A remote fault diagnosis and prediction system and method for a soil granulator

Abstract

The application belongs to the technical field of soil granulator operation and maintenance, and discloses a remote fault diagnosis and prediction system and method for a soil granulator, which comprises a sensing and collecting layer for collecting multi-dimensional parameters of the granulator in real time; an edge computing layer is connected in communication with the sensing and collecting layer by using an industrial edge gateway; a remote transmission layer is used for establishing a bidirectional communication link between the edge computing layer and a cloud diagnosis and prediction layer; the cloud diagnosis and prediction layer adopts a double diagnosis and prediction mode of an improved hybrid intelligent algorithm combined with a fault knowledge base to complete fault type identification, fault location, fault level evaluation and fault trend prediction; and a terminal interaction layer supports multi-terminal cooperation to realize diagnosis result checking, early warning receiving, remote operation and data management. The system and method are used to realize accurate diagnosis, early prediction and remote control of the soil granulator fault, greatly improve the equipment operation and maintenance efficiency, reduce the operation and maintenance cost, and ensure the continuous and stable operation of the granulator.

Description

Technical Field

[0001] This invention relates to the field of soil granulation equipment operation and maintenance technology, and in particular to a remote fault diagnosis and prediction system and method for soil granulators. Background Technology

[0002] Soil granulators are core equipment in modern agriculture and environmental engineering, used to process loose soil, soil conditioners, and other materials into uniform granules. They are widely used in farmland improvement, mine reclamation, and landscaping. The operational stability of a soil granulator directly determines granulation efficiency, granule quality, and project progress. However, its working environment is usually harsh, often in outdoor or workshop settings with high dust levels, large humidity fluctuations, and frequent vibrations. After long-term high-load operation, various malfunctions are prone to occur, such as motor overload, bearing wear, die blockage, pressure roller eccentricity, and drive belt loosening.

[0003] Currently, the maintenance of soil granulators mainly relies on manual inspection. Maintenance personnel regularly go to the site to check the equipment's operating status and test key parameters, relying on their personal experience to judge potential faults and fault types. This traditional model has many inherent drawbacks: First, manual inspections have low coverage, typically below 60%, leading to significant delays in fault detection. Most faults are only discovered after equipment shutdown, resulting in prolonged downtime and severely impacting production schedules. Second, diagnostic accuracy is low, typically below 70%. Fault location relies heavily on the experience of maintenance personnel, making it difficult to effectively identify latent faults such as early bearing wear or inter-turn short circuits in motors, increasing the risk of misdiagnosis and missed diagnosis, and raising maintenance costs. Third, it lacks fault prediction capabilities, allowing only reactive maintenance after a fault occurs, making it impossible to predict fault development trends in advance and hindering preventative maintenance. Fourth, it lacks remote control capabilities. For multiple granulators deployed in a dispersed manner, centralized monitoring and remote diagnosis are impossible, resulting in a heavy workload and low response efficiency for maintenance personnel. Fifth, data cannot be effectively reused. Inspection records and fault cases are mostly paper-based or fragmented electronic records, making it difficult to form a standardized knowledge base, hindering the transfer of maintenance experience, and causing similar faults to recur.

[0004] While some industrial equipment has begun to adopt simple remote monitoring or fault diagnosis technologies, dedicated intelligent remote fault diagnosis and prediction systems for soil granulators remain scarce. Existing technologies either only enable remote monitoring of single parameters, failing to perform fault diagnosis and prediction; or they employ traditional single algorithms for fault identification, resulting in low diagnostic accuracy and weak anti-interference capabilities, making them unsuitable for the complex scenarios of soil granulators with multiple fault types and interference factors; or they fail to consider the core operating conditions of the soil granulator, such as the correlation between material moisture content, granulation pressure, and faults, leading to a disconnect between diagnostic results and actual operation and maintenance needs. Furthermore, existing technologies generally suffer from high data transmission latency, insufficient edge processing capabilities, and poor fault warning timeliness, failing to meet the operation and maintenance requirements for the continuous and stable operation of soil granulators.

[0005] Therefore, developing an intelligent remote fault diagnosis and prediction system that can achieve multi-parameter acquisition, accurate diagnosis, early prediction, and remote control, and is adapted to the harsh working environment of soil granulators, and solve many of the shortcomings of existing technologies, has become an urgent technical problem to be solved in the field of soil granulation equipment operation and maintenance. Summary of the Invention

[0006] The purpose of this invention is to provide a remote fault diagnosis and prediction system and method for soil granulators, which aims to solve the technical problems of existing soil granulator fault diagnosis relying on manual inspection, slow response, low diagnostic accuracy, and inability to predict faults in advance, resulting in long equipment downtime, high operation and maintenance costs, and unstable granulation quality.

[0007] To achieve the above objectives, the present invention provides a remote fault diagnosis and prediction system for a soil granulator, comprising a sensing and acquisition layer, an edge computing layer, a remote transmission layer, a cloud-based diagnosis and prediction layer, and a terminal interaction layer, wherein each layer is sequentially connected in communication. The sensing and acquisition layer uses industrial-grade protected multi-type sensors to collect multi-dimensional parameters of the granulator in real time at a preset acquisition frequency, covering three major categories: equipment operation, environment, and granulation quality. The sensing and acquisition layer also includes a data caching module, which uses a local SD card to store and cache the original acquisition data to avoid data loss due to network interruption. After the network is restored, it will be automatically synchronized to the edge computing layer. The edge computing layer is deployed in the field control box of the soil granulator. It adopts an industrial-grade edge gateway, supports multi-protocol conversion, and communicates with the sensors and data cache modules of the sensing and acquisition layer to perform localized preprocessing, feature extraction and preliminary anomaly identification on the raw acquired data. The remote transport layer is used to establish a bidirectional communication link between the edge computing layer and the cloud diagnostic prediction layer to realize the transmission of data and instructions; The cloud-based diagnostic prediction layer, deployed on a cloud server, adopts a dual diagnostic prediction mode combining an improved MCNN-BiGRU hybrid intelligent algorithm with a fault knowledge base. Based on the data transmitted from the edge computing layer, it completes fault type identification, fault location, fault level assessment, and fault trend prediction. The terminal interaction layer includes industrial control terminals, mobile operation and maintenance terminals, and Web monitoring terminals. It supports multi-terminal collaboration, provides human-machine interaction interfaces for operation and maintenance personnel, and enables them to view diagnostic results, receive early warnings, operate remotely, and manage data.

[0008] Preferably, the sensing and acquisition layer includes: an operation parameter acquisition unit, an environmental parameter acquisition unit, and a granulation quality parameter acquisition unit; Operating parameter acquisition unit: includes vibration sensors, temperature sensors, current and voltage sensors, speed sensors, pressure sensors, and displacement sensors; among them, vibration sensors are deployed in the granulator bearings, pressure rollers, and motor housing to collect vibration acceleration and vibration frequency parameters; temperature sensors are deployed in the motor, bearings, gearbox, and die holes to collect the operating temperature of each component; current and voltage sensors are connected in series in the motor power supply circuit to collect the motor operating current and voltage parameters; speed sensors are deployed in the transmission system to collect the speed of the pressure rollers and main shaft; pressure sensors are deployed in the granulation chamber to collect the granulation pressure; and displacement sensors are deployed between the pressure rollers and the flat die to collect the displacement between the two. Environmental parameter acquisition unit: including temperature and humidity sensors and dust sensors; deployed around the granulator to collect ambient temperature, humidity and dust concentration parameters to help determine the cause of the fault; Granulation quality parameter acquisition unit: including particle size sensor and particle strength sensor; deployed at the discharge port of granulator, it collects average particle size, particle size uniformity and particle compressive strength parameters, and uses granulation quality abnormalities to help judge equipment failure. The perception and acquisition layer also includes a data caching module, which uses a local SD card for storage to cache the original acquired data, avoiding data loss due to network interruption, and automatically synchronizing to the edge computing layer after the network is restored.

[0009] Preferably, the specific implementation process of the edge computing layer includes the following steps: Step 1, Data Preprocessing: A three-level processing flow of outlier removal, noise reduction, and standardization is adopted to eliminate interference factors in the data collection process and ensure data validity; The algorithm employs wavelet thresholding to suppress noise caused by environmental dust and electromagnetic interference while preserving the effective signal. The wavelet basis is selected as the db4 wavelet, with a decomposition level of 5 layers, and the threshold... The calculation formula is: ; in, The standard deviation of noise; The number of samples to collect data; the high-frequency coefficients after decomposition are thresholded, and then the data is reconstructed through wavelet inverse transform to obtain the effective data after noise reduction; Step 2, Feature Extraction: For the preprocessed effective data, extract parameters that characterize the equipment's operating status and fault characteristics, which are divided into three categories: time domain features, frequency domain features, and correlation features; Step 2.1, Time Domain Features: For vibration, temperature, and current time series data, extract six time domain features: mean, variance, peak value, peak factor, kurtosis, and waveform factor to reflect the overall distribution and fluctuation characteristics of the data. Step 2.2, Frequency Domain Features: The vibration and current time-domain signals are converted into frequency-domain signals by Fast Fourier Transform (FFT), and three frequency-domain features, namely characteristic frequency, frequency peak value, and harmonic content, are extracted to reflect the fault characteristics of the equipment components. Step 2.3, Correlation Features: Based on the core operating conditions of the soil granulator, extract multi-parameter correlation features to assist in determining the fault type; Step 3, Preliminary Anomaly Identification: Based on the preset parameter threshold range, the preprocessed valid data and extracted feature data are calibrated using historical fault data to make threshold judgments; if a parameter or feature data exceeds the threshold range, a preliminary anomaly signal is generated, marking the anomaly parameter type, anomaly occurrence time and anomaly amplitude, and immediately pushed to the terminal interaction layer for local early warning. Simultaneously, the preprocessed data, feature data, and preliminary abnormal signals are sent to the cloud-based diagnostic prediction layer via the remote transmission layer for diagnosis and prediction.

[0010] Preferably, the implementation process of the remote transport layer is as follows: First, a hybrid transmission mode combining industrial Ethernet and wireless supplementation is adopted, with the transmission link using gigabit industrial Ethernet; Then, during data transmission, the AES-256 encryption algorithm is used to encrypt the transmitted data to prevent it from being stolen or tampered with. At the same time, a data fragmentation transmission mechanism is used to divide the data packet into multiple small data packets and mark them with sequence numbers. After the transmission is completed, they are reassembled in the cloud to ensure the integrity of data transmission. In addition, a transmission heartbeat mechanism is set up. If the cloud diagnostic prediction layer does not receive the heartbeat signal from the edge computing layer for three consecutive times, it is determined that the communication is interrupted. The communication abnormality warning is immediately pushed to the terminal interaction layer, and the local caching mechanism of the edge computing layer is triggered. The untransmitted data will be automatically synchronized after the communication is restored. Finally, the remote transmission layer also supports bidirectional command transmission. The operation and maintenance commands generated by the cloud diagnostic prediction layer are transmitted to the edge computing layer through the remote transmission layer. After receiving the commands, the edge computing layer sends them to the control system of the granulator to achieve remote control. The device status feedback information of the edge computing layer is transmitted to the cloud through the remote transmission layer to form a closed-loop control.

[0011] Preferably, the cloud-based diagnostic prediction layer includes four modules: a fault knowledge base module, a fault diagnosis module, a fault level assessment module, and a fault prediction module. (1) Fault knowledge base module: Construct a special fault knowledge base for soil granulators, integrate common fault types, fault characteristics, fault causes, fault locations and maintenance solutions for soil granulators, and support self-learning updates. (2) Fault diagnosis module adopts improved MCNN-BiGRU hybrid intelligent algorithm, introduces attention mechanism, filters key fault features, and combines feature data transmitted by edge computing layer and fault knowledge base to complete fault type identification and fault location; (3) Fault level assessment module: Based on fault type, abnormal parameter amplitude and fault development speed, the analytic hierarchy process (AHP) is used to assess the fault level, which is divided into four levels: Level I is minor fault, Level II is general fault, Level III is serious fault and Level IV is emergency fault. (4) Fault prediction module: The improved gray prediction model GM(1,1) is used to predict the time of occurrence and development trend of faults by combining historical operation data, current feature data and fault diagnosis results, so as to realize early warning of faults.

[0012] Preferably, the fault knowledge base consists of two parts: a basic library and a dynamically updated library. 1) Basic Database: 12 common faults of soil granulators are pre-entered, including: motor overload, bearing wear, die blockage, pressure roller eccentricity, transmission belt looseness, gearbox wear, electrical short circuit, abnormal granulation pressure, abnormal speed, poor particle formation, sensor failure, and material blockage; each fault type corresponds to a unique fault code, typical characteristic parameter range, fault cause, and repair steps; 2) Dynamically updated database: Based on historical diagnostic data and maintenance records, the system adds new fault types and optimizes the range of fault characteristic parameters through a self-learning algorithm. When a new type of fault that is not identified occurs, maintenance personnel can enter fault information and maintenance plans through the terminal interaction layer. The system automatically extracts the characteristic parameters of the fault and updates them to the fault knowledge base. The self-learning update frequency is once a week.

[0013] Preferably, an improved MCNN-BiGRU hybrid intelligent algorithm is used, combining feature data transmitted from the edge computing layer and a fault knowledge base, to complete fault type identification and fault location. The specific process is as follows: 1) Feature Input: The 12-dimensional feature data extracted from the edge computing layer, namely 6 time-domain features, 3 frequency-domain features, and 3 correlation features, are used as the algorithm input. An attention mechanism is used to weight each feature; the attention weight calculation formula is as follows: ; in, Let be the attention weight for the i-th feature; The importance score for the i-th feature is calculated through a fully connected layer. As the feature dimension, M=12 in this system; weighted feature data for: ; 2) Multi-scale feature extraction: A multi-scale convolutional neural network (MCNN) is used, employing three sets of parallel convolutional kernels of different sizes to perform multi-scale convolution processing on the weighted feature data; the convolutional kernels are 16×1, 8×1, and 4×1 respectively; the convolution calculation method is as follows: ; in, The output features are from the convolution; These are the kernel weights; For bias terms; It is the ReLU activation function; The kernel size is used; multi-scale convolutional features are integrated through a concatenation operation to obtain a spatial feature matrix. 3) Temporal feature mining: The spatial feature matrix output by MCNN is input into the bidirectional gated recurrent unit BiGRU to capture the temporal dependencies of the feature data; the temporal feature matrix is ​​obtained by concatenating the outputs of the forward GRU and the backward GRU. 4) Fault Identification and Localization: The spatiotemporal fusion features output by MCNN-BiGRU are input into the fully connected layer, and the probability of each fault type is calculated using the softmax function. Simultaneously, by combining the correspondence between fault types and fault locations in the fault knowledge base, the specific fault location is located. The softmax function calculation formula is as follows: ; in, For the input feature x to belong to the th The probability of a type of failure; The first output of the fully connected layer Fault type score; This represents the total number of fault types.

[0014] Preferably, an improved grey prediction model (GM(1,1)) is used, which combines historical operating data, current feature data, and fault diagnosis results to predict the timing and development trend of faults, thereby achieving early warning of faults. The specific process is as follows: First, data preprocessing: Select core characteristic parameters related to the fault, extract historical data from the past 24 hours, and smooth the data using a moving average method to eliminate random interference. The smoothing formula is as follows: ; in, The data is smoothed at time t; This represents the original feature data at time t; Secondly, model establishment: an improved GM(1,1) prediction model is constructed, and a decay coefficient is introduced. , To optimize the model's prediction accuracy, the model's differential equation is as follows: ; in, Generate sequences for the 1-AGO of the feature data; The development coefficient; Gray action quantity; attenuation coefficient The improved 1-AGO generation sequence calculation formula, used to adjust the weights of the accumulated generation, is as follows: ; in, The original feature data sequence; Generate a sequence value for the accumulation at time k; Next, the parameters are solved: the least squares method is used to solve for the model parameters. and The calculation formula is as follows: ; in, For parameter vectors; To generate an accumulation matrix; The original data vector; This is a transpose operation; Then, trend prediction and early warning: by solving differential equations, the predicted sequence of feature data is obtained. The predicted values ​​of the original feature data are generated through cumulative subtraction. The system compares the predicted value with the preset fault threshold. If the predicted value exceeds the fault threshold within the next T hours, a fault prediction report is generated, marking the predicted fault occurrence time, fault type, and fault level, and an early warning message is pushed out. At the same time, the system plots the trend curve of the characteristic parameters to intuitively display the fault development trend. Finally, model optimization: The model parameters are calibrated and updated hourly using the latest historical data to ensure prediction accuracy; when the prediction error exceeds 10%, the attenuation coefficient is automatically adjusted. The parameters are then re-solved to optimize the prediction model.

[0015] Preferably, the specific functions of the terminal interaction layer include: information display, early warning push, remote operation, data management and permission management.

[0016] A method for a remote fault diagnosis and prediction system for a soil granulator includes the following steps: Step S1, System Deployment and Initialization: Deploy the sensors of the sensing and acquisition layer to their corresponding positions on the soil granulator according to the preset locations, and complete sensor calibration; deploy the edge gateway of the edge computing layer to the field control box and complete the communication connection with the sensors; deploy the cloud server of the cloud diagnosis and prediction layer, build a fault knowledge base, and initialize algorithm parameters, including convolution kernel size, initial value of attention weight, and prediction model decay coefficient; deploy various terminals of the terminal interaction layer, and complete permission allocation and network configuration; after system initialization, preset the normal threshold and fault threshold of each parameter, and calibrate the basic parameters of acquisition frequency and transmission frequency. Step S2, Multi-dimensional Data Acquisition: Each sensor in the sensing and acquisition layer collects the operating parameters, environmental parameters, and granulation quality parameters of the granulator in real time according to the preset acquisition frequency. The raw data is stored in the local cache module to avoid data loss due to network interruption. At the same time, the raw data is transmitted to the edge computing layer in real time. Step S3, Edge Preprocessing and Preliminary Identification: The edge computing layer receives the raw collected data and sequentially completes three levels of preprocessing: outlier removal, noise reduction, and standardization to eliminate interference factors. Then, it extracts three types of feature data: time domain, frequency domain, and correlation. Based on a preset threshold, it performs preliminary anomaly identification on the preprocessed data and feature data. If anomalies are found, a preliminary anomaly signal is generated and a local warning is pushed. At the same time, the preprocessed data, feature data, and preliminary anomaly signal are packaged and encrypted and transmitted to the cloud diagnostic prediction layer through the remote transmission layer. Step S4, Cloud-based Diagnosis Prediction and Command Generation: The cloud-based diagnosis prediction layer receives data transmitted from the edge computing layer. The fault diagnosis module uses an improved MCNN-BiGRU algorithm, combined with a fault knowledge base, to identify the fault type and locate the fault location. The fault level assessment module uses the analytic hierarchy process (AHP) to assess the fault level. The fault prediction module uses an improved GM(1,1) model to predict the fault occurrence time and development trend. The system integrates the diagnosis results, level assessment results, and prediction reports to generate operation and maintenance commands, and feeds back the diagnosis results, prediction reports, and operation and maintenance commands to the edge computing layer and terminal interaction layer through the remote transmission layer. Step S5, Interactive Control and Closed-Loop Optimization: The terminal interaction layer receives diagnostic results, prediction reports, and maintenance instructions, and pushes early warning information through various means; maintenance personnel view relevant information and perform on-site repairs or remote operations according to the fault level and maintenance instructions; after the repair is completed, maintenance personnel enter the repair record through the terminal; the cloud-based diagnostic prediction layer updates the fault knowledge base and optimizes the diagnostic prediction algorithm parameters based on the repair record; the edge computing layer adjusts the collection frequency and preprocessing parameters according to the instructions fed back from the cloud to achieve closed-loop optimization of the system.

[0017] Therefore, the present invention employs the above-mentioned remote fault diagnosis and prediction system and method for a soil granulator, and the beneficial effects are as follows: (1) Achieve accurate fault diagnosis and early prediction, significantly reducing downtime: Using an improved intelligent algorithm and a dedicated fault knowledge base, the fault diagnosis accuracy is greater than or equal to 95%, the fault prediction accuracy is greater than or equal to 90%, and the early warning time is less than or equal to 1 hour. It can predict latent and explicit faults in advance, leaving sufficient maintenance time for operation and maintenance personnel and avoiding equipment downtime. After a fault occurs, it can quickly locate the fault location and provide maintenance guidance, shortening the maintenance time by ≥40% and ensuring the continuous and stable operation of the granulator.

[0018] (2) Realize remote intelligent operation and maintenance and reduce operation and maintenance costs: No need for manual regular inspections, support centralized remote control of multiple granulators, greatly reduce the workload of operation and maintenance personnel, and reduce operation and maintenance personnel costs by 25%; avoid ineffective maintenance caused by misjudgment or omission of faults, and reduce maintenance consumables and labor costs; realize preventive operation and maintenance through fault prediction, reduce equipment failure rate, and extend equipment service life.

[0019] (3) Adaptable to the harsh working environment of soil granulators and highly practical: The sensing and acquisition layer adopts IP67 industrial-grade protective sensors, which can adapt to outdoor and workshop scenarios with a lot of dust, large humidity changes and frequent vibrations; the edge computing layer supports local caching and local early warning, and the remote transmission layer supports automatic reconnection after network disconnection and encrypted transmission, ensuring continuous and stable operation of the system in complex environments; the algorithm introduces an anti-interference mechanism, which can effectively suppress environmental interference and improve the accuracy of diagnosis and prediction.

[0020] (4) Achieve multi-parameter fusion and closed-loop optimization to improve operation and maintenance quality: Collect multi-dimensional parameters of equipment operation, environment and granulation quality, and improve the accuracy of fault diagnosis through multi-parameter correlation analysis; combine fault knowledge base self-learning and algorithm parameter adaptive calibration to achieve closed-loop optimization of the system and continuously improve the accuracy of diagnosis and prediction and operation and maintenance efficiency; at the same time, use granulation quality parameters to assist in judging equipment faults and ensure stable granulation quality.

[0021] (5) Easy to operate, highly compatible and widely adaptable: The terminal interaction layer supports multi-terminal collaboration of industrial control terminals, mobile terminals and Web terminals, making operation convenient and adaptable to the operation and maintenance needs of different scenarios; the system supports multiple industrial protocols such as Modbus / TCP and Profinet, and is compatible with soil granulators of different models and manufacturers. It can be widely used in the operation and maintenance of soil granulation equipment in modern agriculture, environmental protection engineering, mine reclamation and other fields, and has high promotion and application value.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the architecture of a remote fault diagnosis and prediction system for a soil granulator according to the present invention. Figure 2 This is a flowchart of a remote fault diagnosis and prediction method for a soil granulator according to the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Example 1 like Figure 1 As shown, a remote fault diagnosis and prediction system for a soil granulator includes a sensing and acquisition layer, an edge computing layer, a remote transmission layer, a cloud-based diagnosis and prediction layer, and a terminal interaction layer. Each layer is connected in sequence to form a closed-loop system of acquisition, preprocessing, transmission, diagnosis and prediction, and interactive control.

[0026] I. Sensing and Acquisition Layer.

[0027] The sensing and acquisition layer is the core of the system's data source. It is adapted to the harsh working environment of soil granulators, which are characterized by high dust levels, large humidity fluctuations, and frequent vibrations. It adopts multiple types of sensors with industrial-grade protection and collects multi-dimensional parameters of the granulator in real time at a preset acquisition frequency. The frequency range is 1-10Hz and can be adaptively adjusted, covering three major categories: equipment operation, environment, and granulation quality, ensuring the comprehensiveness, accuracy, and real-time nature of the collected data.

[0028] The sensing and acquisition layer includes: an operation parameter acquisition unit, an environmental parameter acquisition unit, and a granulation quality parameter acquisition unit.

[0029] The operating parameter acquisition unit includes a vibration sensor (range 0-50 mm / s, accuracy ±0.1 mm / s), a temperature sensor (range -40~200℃, accuracy ±0.5℃), a current / voltage sensor (accuracy ±1%), a speed sensor, a pressure sensor, and a displacement sensor. Specifically, the vibration sensor is deployed in the granulator bearings, pressure rollers, and motor housing to collect vibration acceleration and frequency parameters; the temperature sensor is deployed in the motor, bearings, gearbox, and die holes to collect the operating temperature of each component; the current / voltage sensor is connected in series in the motor power supply circuit to collect the motor's operating current and voltage parameters; the speed sensor is deployed in the transmission system to collect the speed of the pressure rollers and spindle; the pressure sensor is deployed in the granulation chamber to collect the granulation pressure; and the displacement sensor is deployed between the pressure rollers and the flat die to collect the displacement between them.

[0030] Environmental parameter acquisition unit: including temperature and humidity sensors and dust sensors, deployed around the granulator to collect ambient temperature, humidity and dust concentration parameters to help determine the cause of the fault, such as short circuit of electrical components caused by high humidity, and blockage of mold holes caused by high dust.

[0031] Granulation quality parameter acquisition unit: including particle size sensor and particle strength sensor, deployed at the discharge port of granulator, to collect average particle size, particle size uniformity and particle compressive strength parameters. It can also help to judge equipment faults by granulation quality abnormalities. For example, particle size non-uniformity may correspond to eccentric pressure roller, die hole wear, etc.

[0032] The perception and acquisition layer also includes a data caching module, which uses a local SD card for storage. It can cache the original acquired data to avoid data loss due to network interruption and automatically synchronize it to the edge computing layer after the network is restored.

[0033] II. Edge computing layer.

[0034] The edge computing layer is deployed in the field control box of the soil granulator. It adopts an industrial-grade edge gateway, supports multi-protocol conversion, and communicates with the sensors and data cache modules of the sensing and acquisition layer. Its core function is to perform local preprocessing, feature extraction and preliminary anomaly identification on the raw collected data, reduce cloud computing pressure and reduce data transmission volume, and at the same time realize local preliminary early warning in the case of network outage, solving the problems of high data transmission latency and strong cloud dependence in the existing technology.

[0035] The specific implementation process of the edge computing layer includes the following steps: Step 1, Data Preprocessing: A three-level processing flow of outlier removal, noise reduction, and standardization is adopted to eliminate interference factors in the data collection process and ensure data validity.

[0036] Step 1.1, Outlier Removal: Based on the 3σ criterion, identify and remove outlier data that exceeds the normal range, such as sudden changes in values ​​caused by sensor malfunctions, as shown below: ; in, For the first One data point was collected; The average value of the collected parameters; The standard deviation of the parameter collection is denoted as . If the above formula is satisfied, it is determined to be an outlier, which is then removed. The missing value is supplemented by linear interpolation of two adjacent valid data to ensure data continuity.

[0037] Step 1.2, Noise Reduction Processing: For time-series data susceptible to interference such as vibration and current, a wavelet threshold noise reduction algorithm is used to suppress noise caused by environmental dust and electromagnetic interference, while retaining the effective signal; the wavelet basis is selected as db4 wavelet, the decomposition level is 5 levels, and the threshold is... The calculation formula is: ; in, The noise standard deviation is estimated using the high-frequency coefficients after decomposition. The number of samples to be collected; the high-frequency coefficients after decomposition are thresholded (soft threshold), and then the data is reconstructed through wavelet inverse transform to obtain the effective data after noise reduction.

[0038] Step 1.3, Standardization Processing: The min-max standardization method is used to map the preprocessed parameters to the [0,1] interval, eliminating the dimensional differences between different parameters and facilitating subsequent feature extraction and algorithm calculation, as shown below: ; in, These are the standardized values; This is the minimum collected value of the parameter; This represents the maximum value collected for the parameter.

[0039] Step 2, Feature Extraction: For the preprocessed effective data, extract the core parameters that can characterize the equipment's operating status and fault characteristics, which are divided into three categories: time domain features, frequency domain features, and correlation features.

[0040] Step 2.1, Time-domain features: For time-series data such as vibration, temperature, and current, six time-domain features are extracted: mean, variance, peak value, peak factor, kurtosis, and waveform factor. These features reflect the overall distribution and fluctuation characteristics of the data. Taking vibration acceleration as an example, the calculation formulas for each time-domain feature are as follows: Mean: ; variance: ; Peak value: ; Peak factor: ; kurtosis: ; Waveform factor: ; Step 2.2, Frequency Domain Features: Vibration, current, and other time-domain signals are converted into frequency-domain signals using Fast Fourier Transform (FFT). Three frequency-domain features—characteristic frequency, peak frequency, and harmonic content—are extracted to reflect the fault characteristics of equipment components. For example, bearing wear corresponds to an increase in peak values ​​at specific frequencies. The FFT calculation formula is as follows: ; in, These are time-domain signal samples; This represents the number of FFT points, corresponding to the number of samples in the collected data. For the amplitude of the frequency domain signal, through Extract frequency domain feature parameters.

[0041] Step 2.3, Correlation Features: Based on the core operating conditions of the soil granulator, extract multi-parameter correlation features, such as the ratio of granulation pressure to roller speed, the correlation coefficient between motor current and granulation pressure, and the difference between ambient humidity and die temperature, to help determine the fault type. For example, an abnormal increase in the ratio of granulation pressure to roller speed may correspond to die blockage.

[0042] Taking granulation pressure and motor current data as an example, the correlation coefficient is calculated using the following formula: ; in, These are the granulation pressure parameters; These are motor current parameters; , These are the means of the two, respectively; The correlation coefficient has a range of [-1, 1]. The larger the absolute value, the stronger the correlation.

[0043] Step 3, Preliminary Anomaly Identification: Based on the preset parameter threshold range, the preprocessed valid data and extracted feature data are calibrated using historical fault data to determine the threshold. If a parameter or feature data exceeds the threshold range, a preliminary anomaly signal is generated, marking the anomaly parameter type, anomaly occurrence time, and anomaly amplitude, and immediately pushed to the terminal interaction layer for local early warning. At the same time, the preprocessed data, feature data, and preliminary anomaly signal are packaged and sent to the cloud diagnostic prediction layer through the remote transmission layer for in-depth diagnosis and prediction.

[0044] III. Remote Transmission Layer.

[0045] The remote transmission layer is used to establish a bidirectional communication link between the edge computing layer and the cloud diagnostic prediction layer, enabling secure, efficient, and low-latency transmission of data and instructions. It is suitable for scenarios where multiple granulators are deployed in a distributed manner, solving the problems of unstable data transmission, high latency, and poor security in existing technologies.

[0046] First, a hybrid transmission mode combining industrial Ethernet and wireless supplementation is adopted. The core transmission link uses gigabit industrial Ethernet (Profinet / EtherNet / IP) with a low latency of less than or equal to 10ms to ensure real-time data transmission. For granulators deployed in workshop blind spots and outdoor areas, industrial-grade 5G CPE and LoRa are used for wireless supplementation, supporting automatic reconnection after network failure with a low latency of less than or equal to 20ms. LoRa can cover 1-3km and has low power consumption.

[0047] Then, during data transmission, the AES-256 encryption algorithm is used to encrypt the transmitted data to prevent it from being stolen or tampered with. Simultaneously, a data fragmentation mechanism is employed, dividing the packaged large data packets, such as feature data and historical cached data, into multiple smaller data packets, marking them with sequence numbers. After transmission, these smaller packets are reassembled in the cloud to ensure data transmission integrity. Furthermore, a transmission heartbeat mechanism is implemented at 30-second intervals. If the cloud-based diagnostic prediction layer fails to receive a heartbeat signal from the edge computing layer three consecutive times, it is considered a communication interruption. An abnormal communication warning is immediately pushed to the terminal interaction layer, and the local caching mechanism of the edge computing layer is triggered. Once communication is restored, the untransmitted data is automatically synchronized.

[0048] Finally, the remote transmission layer also supports bidirectional command transmission. Operation and maintenance commands generated by the cloud diagnostic prediction layer, such as parameter adjustment commands and shutdown warning commands, are transmitted to the edge computing layer through the remote transmission layer. After receiving them, the edge computing layer sends them to the control system of the granulator to achieve remote management and control. The device status feedback information of the edge computing layer, such as command execution results and device operating status, is also transmitted to the cloud through the remote transmission layer to form a closed-loop management and control system.

[0049] IV. Cloud-based diagnostic and prediction layer.

[0050] The cloud-based diagnostic prediction layer is the core diagnostic prediction unit of the system. Deployed on a cloud server, it adopts a dual diagnostic prediction mode that combines an improved intelligent algorithm with a fault knowledge base. Based on the data transmitted from the edge computing layer, it completes fault type identification, fault location, fault level assessment, and fault trend prediction, solving the problems of low diagnostic accuracy, lack of predictive ability, and poor adaptability in existing technologies.

[0051] The cloud-based diagnostic prediction layer comprises four core modules: a fault knowledge base module, a fault diagnosis module, a fault level assessment module, and a fault prediction module.

[0052] (1) Fault Knowledge Base Module: Construct a dedicated fault knowledge base for soil granulators, integrating common fault types, fault characteristics, fault causes, fault locations, and maintenance solutions for soil granulators. It also supports self-learning updates to continuously improve diagnostic accuracy. The fault knowledge base consists of two parts: a basic database and a dynamically updated database.

[0053] 1) Basic Database: 12 common faults of soil granulators are pre-entered, including: motor overload, bearing wear, die blockage, pressure roller eccentricity, transmission belt looseness, gearbox wear, electrical short circuit, abnormal granulation pressure, abnormal speed, poor particle formation, sensor failure, and material blockage; each fault type corresponds to a unique fault code, typical characteristic parameter range, fault cause, and repair steps.

[0054] For example, typical characteristics of die blockage include: continuous increase in granulation pressure, uneven particle size at the discharge port, increased die temperature, and increased motor current; causes of failure include: excessive material moisture content, excessive material impurities, and failure to clean the die in time; maintenance steps include: stopping the machine to clean the die, adjusting the material moisture content, and installing a grid screen to filter impurities.

[0055] 2) Dynamically updated database: Based on historical diagnostic data and maintenance records, the system continuously adds fault types and optimizes the range of fault characteristic parameters through a self-learning algorithm. When a new type of fault that is not identified occurs, maintenance personnel can enter fault information and maintenance plans through the terminal interaction layer. The system automatically extracts the characteristic parameters of the fault and updates them to the fault knowledge base, thereby achieving dynamic improvement of the knowledge base. The self-learning update frequency is once a week.

[0056] (2) Fault diagnosis module adopts improved MCNN-BiGRU hybrid intelligent algorithm, combined with feature data transmitted from edge computing layer and fault knowledge base, to complete fault type identification and fault location, with a diagnosis accuracy of greater than or equal to 95%. The core improvement is the introduction of attention mechanism to screen key fault features, improve anti-interference ability, and adapt to the multi-interference and multi-fault scenarios of soil granulator.

[0057] 1) Feature Input: The 12-dimensional feature data extracted from the edge computing layer, namely 6 time-domain features, 3 frequency-domain features, and 3 correlation features, are used as the algorithm input. An attention mechanism is used to weight and allocate each feature, highlighting key fault features, such as the vibration frequency domain features of bearing faults and the pressure correlation features of die blockage. The attention weight calculation formula is as follows: ; in, Let be the attention weight for the i-th feature; The importance score for the i-th feature is calculated through a fully connected layer. As the feature dimension, M=12 in this system; weighted feature data for: .

[0058] 2) Multi-scale feature extraction: A multi-scale convolutional neural network (MCNN) is used, employing three sets of parallel convolutional kernels of different sizes to perform multi-scale convolution processing on the weighted feature data; the convolutional kernels are 16×1, 8×1, and 4×1 respectively; the large convolutional kernel captures low-frequency overall features, such as periodic fault patterns in vibration; the small convolutional kernel extracts high-frequency local details, such as the high-frequency impact signal of a micro-crack in a bearing; the convolution calculation method is as follows: ; in, The output features are from the convolution; These are the kernel weights; For bias terms; It is the ReLU activation function; The kernel size is used to integrate multi-scale convolutional features through a concatenation operation to obtain a spatial feature matrix.

[0059] 3) Temporal Feature Mining: The spatial feature matrix output by MCNN is input into a Bidirectional Gated Recurrent Unit (BiGRU) to capture the temporal dependencies of feature data, such as the changing trends of feature parameters during fault development, thus overcoming the limitation of traditional algorithms in capturing temporal correlations. The calculation formula for the hidden layer output of BiGRU is as follows: Reset Door: ; Update Gate: ; Candidate hidden state: ; Hidden state: ; in, This is the hidden state from the previous moment; The input features at the current time; This is the weight matrix; For bias terms; This is a dot product operation; It is the sigmoid activation function; The hyperbolic tangent activation function is used; the temporal feature matrix is ​​obtained by concatenating the outputs of the forward GRU and the backward GRU.

[0060] 4) Fault identification and localization: The spatiotemporal fusion features output by MCNN-BiGRU are input into the fully connected layer. The probability of each fault type is calculated by the softmax function. The fault type with the highest probability greater than 80% is selected as the diagnosis result. If the highest probability is less than 80%, it is judged as a suspected fault and pushed to the maintenance personnel for manual review. At the same time, the specific fault location is located by combining the correspondence between fault type and fault location in the fault knowledge base. For example, bearing wear corresponds to the front bearing / rear bearing. The fault location accuracy (≤1 component) is calculated.

[0061] Formula for calculating the softmax function: ; in, For the input feature x to belong to the th The probability of a type of failure; The first output of the fully connected layer Fault type score; The total number of fault types is C≥12 for this system.

[0062] (3) Fault level assessment module: Based on fault type, abnormal parameter amplitude and fault development speed, the analytic hierarchy process (AHP) is used to assess the fault level, which is divided into four levels: Level I is minor fault, Level II is general fault, Level III is serious fault and Level IV is emergency fault. This provides priority guidance for maintenance personnel and avoids blind repair.

[0063] First, determine the evaluation indicators and weights: select three core evaluation indicators and use the AHP method to determine the weight of each indicator; among them, the weight of the degree of fault impact is 0.5, the weight of the abnormal amplitude is 0.3, and the weight of the fault development speed is 0.2.

[0064] Then, the indicators are quantified and scored: each evaluation indicator is quantified on a scale of 1 to 4 (1 point corresponds to minor and 4 points correspond to urgent), and the specific quantification rules are shown in Table 1.

[0065] Table 1 Quantification Rules

[0066] Finally, the fault level score is calculated using a weighted summation method, as shown below: ; in, Score the fault level; Let be the weight of the i-th evaluation indicator; Let be the quantitative score of the i-th evaluation indicator; the fault level is determined based on the score as shown in Table 2.

[0067] Table 2 Fault Level Table

[0068] (4) Fault prediction module: The improved gray prediction model (GM(1,1)) is adopted. It combines historical operation data, current feature data and fault diagnosis results to predict the time of occurrence and development trend of faults, realize early warning of faults, and the early warning time is less than or equal to 1 hour. This solves the defects of existing technologies that cannot achieve fault prediction and can only be passively repaired. For hidden faults, such as early wear of bearings, the prediction accuracy is greater than or equal to 90%, and the early warning time is greater than or equal to 30 minutes, which provides sufficient maintenance time for operation and maintenance personnel.

[0069] First, data preprocessing: Select core characteristic parameters related to the fault (such as vibration peak value and average temperature value corresponding to bearing wear), extract historical data from the past 24 hours, and use the moving average method for smoothing to eliminate random interference. The smoothing formula is as follows: ; in, The data is smoothed at time t; The original feature data at time t is used, and a 5-point moving average is employed to ensure data smoothness.

[0070] Secondly, model establishment: an improved GM(1,1) prediction model is constructed, and a decay coefficient is introduced. ( (The optimal value is 0.6, calibrated using historical data), optimizing the model's prediction accuracy and addressing the issue of low prediction accuracy for nonlinear data in the traditional GM(1,1) model. The model's differential equation is as follows: ; in, Generate a sequence (cumulative generation sequence) for the 1-AGO of the feature data. The development coefficient; This represents the amount of gray action.

[0071] Attenuation coefficient The improved 1-AGO generation sequence calculation formula, used to adjust the weights of the accumulated generation, is as follows: ; in, The original feature data sequence; Generate the cumulative sequence value at time k.

[0072] Next, the parameters are solved: the least squares method is used to solve for the model parameters. and The calculation formula is as follows: ; in, For parameter vectors; To generate an accumulation matrix; The original data vector; This is a transpose operation.

[0073] Then, trend prediction and early warning: by solving differential equations, the predicted sequence of feature data is obtained. The predicted values ​​of the original feature data are generated through cumulative subtraction. The predicted value is compared with the preset fault threshold. If the predicted value exceeds the fault threshold within the next T hours (T≤1), a fault prediction report is generated, marking the predicted fault occurrence time, fault type, and fault level, and an early warning message is pushed. At the same time, the trend curve of the characteristic parameters is plotted to intuitively show the fault development trend.

[0074] Finally, model optimization: The model parameters are calibrated and updated hourly using the latest historical data to ensure prediction accuracy; when the prediction error exceeds 10%, the attenuation coefficient is automatically adjusted. The parameters are then re-solved to optimize the prediction model.

[0075] V. Terminal Interaction Layer.

[0076] The terminal interaction layer includes industrial control terminals, mobile operation and maintenance terminals (mobile APP, tablet) and Web monitoring terminals, supporting multi-terminal collaboration, providing human-machine interaction interfaces for operation and maintenance personnel, enabling them to view diagnostic results, receive early warnings, perform remote operations and manage data, and adapt to the operation and maintenance needs of different scenarios.

[0077] The specific functions of the terminal interaction layer include: information display, early warning push, remote operation, data management and access control.

[0078] Information Display: Real-time display of the soil granulator's operating parameters, environmental parameters, and granulation quality parameters; display of fault diagnosis results, fault prediction reports, and fault trend curves; display of equipment operating status, including normal, abnormal, warning, and fault; supports centralized display of multiple granulators, and allows switching to view detailed information of a single device.

[0079] Warning push notifications: Through various means such as sound, light, pop-up windows, SMS, and APP push, we push fault warnings, fault alarms, and communication anomaly warnings, mark the warning level, and ensure that maintenance personnel receive them in a timely manner.

[0080] Remote operation: Maintenance personnel can issue remote operation commands through terminals, including equipment start / stop commands and parameter adjustment commands, such as data acquisition frequency, early warning threshold, and fault reset commands. After the commands are issued, the execution results are fed back in real time to ensure the effectiveness of the operation. For emergency faults, shutdown commands can be issued remotely to prevent the fault from escalating.

[0081] Data Management: Supports querying, statistics, and exporting historical operation data, fault diagnosis data, fault prediction data, and maintenance records. It can be filtered by time, equipment number, and fault type. It also supports the entry of maintenance records, including fault handling time, handling personnel, handling plan, and handling results, providing data support for fault knowledge base updates and model optimization.

[0082] Access Control: The RBAC access control model is adopted to set different levels of operation and maintenance permissions, including administrators, senior operation and maintenance personnel, and ordinary operation and maintenance personnel. Administrators have full operation permissions, senior operation and maintenance personnel have diagnostic, prediction, and remote operation permissions, and ordinary operation and maintenance personnel only have information viewing and maintenance record entry permissions. Operation logs are retained for more than 5 years, are traceable, and ensure system security.

[0083] Example 2 like Figure 2 As shown, based on the prediction system proposed above, this embodiment also proposes a remote fault diagnosis and prediction method for soil granulators, including the following steps: Step S1: System Deployment and Initialization.

[0084] Step S11: Deploy each sensor of the sensing and acquisition layer at the corresponding part of the soil granulator according to the preset position, and complete the sensor calibration; Step S12: Deploy the edge gateway of the edge computing layer in the field control box and complete the communication connection with the sensor; Step S13: Deploy the cloud server for the cloud diagnostic prediction layer, build a fault knowledge base, and initialize algorithm parameters, including convolution kernel size, initial value of attention weights, and prediction model decay coefficient. Step S14: Deploy various terminals in the terminal interaction layer and complete permission allocation and network configuration; Step S15: After system initialization, preset the normal threshold and fault threshold of each parameter, and calibrate basic parameters such as acquisition frequency and transmission frequency.

[0085] Step S2: Multi-dimensional data collection.

[0086] The sensors in the sensing and acquisition layer collect the operating parameters, environmental parameters, and granulation quality parameters of the granulator in real time at a preset acquisition frequency. The raw data is stored in the local cache module to avoid data loss due to network interruption. At the same time, the raw data is transmitted to the edge computing layer in real time.

[0087] Step S3: Edge preprocessing and preliminary identification.

[0088] Step S31: The edge computing layer receives the raw collected data and sequentially completes three levels of preprocessing: outlier removal, noise reduction, and standardization to eliminate interference factors. Step S32: Subsequently, extract three types of feature data: time domain, frequency domain, and correlation. Step S33: Based on a preset threshold, perform preliminary anomaly identification on the preprocessed data and feature data. If anomalies are found, generate a preliminary anomaly signal and push a local warning. Step S34: Simultaneously, the preprocessed data, feature data, and preliminary abnormal signals are packaged and encrypted and transmitted to the cloud diagnostic prediction layer through the remote transmission layer.

[0089] Step S4: Cloud-based diagnostic prediction and instruction generation.

[0090] Step S41: The cloud-based diagnostic prediction layer receives data transmitted from the edge computing layer. The fault diagnosis module uses the improved MCNN-BiGRU algorithm, combined with the fault knowledge base, to complete the identification of fault type and location of fault location. Step S42: The fault level assessment module uses the analytic hierarchy process (AHP) to assess the fault level. Step S43: The fault prediction module uses an improved GM(1,1) model to predict the fault occurrence time and development trend. Step S44: The system integrates the diagnostic results, level assessment results, and prediction reports to generate operation and maintenance instructions (such as maintenance guidance, parameter adjustment, and downtime warning), and feeds back the diagnostic results, prediction reports, and operation and maintenance instructions to the edge computing layer and terminal interaction layer through the remote transmission layer.

[0091] Step S5: Interaction Control and Closed-Loop Optimization.

[0092] Step S51: The terminal interaction layer receives diagnostic results, prediction reports, and operation and maintenance instructions, and pushes early warning information through various means; Step S52: The maintenance personnel check the relevant information and perform on-site repairs or remote operations (such as remote parameter adjustment or remote shutdown) according to the fault level and maintenance instructions. Step S53: After the maintenance is completed, the maintenance personnel enter the maintenance record through the terminal; Step S54: The cloud-based diagnostic prediction layer updates the fault knowledge base and optimizes the diagnostic prediction algorithm parameters based on maintenance records. Step S55: The edge computing layer adjusts the acquisition frequency, preprocessing parameters, etc., according to the instructions fed back from the cloud, to achieve closed-loop optimization of the system and continuously improve the accuracy of diagnosis and prediction and the efficiency of operation and maintenance.

[0093] During the operation of the system of this invention, the remote transmission layer maintains bidirectional communication between the edge computing layer and the cloud diagnosis and prediction layer in real time. The heartbeat mechanism monitors the communication status in real time. If communication is interrupted, the edge computing layer continues to complete local preprocessing and preliminary early warning, caches data locally, and automatically synchronizes it to the cloud after communication is restored to ensure continuous and stable operation of the system. The fault knowledge base is automatically updated every week, and the diagnostic prediction model calibrates parameters every hour to continuously optimize system performance.

[0094] Therefore, the present invention adopts the above-mentioned remote fault diagnosis and prediction system and method for soil granulators. Through the synergistic effect of each module and intelligent improvement, it realizes accurate diagnosis, early prediction and remote control of soil granulator faults, greatly improves equipment operation and maintenance efficiency, reduces operation and maintenance costs, and ensures continuous and stable operation of the granulator.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A remote fault diagnosis and prediction system for a soil granulator, characterized in that: It includes a perception and acquisition layer, an edge computing layer, a remote transmission layer, a cloud-based diagnostic and prediction layer, and a terminal interaction layer, with each layer communicating with each other in sequence; The sensing and acquisition layer uses industrial-grade protected multi-type sensors to collect multi-dimensional parameters of the granulator in real time at a preset acquisition frequency, covering three major categories: equipment operation, environment, and granulation quality. The sensing and acquisition layer also includes a data caching module, which uses a local SD card to store and cache the original acquisition data to avoid data loss due to network interruption. After the network is restored, it will be automatically synchronized to the edge computing layer. The edge computing layer is deployed in the field control box of the soil granulator. It adopts an industrial-grade edge gateway, supports multi-protocol conversion, and communicates with the sensors and data cache modules of the sensing and acquisition layer to perform localized preprocessing, feature extraction and preliminary anomaly identification on the raw acquired data. The remote transport layer is used to establish a bidirectional communication link between the edge computing layer and the cloud diagnostic prediction layer to realize the transmission of data and instructions; The cloud-based diagnostic prediction layer, deployed on a cloud server, adopts a dual diagnostic prediction mode combining an improved MCNN-BiGRU hybrid intelligent algorithm with a fault knowledge base. Based on the data transmitted from the edge computing layer, it completes fault type identification, fault location, fault level assessment, and fault trend prediction. The terminal interaction layer includes industrial control terminals, mobile operation and maintenance terminals, and Web monitoring terminals. It supports multi-terminal collaboration, provides human-machine interaction interfaces for operation and maintenance personnel, and enables them to view diagnostic results, receive early warnings, operate remotely, and manage data.

2. The remote fault diagnosis and prediction system for a soil granulator according to claim 1, characterized in that, The sensing and acquisition layer includes: an operation parameter acquisition unit, an environmental parameter acquisition unit, and a granulation quality parameter acquisition unit; Operating parameter acquisition unit: includes vibration sensors, temperature sensors, current and voltage sensors, speed sensors, pressure sensors, and displacement sensors; among them, vibration sensors are deployed in the granulator bearings, pressure rollers, and motor housing to collect vibration acceleration and vibration frequency parameters; temperature sensors are deployed in the motor, bearings, gearbox, and die holes to collect the operating temperature of each component; current and voltage sensors are connected in series in the motor power supply circuit to collect the motor operating current and voltage parameters; speed sensors are deployed in the transmission system to collect the speed of the pressure rollers and spindle; pressure sensors are deployed in the granulation chamber to collect the granulation pressure; and displacement sensors are deployed between the pressure rollers and the flat die to collect the displacement between the two. Environmental parameter acquisition unit: including temperature and humidity sensors and dust sensors; deployed around the granulator to collect ambient temperature, humidity and dust concentration parameters to help determine the cause of the fault; Granulation quality parameter acquisition unit: including particle size sensor and particle strength sensor; deployed at the discharge port of granulator, it collects average particle size, particle size uniformity and particle compressive strength parameters, and uses granulation quality anomalies to help judge equipment failure; The perception and acquisition layer also includes a data caching module, which uses a local SD card for storage to cache the original acquired data, avoiding data loss due to network interruption, and automatically synchronizing to the edge computing layer after the network is restored.

3. The remote fault diagnosis and prediction system for a soil granulator according to claim 1, characterized in that, The specific implementation process of the edge computing layer includes the following steps: Step 1, Data Preprocessing: A three-level processing flow of outlier removal, noise reduction, and standardization is adopted to eliminate interference factors in the data collection process and ensure data validity; The algorithm employs wavelet thresholding to suppress noise caused by environmental dust and electromagnetic interference while preserving the effective signal. The wavelet basis is selected as the db4 wavelet, with a decomposition level of 5 layers, and the threshold... The calculation formula is: ； in, The standard deviation of noise; The number of samples to collect data; the high-frequency coefficients after decomposition are thresholded, and then the data is reconstructed through wavelet inverse transform to obtain the effective data after noise reduction; Step 2, Feature Extraction: For the preprocessed effective data, extract parameters that characterize the equipment's operating status and fault characteristics, which are divided into three categories: time domain features, frequency domain features, and correlation features; Step 2.1, Time Domain Features: For vibration, temperature, and current time series data, extract six time domain features: mean, variance, peak value, peak factor, kurtosis, and waveform factor to reflect the overall distribution and fluctuation characteristics of the data. Step 2.2, Frequency Domain Features: The vibration and current time-domain signals are converted into frequency-domain signals by Fast Fourier Transform (FFT), and three frequency-domain features, namely characteristic frequency, frequency peak value, and harmonic content, are extracted to reflect the fault characteristics of the equipment components. Step 2.3, Correlation Features: Based on the core operating conditions of the soil granulator, extract multi-parameter correlation features to assist in determining the fault type; Step 3, Preliminary Anomaly Identification: Based on the preset parameter threshold range, the preprocessed valid data and extracted feature data are calibrated using historical fault data to make threshold judgments; if a parameter or feature data exceeds the threshold range, a preliminary anomaly signal is generated, marking the anomaly parameter type, anomaly occurrence time and anomaly amplitude, and immediately pushed to the terminal interaction layer for local early warning. Simultaneously, the preprocessed data, feature data, and preliminary abnormal signals are sent to the cloud-based diagnostic prediction layer via the remote transmission layer for diagnosis and prediction.

4. The remote fault diagnosis and prediction system for a soil granulator according to claim 1, characterized in that, The implementation process of the remote transport layer is as follows: First, a hybrid transmission mode combining industrial Ethernet and wireless supplementation is adopted, with the transmission link using gigabit industrial Ethernet; Then, during data transmission, the AES-256 encryption algorithm is used to encrypt the transmitted data to prevent it from being stolen or tampered with. At the same time, a data fragmentation transmission mechanism is used to divide the data packet into multiple small data packets and mark them with sequence numbers. After the transmission is completed, they are reassembled in the cloud to ensure the integrity of data transmission. In addition, a transmission heartbeat mechanism is set up. If the cloud diagnostic prediction layer does not receive the heartbeat signal from the edge computing layer for three consecutive times, it is determined that the communication is interrupted. The communication abnormality warning is immediately pushed to the terminal interaction layer, and the local caching mechanism of the edge computing layer is triggered. The untransmitted data will be automatically synchronized after the communication is restored. Finally, the remote transmission layer also supports bidirectional command transmission. The operation and maintenance commands generated by the cloud diagnostic prediction layer are transmitted to the edge computing layer through the remote transmission layer. After receiving the commands, the edge computing layer sends them to the control system of the granulator to achieve remote control. The device status feedback information of the edge computing layer is transmitted to the cloud through the remote transmission layer to form a closed-loop control.

5. The remote fault diagnosis and prediction system for a soil granulator according to claim 1, characterized in that, The cloud-based diagnostic prediction layer includes four modules: a fault knowledge base module, a fault diagnosis module, a fault level assessment module, and a fault prediction module. (1) Fault knowledge base module: Construct a special fault knowledge base for soil granulators, integrate common fault types, fault characteristics, fault causes, fault locations and maintenance solutions for soil granulators, and support self-learning updates. (2) Fault diagnosis module adopts improved MCNN-BiGRU hybrid intelligent algorithm, introduces attention mechanism, filters key fault features, and combines feature data transmitted by edge computing layer and fault knowledge base to complete fault type identification and fault location; (3) Fault level assessment module: Based on fault type, abnormal parameter amplitude and fault development speed, the analytic hierarchy process (AHP) is used to assess the fault level, which is divided into four levels: Level I is minor fault, Level II is general fault, Level III is serious fault and Level IV is emergency fault. (4) Fault prediction module: The improved gray prediction model GM(1,1) is used to predict the time of occurrence and development trend of faults by combining historical operation data, current feature data and fault diagnosis results, so as to realize early warning of faults.

6. The remote fault diagnosis and prediction system for a soil granulator according to claim 5, characterized in that, The fault knowledge base consists of two parts: a basic library and a dynamically updated library. 1) Basic Database: 12 common faults of soil granulators are pre-entered, including: motor overload, bearing wear, die blockage, pressure roller eccentricity, transmission belt looseness, gearbox wear, electrical short circuit, abnormal granulation pressure, abnormal speed, poor particle formation, sensor failure, and material blockage; each fault type corresponds to a unique fault code, typical characteristic parameter range, fault cause, and repair steps; 2) Dynamically updated database: Based on historical diagnostic data and maintenance records, the system adds new fault types and optimizes the range of fault characteristic parameters through a self-learning algorithm. When a new type of fault that is not identified occurs, maintenance personnel can enter fault information and maintenance plans through the terminal interaction layer. The system automatically extracts the characteristic parameters of the fault and updates them to the fault knowledge base. The self-learning update frequency is once a week.

7. A remote fault diagnosis and prediction system for a soil granulator according to claim 5, characterized in that, An improved MCNN-BiGRU hybrid intelligent algorithm is adopted, combining feature data transmitted from the edge computing layer and a fault knowledge base, to complete fault type identification and fault location. The specific process is as follows: 1) Feature Input: The 12-dimensional feature data extracted from the edge computing layer, namely 6 time-domain features, 3 frequency-domain features, and 3 correlation features, are used as the algorithm input. An attention mechanism is used to weight each feature; the attention weight calculation formula is as follows: ； in, Let be the attention weight for the i-th feature; The importance score for the i-th feature is calculated through a fully connected layer. As the feature dimension, M=12 in this system; weighted feature data for: ； 2) Multi-scale feature extraction: A multi-scale convolutional neural network (MCNN) is used, employing three sets of parallel convolutional kernels of different sizes to perform multi-scale convolution processing on the weighted feature data; the convolutional kernels are 16×1, 8×1, and 4×1 respectively; the convolution calculation method is as follows: ； in, The output features are from the convolution. These are the kernel weights; For bias terms; It is the ReLU activation function; The kernel size is used; multi-scale convolutional features are integrated through a concatenation operation to obtain a spatial feature matrix. 3) Temporal feature mining: The spatial feature matrix output by MCNN is input into the bidirectional gated recurrent unit BiGRU to capture the temporal dependencies of the feature data; the temporal feature matrix is ​​obtained by concatenating the outputs of the forward GRU and the backward GRU. 4) Fault Identification and Localization: The spatiotemporal fusion features output by MCNN-BiGRU are input into the fully connected layer, and the probability of each fault type is calculated using the softmax function. Simultaneously, by combining the correspondence between fault types and fault locations in the fault knowledge base, the specific fault location is located. The softmax function calculation formula is as follows: ； in, For the input feature x to belong to the th The probability of a type of failure; The first output of the fully connected layer Fault type score; This represents the total number of fault types.

8. A remote fault diagnosis and prediction system for a soil granulator according to claim 5, characterized in that, An improved grey prediction model (GM(1,1)) is adopted, which combines historical operating data, current feature data, and fault diagnosis results to predict the timing and development trend of faults, thereby achieving early warning of faults. The specific process is as follows: First, data preprocessing: Select core characteristic parameters related to the fault, extract historical data from the past 24 hours, and smooth the data using a moving average method to eliminate random interference. The smoothing formula is as follows: ； in, The data is smoothed at time t; This represents the original feature data at time t; Secondly, model establishment: an improved GM(1,1) prediction model is constructed, and a decay coefficient is introduced. , To optimize the model's prediction accuracy, the model's differential equation is as follows: ； in, Generate sequences for the 1-AGO of the feature data; The development coefficient; Gray action quantity; attenuation coefficient The improved 1-AGO generation sequence calculation formula, used to adjust the weights of the accumulated generation, is as follows: ； in, The original feature data sequence; Generate a sequence value for the accumulation at time k; Next, the parameters are solved: the least squares method is used to solve for the model parameters. and The calculation formula is as follows: ； in, For parameter vectors; To generate an accumulation matrix; The original data vector; This is a transpose operation; Then, trend prediction and early warning: by solving differential equations, the predicted sequence of feature data is obtained. The predicted values ​​of the original feature data are generated through cumulative subtraction. The system compares the predicted value with the preset fault threshold. If the predicted value exceeds the fault threshold within the next T hours, a fault prediction report is generated, marking the predicted fault occurrence time, fault type, and fault level, and an early warning message is pushed out. At the same time, the system plots the trend curve of the characteristic parameters to intuitively display the fault development trend. Finally, model optimization: The model parameters are calibrated and updated hourly using the latest historical data to ensure prediction accuracy; when the prediction error exceeds 10%, the attenuation coefficient is automatically adjusted. The parameters are then re-solved to optimize the prediction model.

9. A remote fault diagnosis and prediction system for a soil granulator according to claim 1, characterized in that, The specific functions of the terminal interaction layer include: information display, early warning push, remote operation, data management and access control.

10. A method for a remote fault diagnosis and prediction system for a soil granulator according to any one of claims 1-9, characterized in that, Includes the following steps: Step S1, System Deployment and Initialization: Deploy the sensors of the sensing and acquisition layer to their corresponding positions on the soil granulator according to the preset locations, and complete sensor calibration; deploy the edge gateway of the edge computing layer to the field control box and complete the communication connection with the sensors; deploy the cloud server of the cloud diagnosis and prediction layer, build a fault knowledge base, and initialize algorithm parameters, including convolution kernel size, initial value of attention weight, and prediction model decay coefficient; deploy various terminals of the terminal interaction layer, and complete permission allocation and network configuration; after system initialization, preset the normal threshold and fault threshold of each parameter, and calibrate the basic parameters of acquisition frequency and transmission frequency. Step S2, Multi-dimensional Data Acquisition: Each sensor in the sensing and acquisition layer collects the operating parameters, environmental parameters, and granulation quality parameters of the granulator in real time according to the preset acquisition frequency. The raw data is stored in the local cache module to avoid data loss due to network interruption. At the same time, the raw data is transmitted to the edge computing layer in real time. Step S3, Edge Preprocessing and Preliminary Identification: The edge computing layer receives the raw collected data and sequentially completes three levels of preprocessing: outlier removal, noise reduction, and standardization to eliminate interference factors; then, it extracts three types of feature data: time domain, frequency domain, and correlation. Based on preset thresholds, preliminary anomaly identification is performed on preprocessed data and feature data. If anomalies are found, a preliminary anomaly signal is generated and a local warning is pushed. At the same time, the preprocessed data, feature data, and preliminary anomaly signal are packaged and transmitted to the cloud diagnostic prediction layer through a remote transmission layer with encryption. Step S4, Cloud-based Diagnosis Prediction and Command Generation: The cloud-based diagnosis prediction layer receives data transmitted from the edge computing layer. The fault diagnosis module uses an improved MCNN-BiGRU algorithm, combined with a fault knowledge base, to identify the fault type and locate the fault location. The fault level assessment module uses the analytic hierarchy process (AHP) to assess the fault level. The fault prediction module uses an improved GM(1,1) model to predict the fault occurrence time and development trend. The system integrates the diagnosis results, level assessment results, and prediction reports to generate operation and maintenance commands, and feeds back the diagnosis results, prediction reports, and operation and maintenance commands to the edge computing layer and terminal interaction layer through the remote transmission layer. Step S5, Interactive Control and Closed-Loop Optimization: The terminal interaction layer receives diagnostic results, prediction reports, and maintenance instructions, and pushes early warning information through various means; maintenance personnel view relevant information and perform on-site repairs or remote operations according to the fault level and maintenance instructions; after the repair is completed, maintenance personnel enter the repair record through the terminal; the cloud-based diagnostic prediction layer updates the fault knowledge base and optimizes the diagnostic prediction algorithm parameters based on the repair record; the edge computing layer adjusts the collection frequency and preprocessing parameters according to the instructions fed back from the cloud to achieve closed-loop optimization of the system.

Images