLM ind This indicates that the process is performed using an industrial-grade LLM, S deg For the local degradation sequence of the equipment, Encode is the degradation sequence encoder, Embed(L, N) is the operating condition factor embedding vector, and the operating condition factors include load L and speed N, α age The aging factor of the equipment. This is the operating condition coupled time decay error term, used to compensate for error accumulation under high operating conditions, and is calculated using the following formula:

[0041] ;

[0042] L rated For rated load, N rated This is the rated speed.

[0043] Furthermore, in step S5, the LLM inference weights are iteratively optimized using the following formula:

[0044] ;

[0045] Among them, W new With W old These represent the inference weights before and after the update, respectively, where μ is the learning rate. For loss function, Let β be the gradient of the loss function with respect to the inference weights W. conf To maintain the confidence level of the feedback.

[0046] Furthermore, in step S6, when generating maintenance recommendations, the Maintenance Priority Index (MPI) is calculated using the following formula:

[0047] ;

[0048] in, All values ​​are weights, where S is the severity index of the fault and I is the production impact coefficient. The production task urgency factor is represented by D, where D is the time taken to detect equipment failure, and R is the remaining equipment availability. This is the resource adaptation coefficient.

[0049] Another aspect of the present invention provides an industrial equipment predictive maintenance system based on the MCP protocol and LLM, comprising:

[0050] The equipment layer includes industrial equipment and sensor modules used to collect operational data from the industrial equipment.

[0051] The edge computing layer, including the MCP server, is used to communicate with the sensor modules in the device layer, acquire the operational data collected by the sensor modules, and perform data cleaning and preprocessing.

[0052] The cloud platform, including the LLM model, knowledge base management module, and maintenance decision engine, is used to implement the core processing of the system and generate relevant data of industrial equipment based on the processing results.

[0053] The smart terminal is wirelessly connected to the cloud platform, enabling operation and maintenance personnel to obtain relevant data of industrial equipment through the interactive interface of the smart terminal;

[0054] Furthermore, the aforementioned predictive maintenance method for industrial equipment based on the MCP protocol and LLM is adopted to achieve predictive maintenance of industrial equipment.

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] The predictive maintenance system and method for industrial equipment based on the MCP protocol and LLM provided by this invention realizes the transformation from traditional periodic maintenance and reactive repair to predictive maintenance by introducing the MCP protocol and LLM technology into the field of industrial equipment maintenance. It can effectively integrate multi-source data, accurately diagnose equipment faults, accurately predict the remaining service life of equipment, and generate intelligent maintenance suggestions. It has the advantages of high fault diagnosis accuracy, long warning time, and strong adaptability. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a predictive maintenance method for industrial equipment based on the MCP protocol and LLM provided by the present invention. Detailed Implementation

[0058] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates how this invention is implemented.

[0059] In one specific embodiment, refer to Figure 1 As shown, this invention provides a predictive maintenance method for industrial equipment based on the MCP protocol and LLM, comprising the following steps:

[0060] S1. Collect multimodal operating data of industrial equipment through sensor modules and transmit the operating data to the MCP server through the MCP protocol;

[0061] S2. On the MCP server, the collected operational data is preprocessed to improve data quality;

[0062] S3. Extract time-domain and frequency-domain features from the preprocessed data;

[0063] S4. Transmit the extracted features to the LLM model in the cloud via the MCP protocol;

[0064] S5. Based on the reasoning capability of the LLM model, it performs fault mode identification and diagnosis, analyzes the causes of faults, and predicts the remaining service life of industrial equipment or key components by combining historical data.

[0065] S6. Based on the fault diagnosis results, cause analysis and life prediction, generate maintenance suggestions and plans;

[0066] S7. Output the diagnostic results, cause analysis, life prediction and maintenance recommendations to the user.

[0067] Preferably, in one specific embodiment:

[0068] In step S1, multimodal operating data (such as vibration, temperature, current, etc.) of industrial equipment are collected through sensor modules, and the operating data is transmitted to the MCP server through the MCP protocol.

[0069] When transmitting operational data to the MCP server via the MCP protocol, the traditional MCP protocol's transmission mechanism is engineered to address the challenges of electromagnetic interference packet loss due to concurrent transmission from multiple sensors in industrial environments and the varying transmission requirements caused by fluctuations in equipment operating conditions. Specifically, an industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol is employed:

[0070] Considering the varying demands of industrial equipment (such as CNC machine tools and fracturing equipment) on the real-time performance of data transmission due to vibration intensity and load changes, a timeout period T is constructed to achieve dynamic timeout scheduling. timeout The coupling model with the device's dynamic health index (DHI) and data priority (P) is as follows:

[0071] ;

[0072] Among them, T base Basic timeout (T) base Based on the calibration of equipment communication link characteristics, the Modbus TCP link of CNC machine tools is set to 50ms, and the industrial Ethernet link of fracturing equipment is set to 80ms.

[0073] DHI max This represents the theoretical upper limit of the dynamic health index. The normalized value of the Dynamic Health Index (DHI) is 1, which indicates a healthy state without faults.

[0074] α is the operating condition influence coefficient, which can be calibrated in reverse using historical fault data. For example, it can be 0.8 for the spindle module of a CNC machine tool and 1.2 for the pump head module of a fracturing equipment.

[0075] β(P) is the priority weighting function, which realizes that the higher the priority, the shorter the timeout window. For example, the value of P can be 1, 2, or 3. P=1 represents regular monitoring data, and β(1) takes 0.5; P=2 represents fault warning data, and β(2) takes 1; P=3 represents emergency control commands, and β(3) takes 1.5.

[0076] γ EMI The electromagnetic interference compensation factor, calculated based on real-time electromagnetic interference monitoring values, is as follows:

[0077] ;

[0078] V EMI V represents the actual interference voltage. EMI,th The interference threshold and electromagnetic interference compensation factor γ EMI Used to compensate for transmission delays caused by electromagnetic interference in industrial environments.

[0079] Furthermore, the Dynamic Health Index (DHI) is calculated using the following formula:

[0080] ;

[0081] Where q represents the number of health features, which can be screened based on equipment failure modes. For example, the health features of CNC machine tools include vibration peak, kurtosis, temperature deviation, current fluctuation, etc.

[0082] The characteristic weights can be calibrated through fault tree analysis (FTA). For example, the peak value of spindle vibration can be 0.3, and the temperature deviation can be 0.2.

[0083] The trend influence coefficient is determined based on the equipment type and failure mode. For example, it can be taken as 0.6 for CNC machine tool spindle failure and 0.8 for compressor bearing failure.

[0084] To maintain the feedback correction factor, it is calibrated based on the results of manual maintenance. The value is 1.1 if the fault is eliminated after maintenance, and 0.9 if the fault is not alleviated after maintenance, so as to achieve closed-loop alignment between DHI and actual working conditions.

[0085] The value represents the rate of change in characteristic health values; negative values ​​indicate deterioration of health, while positive values ​​indicate recovery.

[0086] S j The normalized health value for the j-th feature can be mapped through the fault threshold interval:

[0087] ;

[0088] Let j be the detection value of the j-th feature. This is the normal threshold. This is the fault threshold.

[0089] Preferably, the industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol further includes:

[0090] To address sudden noise interference in data transmission during industrial scenarios, a dual verification factor C based on CRC32 and data characteristics is employed for interference-resistant integrity verification, as shown in the following formula:

[0091] ;

[0092] Among them, Data i The running data of the i-th frame, such as vibration acceleration and temperature value, needs to be normalized and preprocessed.

[0093] Check i This represents the CRC32 checksum of the i-th frame of data, which can be calculated using the polynomial 0xEDB88320 and is compatible with industrial communication standards.

[0094] n is the number of data frames within the sliding window, determined based on the industrial data transmission bandwidth, and can be fixed at 32 frames to balance verification efficiency and accuracy.

[0095] γ data The importance coefficient of the data can be calibrated based on fault sensitivity. For example, vibration and shock data can be set to 1.2, temperature change data can be set to 0.9, and current fluctuation data can be set to 1.0.

[0096] The verification rule is as follows: when C < 0.01, the data is considered to be transmitted without distortion; when 0.01 ≤ C < 0.05, a partial retransmission is triggered; and when C ≥ 0.05, a full retransmission within the window is triggered to ensure industrial-grade data reliability.

[0097] In step S2, the collected runtime data is preprocessed on the MCP server to improve data quality.

[0098] Preferably, the preprocessing of the collected operational data specifically includes denoising and filtering the operational data, and fusing the multimodal operational data using a time-weighted fusion algorithm to achieve efficient fusion of multimodal data.

[0099] When fusing multimodal runtime data using a time-weighted fusion algorithm, specifically:

[0100] The fused feature vector F is calculated by performing time-series alignment and propagation delay compensation on the multimodal run count using the following formula. fusion :

[0101] ;

[0102] Where m is the number of modes. For example, it can cover 3 types of core mode vibration, temperature and current, so m=3. Let X be the weight of the k-th mode. k Let be the original data matrix for the k-th mode.

[0103] The modality number time-series decay factor, taking into account data timeliness, ,in For data collection interval, The attenuation coefficient can be set to 0.05 for vibration and shock data and 0.02 for temperature conduction data.

[0104] It serves as an intermodal time delay compensation factor, based on the calibration of the physical structure of industrial equipment, to compensate for the response time delay of different sensors.

[0105] PCA stands for Principal Component Analysis, which performs dimensionality reduction, retains principal components with a cumulative contribution rate of ≥95%, and removes redundant noise from industrial data.

[0106] Furthermore, to avoid bias in subjective weight setting, the failure contribution of each mode was calculated by combining industrial equipment failure cases. The weights of each mode are adjusted using the following formula. :

[0107] ;

[0108] ;

[0109] Among them, E k Let p be the information entropy of the k-th mode. The smaller the entropy value, the higher the data's ability to distinguish faults. ki y represents the normalized proportion of the i-th feature in the k-th mode; y represents the number of features in the k-th mode. The contribution of the k-th mode to the fault can be based on historical fault data statistics. For example, in bearing faults, the contribution of the vibration mode is taken as 0.7, the temperature mode as 0.2, and the current mode as 0.1, to ensure that the weights match the correlation with the fault.

[0110] Traditional PCA fusion algorithms ignore the temporal correlation of industrial data (such as the transmission lag of vibration and temperature) and the drift characteristics of operating conditions. This algorithm achieves time lag compensation through a time decay factor, which is equivalent to correcting the weight values ​​in the traditional entropy weight method.

[0111] In step S3, time-domain features and frequency-domain features are extracted from the preprocessed data.

[0112] In step S4, the extracted features are transmitted to the LLM model in the cloud via the MCP protocol. Utilizing the dynamic context management capabilities of the MCP protocol, device operating status information is updated in real time, providing comprehensive and accurate reasoning basis for the LLM model.

[0113] In step S5, based on the reasoning capability of the LLM model, fault mode identification and diagnosis are performed, the causes of faults are analyzed, and the remaining service life of industrial equipment or key components is predicted by combining historical data.

[0114] Preferably, addressing the issues of domain knowledge gaps and weak generalization ability of generalized LLM models in industrial fault reasoning, an industrial-specific Prompt engineering and operating condition coupled Remaining Useful Life (RUL) prediction model is designed. Based on the traditional LM4 RUL framework, real-time operating condition factors (load L, speed N) and equipment aging coefficients are introduced. The remaining useful life (RUL) is predicted using the following formula:

[0115] ;

[0116] Among them, LLM ind It indicates that the process is performed using an industrial-specific LLM, which can be fine-tuned based on Llama3-70B using 150,000+ industrial fault-life data. During fine-tuning, an industrial-domain word segmentation dictionary is used, and the accuracy of professional terminology recognition is ≥96%.

[0117] S deg For local degradation sequences of equipment, such as the daily variation sequence of bearing vibration peak values, the length can be 90 days; Encode is the degradation sequence encoder, and Embed(L, N) is the condition factor embedding vector.

[0118] α age This is the equipment aging factor, used to compensate for accelerated degradation caused by equipment aging, and can be based on the equipment's service life. calculate: .

[0119] This is the operating condition coupled time decay error term, used to compensate for error accumulation under high operating conditions, and is calculated using the following formula:

[0120] ;

[0121] L rated For rated load, N rated This is the rated speed.

[0122] Preferably, in step S5, the LLM inference weights are iteratively optimized using the following formula:

[0123] ;

[0124] Among them, W new With W old These are the inference weights before and after the update, respectively; μ is the learning rate, calibrated based on the sparsity of industrial data, for example, it can be taken as 0.001 to avoid weight oscillation; For the loss function, the fault classification task can use weighted cross-entropy loss, while the RUL prediction task uses MAE loss. The gradient of the loss function with respect to the inference weights W can be calculated using the AdamW optimizer, with a weight decay coefficient of 0.01; β conf To maintain the confidence level of feedback, the integrity of maintenance records can be marked, such as 1.0 for "clear fault type + component disassembly verification" and 0.6 for "visual inspection only".

[0125] In step S6, maintenance recommendations and plans are generated based on the fault diagnosis results, cause analysis, and life prediction, thereby realizing the transformation from "passive response" to "proactive prevention".

[0126] Traditional maintenance decision-making focuses solely on fault severity. This invention incorporates production task urgency, maintenance resource constraints, and fault propagation risk to construct a three-dimensional coupled maintenance prioritization model. When generating maintenance recommendations, the Maintenance Priority Index (MPI) is calculated using the following formula:

[0127] ;

[0128] in, All values ​​are weights, with S representing the severity of the fault. A five-level quantification system can be used, with S ranging from 1 to 5. S=1 indicates that the fault has no substantial impact on system operation, while S=5 corresponds to a catastrophic fault that can cause the complete shutdown of core equipment (such as a broken machine tool spindle). The higher the severity of the fault, the higher the maintenance priority.

[0129] I is the production impact coefficient, which can be calculated using the following formula:

[0130] ;

[0131] The unit time output value loss when the equipment is down (unit: yuan / hour). Estimated downtime (in hours). The daily total output value of the production line (unit: yuan) is limited to the following range: The greater the impact of the I reaction on production, the higher the maintenance priority.

[0132] The production task urgency factor is dynamically assessed based on the order fulfillment cycle, and the calculation formula is as follows:

[0133] ;

[0134] This is the order delivery deadline. For the current time point, To generate the lead time.

[0135] D represents the time it takes for the equipment fault to be detected. The longer the fault has existed, the higher the maintenance priority. R represents the remaining availability of the equipment. The higher the remaining availability of the equipment, the lower the maintenance priority. The resource adaptation coefficient measures the degree of matching between the currently available maintenance resources (such as manpower, materials, time, etc.) and the maintenance needs of the equipment.

[0136] In step S7, the diagnostic results, cause analysis, life prediction, and maintenance recommendations are output to the user.

[0137] Another aspect of the present invention provides an industrial equipment predictive maintenance system based on the MCP protocol and LLM, comprising:

[0138] The equipment layer includes industrial equipment and sensor modules for collecting operational data from that equipment. Industrial equipment can include various types such as CNC machine tools, fracturing equipment, drilling pumps, mud pumps, compressors, and bearings; sensor modules can include vibration sensors, temperature sensors, current sensors, etc., to collect various types of operational data.

[0139] The edge computing layer, including the MCP server, is used to communicate with the sensor modules in the device layer, acquire the operational data collected by the sensor modules, and perform data cleaning and preprocessing.

[0140] The cloud platform, comprising an LLM model, a knowledge base management module, and a maintenance decision engine, is used to implement the core processing of the system and generate relevant data for industrial equipment based on the processing results. The knowledge base management module stores knowledge such as equipment failure modes, maintenance strategies, and historical maintenance records; the maintenance decision engine generates maintenance suggestions and plans based on the relevant equipment data and the knowledge base.

[0141] The smart terminal, which is wirelessly connected to the cloud platform, enables operation and maintenance personnel to obtain relevant data of industrial equipment through the interactive interface of the smart terminal.

[0142] Furthermore, the aforementioned predictive maintenance method for industrial equipment based on the MCP protocol and LLM is adopted to achieve predictive maintenance of industrial equipment.

[0143] This invention deploys an MCP server and lightweight data processing logic at the edge, and an LLM model and core business logic in the cloud. This deployment method balances computing resource consumption and system response speed, making it suitable for large-scale device deployments. For enterprises with high data security requirements, a private cloud deployment approach can be adopted, where all system components are deployed in the enterprise's internal private cloud environment to ensure data security and system controllability. Alternatively, a hybrid cloud deployment approach can be used, with some non-sensitive data and functions deployed in the public cloud, and sensitive data and core functions deployed in the private cloud, fully leveraging the advantages of both public and private clouds to achieve the best balance between performance and security.

[0144] Furthermore, the Resources module of the MCP protocol can achieve millisecond-level latency data stream injection via SSE, enabling the model to have a complete information graph during inference, supporting unified schema mapping for cross-modal data (text, images, structured data), and achieving dynamic context management. Sandboxed tool calls can also be implemented through the Tools module of the MCP protocol, supporting everything from simple API calls to complex system operations; a declarative DSL-based definition of the tool call pipeline enables dynamic task orchestration.

[0145] The LLM model supports a multi-LLM aggregator architecture, with different LLMs responsible for different diagnostic tasks. It also employs a federated learning mechanism to achieve knowledge sharing and model optimization across equipment and factories. Combined with a knowledge base management module, internal enterprise data (such as maintenance manuals and fault case libraries) can be used to fine-tune the LLM, improving the accuracy of terminology recognition. Before use, a domain knowledge graph (such as an equipment fault relationship network) can be pre-configured to force the LLM model to follow logical reasoning paths. Each tool call result updates the knowledge base and model parameters in reverse, allowing the system to continuously optimize the diagnostic model and prediction algorithm through the accumulation of equipment operation data and maintenance experience.

[0146] The following are explanations of several specific application examples:

[0147] 1) Predictive maintenance of CNC machine tools

[0148] Three-dimensional vibration sensors are installed at the front and rear ends of the spindle of the CNC machine tool, with a sampling frequency set to 20kHz; an infrared temperature sensor is installed to monitor the surface temperature of the bearing housing; and the sensor data is transmitted to the MCPServer of the edge computing layer via the MCP protocol.

[0149] Time-domain analysis of vibration signals is performed to calculate peak value, root mean square value and other indicators; frequency-domain analysis of vibration signals is performed to calculate frequency band energy, envelope spectrum and other indicators to identify bearing fault characteristic frequencies; and temperature change trend and current fluctuation characteristics are extracted.

[0150] The extracted feature data is transmitted to the LLM model in the cloud via the MCP protocol. The LLM model performs fault diagnosis based on the input feature data, identifying faults such as bearing wear and misalignment. It also predicts the remaining service life of the equipment by combining historical data, and issues an early warning when the predicted remaining service life is less than 72 hours.

[0151] Based on the diagnostic results and predictive data, maintenance recommendations are generated, such as "stop the machine immediately to check and replace the spindle bearing"; detailed maintenance steps and spare parts lists are generated; the maintenance recommendations are sent to the edge computing layer via the MCP protocol and finally displayed on the terminal user interface.

[0152] 2) Predictive maintenance of the compressor

[0153] Vibration and temperature sensors are installed in key parts of the compressor to collect vibration signals, temperature data, and operating parameters; the data is then transmitted to the edge computing layer via the MCP protocol.

[0154] Spectral analysis of vibration signals is performed to identify the characteristic frequencies of common compressor faults, such as frequency components corresponding to problems like imbalance and misalignment; temperature change characteristics and pressure fluctuation characteristics are extracted; and the Dynamic Health Index (DHI) is calculated to assess the overall health status of the compressor.

[0155] The extracted feature data is transmitted to the cloud via the MCP protocol; the LLM model analyzes the input data to identify problems such as imbalance, misalignment, or valve failure in the compressor; and predicts the remaining service life of key compressor components, such as bearings and impellers.

[0156] Based on diagnostic results and predictive data, maintenance recommendations are generated, such as "adjust compressor impeller balance" or "replace valve assembly"; a maintenance priority assessment is generated, indicating which maintenance tasks need to be prioritized; and maintenance recommendations and plans are sent to end users via the MCP protocol.

[0157] 3) Predictive maintenance of bearings

[0158] A vibration sensor is installed on the bearing housing to collect vibration signals; the vibration data is transmitted to the edge computing layer via the MCP protocol.

[0159] Time-frequency analysis is performed on the vibration signal to extract the characteristic frequencies of bearing faults, such as the inner ring fault frequency and the outer ring fault frequency; the kurtosis, peak factor and other indicators of the vibration signal are calculated to assess the degree of bearing wear.

[0160] The extracted feature data is transmitted to the cloud via the MCP protocol; the LLM model analyzes the input data to identify the bearing failure type and severity; and the LM4RUL framework is used to predict the remaining service life of the bearing.

[0161] Based on diagnostic results and predictive data, maintenance recommendations are generated, such as "replace the bearing immediately" or "closely monitor and plan to replace it in the next maintenance cycle"; detailed maintenance steps and spare parts lists are generated; and maintenance recommendations are sent to end users via the MCP protocol.

[0162] This invention provides a predictive maintenance system and method for industrial equipment based on the MCP protocol and LLM. It applies the MCP protocol to the field of predictive maintenance for industrial equipment, achieving efficient fusion and real-time transmission of multi-source data. Utilizing the dynamic context management capabilities of the MCP protocol, it provides LLM with more comprehensive and accurate reasoning basis. Through the tool invocation mechanism of the MCP protocol, it achieves full-process automation from data acquisition to maintenance decision-making. Applying the understanding and reasoning capabilities of LLM to industrial equipment fault diagnosis creates a new model for equipment maintenance. Utilizing the long-sequence modeling capabilities of LLM, it achieves accurate prediction of the remaining service life of equipment. Through the self-learning capabilities of LLM, the system can continuously adapt to new equipment and fault modes. Furthermore, this invention effectively integrates multiple types of data such as vibration, temperature, and current, enabling a comprehensive assessment of the equipment's health status. It realizes a shift from single-signal analysis to multi-dimensional feature fusion, improving the accuracy of fault diagnosis.

[0163] Furthermore, the predictive maintenance system and method for industrial equipment based on the MCP protocol and LLM provided by this invention can effectively reduce maintenance costs and unnecessary disassembly and inspection losses. Through early warning and precise maintenance, it improves spare parts inventory turnover, further reducing costs. Predictive maintenance solutions can effectively reduce unplanned downtime, significantly improve equipment utilization, thereby increasing production efficiency and greatly improving fault handling efficiency. By timely identifying potential problems and taking preventative measures, it can effectively extend equipment lifespan, reduce equipment failure rates, reduce the extent of equipment damage, and lower maintenance costs. Predictive maintenance solutions can also improve production safety, reduce energy consumption and resource waste, reduce waste generation, and promote sustainable development.

[0164] It is evident that, compared to traditional vibration analysis systems, this invention possesses significant advantages in terms of technological innovation, economic benefits, and social benefits. It can effectively reduce maintenance costs, improve production efficiency, extend equipment lifespan, and drive a transformation in industrial equipment maintenance models. With the deepening development of Industry 4.0 and intelligent manufacturing, predictive maintenance will become the mainstream model for industrial equipment maintenance. This invention, through the innovative application of the MCP protocol and LLM technology, provides strong technical support for this trend and has broad application prospects and market value.

[0165] In summary, the predictive maintenance system and method for industrial equipment based on the MCP protocol and LLM provided by this invention, by introducing the MCP protocol and LLM technology into the field of industrial equipment maintenance, realizes the transformation from traditional periodic maintenance and reactive repair to predictive maintenance. It can effectively integrate multi-source data, accurately diagnose equipment faults, accurately predict the remaining service life of equipment, and generate intelligent maintenance suggestions. It has advantages such as high fault diagnosis accuracy, long warning time, and strong adaptability.

[0166] Finally, it should be noted that the above description is only an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A predictive maintenance method for industrial equipment based on MCP protocol and LLM, characterized in that, Includes the following steps: S1. Collect multimodal operating data of industrial equipment through sensor modules and transmit the operating data to the MCP server through the MCP protocol; S2. On the MCP server, the collected operational data is preprocessed to improve data quality; S3. Extract time-domain and frequency-domain features from the preprocessed data; S4. Transmit the extracted features to the LLM model in the cloud via the MCP protocol; S5. Based on the reasoning capability of the LLM model, it performs fault mode identification and diagnosis, analyzes the causes of faults, and predicts the remaining service life of industrial equipment or key components by combining historical data. S6. Based on the fault diagnosis results, cause analysis and life prediction, generate maintenance suggestions and plans; S7. Output the diagnostic results, cause analysis, lifespan prediction, and maintenance recommendations to the user; In step S1, when transmitting runtime data to the MCP server via the MCP protocol, an industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol is employed. Specifically: Considering the varying demands of industrial equipment operation on real-time data transmission due to vibration intensity and load changes, a timeout period T is constructed to achieve dynamic timeout scheduling. timeout The coupling model with the device's dynamic health index (DHI) and data priority (P) is as follows: ； Among them, T base Based on timeout, DHI max The theoretical upper limit of the dynamic health index is given by α, where α is the operating condition influence coefficient, β(P) is the priority weighting function, and γ is the weighting factor. EMI As the electromagnetic interference compensation factor, we have: ； V EMI The actual interference voltage, V EMI,th The interference threshold and electromagnetic interference compensation factor γ EMI Used to compensate for transmission delays caused by electromagnetic interference in industrial environments; Furthermore, the Dynamic Health Index (DHI) is calculated using the following formula: ； Where q is the number of health characteristics, and S j Let j be the normalized health value of the j-th feature. For feature weights, The trend influence coefficient. The rate of change of characteristic health values, To maintain the feedback correction factor.

2. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 1, characterized in that, In step S1, the industrial-grade adaptive data transmission optimization algorithm based on the MCP protocol further includes: To address sudden noise interference in data transmission during industrial scenarios, a dual verification factor C based on CRC32 and data characteristics is employed for interference-resistant integrity verification, as shown in the following formula: ； Among them, Data i This represents the runtime data for the i-th frame, Check. i This represents the CRC32 checksum of the i-th frame of data, where n is the number of data frames within the sliding window, and γ... data This represents the importance coefficient of the data. The verification rules are as follows: when C < 0.01, the data is considered to be transmitted without distortion; when 0.01 ≤ C < 0.05, a partial retransmission is triggered; and when C ≥ 0.05, a full retransmission within the window is triggered.

3. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 1, characterized in that, In step S1, the multimodal operating data includes at least vibration, temperature, and current data; In step S2, the preprocessing of the collected running data specifically includes denoising and filtering the running data, and fusing the multimodal running data using a time-weighted fusion algorithm.

4. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 3, characterized in that, In step S2, when fusing the multimodal runtime data using a time-weighted fusion algorithm, specifically: The fused feature vector F is calculated by performing time-series alignment and propagation delay compensation on the number of multimodal runs using the following formula. fusion : ； Where m is the number of modes. Let X be the weight of the k-th mode. k This is the original data matrix for the k-th mode. The modal number time-series decay factor, The intermodal time delay compensation factor is represented by PCA, which stands for Principal Component Analysis. Furthermore, the weights of each mode are adjusted using the following formula. : ； ； Among them, E k Let p be the information entropy of the k-th mode. ki Let y represent the normalized proportion of the i-th feature in the k-th mode, and y represent the number of features in the k-th mode. The contribution of the fault to the k-th mode.

5. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 1, characterized in that, In step S5, the remaining useful life (RUL) is predicted using the following formula: ； Among them, LLM ind This indicates that the process is performed using an industrial-grade LLM, S deg For the local degradation sequence of the equipment, Encode is the degradation sequence encoder, Embed(L, N) is the operating condition factor embedding vector, and the operating condition factors include load L and speed N, α age The aging factor of the equipment. This is the operating condition coupled time decay error term, used to compensate for error accumulation under high operating conditions, and is calculated using the following formula: ； L rated For rated load, N rated This is the rated speed.

6. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 5, characterized in that, In step S5, the LLM inference weights are iteratively optimized using the following formula: ； Among them, W new With W old These represent the inference weights before and after the update, respectively, where μ is the learning rate. For loss function, Let β be the gradient of the loss function with respect to the inference weights W. conf To maintain the confidence level of the feedback.

7. The predictive maintenance method for industrial equipment based on MCP protocol and LLM according to claim 1, characterized in that, In step S6, when generating maintenance recommendations, the Maintenance Priority Index (MPI) is calculated using the following formula: ； in, All values ​​are weights, where S is the severity index of the fault and I is the production impact coefficient. The production task urgency factor is represented by D, where D is the time taken to detect equipment failure, and R is the remaining equipment availability. This is the resource adaptation coefficient.

8. A predictive maintenance system for industrial equipment based on the MCP protocol and LLM, characterized in that, include: The equipment layer includes industrial equipment and sensor modules used to collect operational data from the industrial equipment. The edge computing layer, including the MCP server, is used to communicate with the sensor modules in the device layer, acquire the operational data collected by the sensor modules, and perform data cleaning and preprocessing. The cloud platform, including the LLM model, knowledge base management module, and maintenance decision engine, is used to implement the core processing of the system and generate relevant data of industrial equipment based on the processing results. The smart terminal is wirelessly connected to the cloud platform, enabling operation and maintenance personnel to obtain relevant data of industrial equipment through the interactive interface of the smart terminal; Furthermore, the predictive maintenance method for industrial equipment based on the MCP protocol and LLM as described in any one of claims 1-7 is used to realize predictive maintenance of industrial equipment.

Images