Air compression station operation and maintenance method and robot based on GRU model

Abstract

The application provides an air compression station operation and maintenance method and robot based on a GRU model. Through a collection device located in an air compression station, operation multi-source data of equipment in the air compression station is collected. A first algorithm module deployed on the collection device is used to process the multi-source data to obtain intermediate data. The intermediate data is transmitted to an analysis device located in the cloud through a communication network. A second algorithm module deployed on the analysis device is used to process the received intermediate data to obtain output conclusion data. The output conclusion data is issued to an execution device or a monitoring terminal in the air compression station to perform equipment control or operation early warning. The application solves the technical problem that it is difficult to simultaneously realize intelligent monitoring and early warning with low delay and high reliability under the premise of ensuring the safety of core operation and maintenance data of the air compression station in the prior art.

Description

Technical Field

[0001] This application relates to the fields of industrial Internet of Things and intelligent operation and maintenance technology, and in particular to an air compressor station operation and maintenance method and robot based on the GRU model. Background Technology

[0002] As the power core of industrial production facilities, the stable and efficient operation of air compressor stations is crucial to ensuring the continuity of the entire production line. Traditional air compressor station maintenance mainly relies on regular manual inspections and on-site instrument monitoring. This approach suffers from problems such as delayed early warnings, easy omissions, high labor costs, and the inability to achieve proactive maintenance. With the advancement of Industry 4.0 and smart manufacturing, IoT-based remote monitoring systems have emerged. These systems typically deploy various sensors (such as pressure, temperature, and vibration sensors) within the air compressor station to collect real-time equipment operating data. This data is then transmitted via network to a remote cloud platform for calculation and analysis, utilizing cloud-based artificial intelligence algorithms for condition monitoring, fault diagnosis, and predictive maintenance.

[0003] However, this centralized model of directly uploading raw data to the cloud for processing reveals significant security and real-time shortcomings in the specific industrial application scenario of air compressor stations. Firstly, regarding data transmission security, the operating parameters of air compressor stations are core production data for enterprises, directly reflecting production status and technological levels, and are highly sensitive. Continuously transmitting raw data over public networks, even with encryption, carries the risk of interception and decryption, posing a serious data leakage hazard. Secondly, regarding system real-time performance, industrial control requires extremely low response latency. The entire process of uploading raw data to the cloud, the cloud model completing calculations, and then issuing control commands introduces a significant network transmission latency, which may not meet the stringent requirement of millisecond-level response to sudden failures. Furthermore, models such as GRU need to maintain a hidden state representing the "memory" of historical information when processing time-series data. In traditional cloud processing models, the updating of this state depends entirely on a continuous, uninterrupted stream of raw data. Once network fluctuations cause data packet loss, the continuity of the model state is disrupted, directly affecting the accuracy of the analysis.

[0004] Therefore, existing technologies struggle to simultaneously achieve low-latency, high-reliability intelligent monitoring and early warning while ensuring the security of core operational data in air compressor stations. There is an urgent need for an innovative technological solution that can effectively address the security risks during data transmission and improve the real-time response capabilities and reliability of intelligent operation and maintenance systems. Summary of the Invention

[0005] This application provides an air compressor station operation and maintenance method and robot based on the GRU model, which is used to solve the technical problem in the prior art that it is difficult to achieve low-latency, high-reliability intelligent monitoring and early warning while ensuring the security of core operation and maintenance data of air compressor stations.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of this invention discloses an operation and maintenance method for air compressor stations based on the GRU model, comprising the following steps:

[0008] Step S1: Collect multi-source operational data of the equipment in the air compressor station using the data acquisition equipment located in the air compressor station;

[0009] Step S2: The multi-source data is processed using the first algorithm module deployed on the acquisition device to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector of the current time step;

[0010] Step S3: Transmit the intermediate data to the analysis device located in the cloud via a communication network;

[0011] Step S4: The received intermediate data is processed using the second algorithm module deployed on the analysis device to obtain output conclusion data; wherein, the second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as the hidden state input for its calculation and performs forward calculation to obtain the output conclusion data;

[0012] Step S5: Send the output conclusion data to the execution equipment or monitoring terminal in the air compressor station to perform equipment control or operation early warning.

[0013] The second aspect of this invention discloses an air compressor station operation and maintenance robot based on the GRU model, the robot comprising:

[0014] The data acquisition module is configured at the air compressor station site and is used to collect multi-source operating data of the equipment in the air compressor station.

[0015] An edge computing module is integrated into the data acquisition device and has a built-in first algorithm module for processing the multi-source data to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector of the current time step;

[0016] A data communication module is used to transmit the intermediate data to an analysis device located in the cloud via a communication network;

[0017] The cloud-based analysis module, deployed on a cloud server, has a built-in second algorithm module for processing the received intermediate data to obtain output conclusion data. The second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as the hidden state input for its calculation and performs forward calculation to obtain the output conclusion data.

[0018] The instruction execution and feedback module is configured in the air compressor station and is used to receive the output conclusion data and send it to the execution equipment or monitoring terminal to perform equipment control or operation early warning.

[0019] A third aspect of the present invention discloses a terminal, including a processor and a storage medium; characterized in that:

[0020] The storage medium is used to store instructions;

[0021] The processor is configured to operate according to the instructions to perform the steps of the method described in the first aspect.

[0022] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the method described in the first aspect.

[0023] Compared with the prior art, this application has the following advantages:

[0024] (1) By segmenting the complete GRU model, only the first part is deployed on the edge. This part of the algorithm converts high-value raw data into intermediate data without any direct business semantics in real time. This intermediate data is an abstract mathematical representation of the model, which is completely uninterpretable for third parties who cannot obtain the corresponding GRU model, thus achieving "natural encryption" of the data. The cloud only receives and processes this intermediate data, which fundamentally ensures the security of core operation and maintenance data and solves the risk of leakage of sensitive industrial data in the transmission of public networks.

[0025] (2) By performing the first-stage calculation of the GRU model at the edge, the amount of data that needs to be uploaded is greatly reduced, effectively reducing network bandwidth usage and transmission time. At the same time, after receiving intermediate data, the cloud does not need to perform complex feature extraction and state iteration calculations, but can directly perform lightweight forward calculations and quickly generate control commands, which greatly shortens the overall delay from data acquisition to command issuance, making it possible to quickly warn and control sudden failures in the air compressor station.

[0026] (3) The maintenance and iterative update process of the model's hidden state is moved to the edge of the factory's intranet, where the network conditions are more stable. The edge computing module persistently stores the hidden state locally, and its update process is completely decoupled from the cloud network conditions. Even if a temporary network interruption occurs, the state iteration on the edge side will not be affected, ensuring the accuracy of the model's "memory". After the network is restored, the cloud only needs to receive the latest state vector to immediately resume computation, enhancing the robustness and reliability of the intelligent operation and maintenance system and avoiding analysis interruptions or state distortions caused by network fluctuations. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0028] Figure 1 This is a flowchart of the air compressor station operation and maintenance method based on the GRU model provided in this embodiment of the disclosure;

[0029] Figure 2 This is a schematic diagram of the structure of an air compressor station maintenance robot based on the GRU model provided in this embodiment.

[0030] Figure 3 This is a schematic diagram illustrating the process of determining a stability graph provided in an embodiment of this disclosure.

[0031] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0033] To enable those skilled in the art to better understand and implement the present invention, the aforementioned air compressor station operation and maintenance method based on the GRU model and the air compressor station operation and maintenance robot based on the GRU model will be described in detail below through several specific embodiments from different dimensions such as system architecture, data flow redirection, and core algorithm principles.

[0034] Example 1

[0035] Figure 1A flowchart of the air compressor station operation and maintenance method based on the GRU model provided in this disclosure is shown below. Figure 1 As shown, the process may include the following steps:

[0036] Step S1: Collect multi-source data on the operation of the equipment in the air compressor station using the data acquisition equipment located in the air compressor station.

[0037] This step involves constructing a distributed sensor network capable of comprehensively and accurately sensing the operational status of the air compressor station, and providing a high-quality data source for subsequent intelligent analysis. The specific implementation process is as follows:

[0038] At the hardware deployment level, a series of industrial-grade sensors need to be installed at key process nodes in the air compressor station. A high-precision piezoelectric pressure transmitter should be installed on the main outlet pipe of the air compressor, with a measurement range covering the station's rated operating pressure (e.g., 0-1.6MPa). The output signal should be a standard 4-20mA analog current signal, which has significantly better anti-interference capabilities than voltage signals during long-distance transmission. Armored K-type thermocouples or PT100 resistance temperature sensors should be deployed at key locations such as the air compressor head, oil-gas tank, and cooling oil outlet to monitor operating temperature and prevent equipment efficiency degradation or lubrication failure due to overheating. ICP-type piezoelectric acceleration vibration sensors should be installed at the motors and bearing housings of the air compressor and dryer, attached to the equipment housing via magnetic attraction or screws, to collect vibration acceleration signals during equipment operation. Their frequency response range should cover the characteristic frequencies of potential equipment failures (typically 0-5kHz). An open-type Hall effect current sensor is integrated into the power supply circuit of the air compressor to measure its operating current non-contactly, thereby indirectly reflecting its load rate and power consumption. All sensors are connected to a distributed I / O acquisition module located in the station control cabinet via shielded cables. This module is responsible for converting analog signals into digital quantities and communicating with the industrial-grade edge computing gateway, which is the core acquisition device, via an RS-485 bus.

[0039] At the data acquisition and preprocessing level, a dedicated data acquisition service process runs within the edge computing gateway. This process polls all I / O modules at a fixed sampling period (e.g., 100ms), reading the raw data values ​​from their registers. The read raw data first enters a data cleaning buffer. Here, the first level of data validity verification is implemented: based on physical common sense and equipment nameplate parameters, a hard threshold is set to automatically remove obvious outliers caused by sensor disconnection, short circuit, or electromagnetic interference. Subsequently, the data enters a smoothing and filtering stage, using first-order hysteresis filtering or moving average filtering to effectively suppress high-frequency noise in the field and generate smoother instantaneous values ​​that better reflect trends. Next is dimensional normalization. Due to the significant differences in the dimensions and numerical ranges of data such as pressure (MPa), temperature (°C), current (A), and vibration (m / s²), they need to be mapped to a unified numerical range, such as [0,1], to facilitate efficient processing by the subsequent neural network model. This process is performed using the following formula: ,in and These are the lower and upper limits of the sensor's measurement range, respectively.

[0040] Finally, at time t, all the preprocessed data are combined into a standardized multi-source data vector. ,in, This is the system outlet manifold pressure value. This refers to the air compressor head temperature. For the air compressor cooling oil temperature, This represents the vertical vibration acceleration value of the air compressor. This represents the horizontal vibration acceleration value of the air compressor. The vector represents the operating current of the air compressor's main motor. It contains 10-15 dimensions of feature data, comprehensively characterizing the operating state of the air compressor system at time t. This vector is placed in a thread-safe circular buffer, awaiting retrieval by subsequent processing modules. Simultaneously, the raw data and timestamps are compressed and stored locally, retaining at least 30 days of historical data for post-processing analysis and model optimization.

[0041] Step S2: The multi-source data is processed using the first algorithm module deployed on the acquisition device to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector at the current time step.

[0042] In an embodiment of the present invention, the complete recurrent neural network model may be a GRU model.

[0043] This step transforms high-value raw data into an internal model state without business semantics before the data leaves the edge device, thereby achieving inherent security. The specific implementation process is as follows:

[0044] At the hardware resource configuration level, the edge computing gateway needs to reserve sufficient computing resources for the first algorithm module. This typically requires allocating a dedicated CPU core (CPU-bound via cgroups or tasksets), 1-2GB of dedicated memory, and necessary GPU / NPU acceleration resources (if the model computation is large). These resource guarantees ensure that the algorithm module can meet real-time requirements and complete all computational tasks within a 100ms period.

[0045] At the software architecture level, the first algorithm module runs as an independent service process, exchanging data efficiently with the data acquisition process via shared memory or Unix domain sockets. This service employs a multi-threaded architecture: a dedicated I / O thread is responsible for reading the latest multi-source data vectors from a shared memory circular buffer. One or more worker threads are responsible for executing the actual computational tasks; a management thread is responsible for monitoring the module's health status, maintaining state persistence, and communicating with other modules in the system.

[0046] At the model deployment and runtime level, the first algorithm module loads the model file, which has been trained and segmented in the cloud, from an encrypted storage partition upon startup. This model file contains a complete network structure definition (such as the number of GRU units and the dimension of hidden layers) as well as all weight parameters after training convergence. The model format is typically chosen to be suitable for edge deployment, such as TensorFlow Lite, ONNX Runtime, or OpenVINO IR format. These formats are heavily optimized for edge devices and can significantly improve inference efficiency.

[0047] When the preprocessed data vector Once ready, the first algorithm module is immediately triggered to perform forward inference computation. Although this process does not involve the specific GRU gating calculation formula, its computational logic still needs to be described in detail: the module receives... As input, it also reads historical state information generated in the previous computation cycle from persistent storage (such as eMMC or FRAM). Then, through a series of complex matrix multiplications and nonlinear transformations (the specific forms of these transformations are determined by the trained model weights), a completely new state representation vector for the current time step is generated. This vector is typically 128-dimensional or 256-dimensional, with each dimension being a 32-bit floating-point number, together forming a highly abstract representation of the current state of the system.

[0048] After the calculation is complete, the module performs two key operations: First, it converts the newly generated state vector into a vector. As intermediate data output for this calculation, it is passed to the communication service module using zero-copy technology, ready for transmission; secondly, the new state vector is used locally. Overwriting old historical states requires ensuring atomicity in the state update operation to prevent interference from read operations during the update process. This is typically achieved through double buffering or read-write lock mechanisms. State persistence employs a write-through strategy, immediately writing each calculated state to non-volatile memory to prevent state loss due to unexpected system power outages.

[0049] At this point, the original, easily readable runtime data has been transformed into a highly abstract state vector lacking physical meaning. For any third party without knowledge of the corresponding GRU model, this intermediate data... Completely uninterpretable, with each dimension's value not corresponding to any specific physical quantity, achieving a "natural encryption" security effect. At the same time, this approach has significant advantages over traditional encryption algorithms (such as AES and RSA): low computational overhead, no additional latency introduced; no key distribution and management issues; and security based on the black-box nature of the model itself, rather than the complexity of mathematical problems.

[0050] Step S3: Transmit the intermediate data to the analysis device located in the cloud via a communication network.

[0051] This step establishes a secure and reliable data transmission channel to reliably deliver the "ciphertext" state generated by edge computing to the cloud. The specific implementation process is as follows:

[0052] At the network connectivity level, edge computing gateways need to establish a stable connection with the cloud. Depending on the on-site network conditions, several connection methods can be chosen: wired industrial Ethernet is preferred, providing bandwidth of over 100Mbps and millisecond-level latency; in situations without cabling, 4G / 5G cellular networks can be used, connected via a built-in industrial-grade cellular module (supporting automatic switching between multiple carrier SIM cards); in extremely remote areas, satellite communication can even be considered as a backup. Regardless of the method used, a VPN tunnel (such as IPsec or OpenVPN) or a carrier's APN leased line service is required to provide better isolation and security than the public internet.

[0053] At the data encapsulation level, the communication service module within the edge computing gateway obtains the intermediate data to be sent from the first algorithm module. This module first... Network transmission encapsulation follows a layered protocol stack principle: at the application layer, the encapsulation process... Each floating-point number is serialized into a byte stream in big-endian order. An application layer protocol header is appended to this byte stream, containing a collection timestamp 't' accurate to milliseconds, a unique gateway device ID, a data packet sequence number, and a CRC32 checksum. At the transport layer, TCP is chosen over UDP. Although UDP has lower overhead, TCP's reliable connections, packet acknowledgments, timeout retransmissions, and flow control mechanisms are crucial for ensuring the continuous, error-free, and in-order delivery of intermediate state data. At the network layer, IP is used for routing. At the link layer, the data is encapsulated into Ethernet frames or PPP frames depending on the actual network type.

[0054] At the transmission optimization level, several optimization measures are implemented to address the characteristics of intermediate data: lossless compression of application layer payloads is performed using data compression algorithms (such as zlib or lz4), typically achieving a compression rate of over 50%; packet aggregation is implemented to package state vectors from multiple time points into a single large data packet for transmission, reducing protocol overhead; dynamic QoS policies are set to assign the highest priority to state data, ensuring priority transmission during network congestion; and adaptive bitrate adjustment is implemented to dynamically adjust the transmission frequency and data precision based on network conditions, reducing the transmission frequency or using lossy compression when network quality is poor.

[0055] At the reliability level, end-to-end integrity protection is implemented: at the gateway, a sending buffer and retransmission mechanism are maintained to ensure that each data packet receives an ACK confirmation from the cloud; at the cloud, data packet sequence number checking and reordering are implemented to prevent data packet out-of-order processing due to network jitter; a heartbeat mechanism and connection status monitoring are set up to quickly rebuild the connection and restore the data flow when the connection is interrupted. Simultaneously, security measures are implemented: firewalls are deployed at all network interfaces, allowing communication only with designated ports of the cloud authentication server; two-way certificate authentication is supported to prevent man-in-the-middle attacks; and transport layer encryption (such as TLS 1.3) is implemented. Although the intermediate data itself already has "natural encryption" characteristics, the additional encryption layer can prevent traffic analysis attacks.

[0056] Even if data packets are maliciously intercepted during the entire transmission process, attackers will not be able to extract information from the intermediate data. The system can extract any original business information (such as pressure or temperature values ​​at time t) through parsing, yielding only an incomprehensible mathematical vector. This ensures the security of core operational data at the transmission level. This security mechanism offers significant advantages over traditional data encryption methods, avoiding the risk of encryption algorithms being cracked, reducing the computational overhead of encryption and decryption, and eliminating the complexity of key distribution and management.

[0057] Step S4: The received intermediate data is processed using the second algorithm module deployed on the analysis device to obtain output conclusion data; wherein, the second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as the hidden state input for its calculation and performs forward calculation to obtain the output conclusion data.

[0058] This step leverages powerful cloud computing capabilities to deduce the final operational decision based on the received intermediate states. The specific implementation process is as follows:

[0059] Cloud analytics devices are typically clustered cloud servers with a load balancer at the front end to handle connections from a large number of edge devices. The load balancer then routes packets to a dedicated analytics service instance. This instance runs a second algorithm module that loads a second GRU sub-model corresponding to the edge front end. This model is the latter half of the complete GRU model and typically includes possible subsequent layers of GRU units (such as another GRU layer or an LSTM layer) and a final task-specific output layer (such as a fully connected layer with Softmax for classification, or a fully connected layer for regression).

[0060] After receiving the data packet, the service instance first parses the application layer header, verifies the data integrity, and extracts the clean intermediate data. Subsequently, the second algorithm module will It is loaded as the initial hidden state for its own GRU computation. Since the model is trained as a whole, the cloud module can seamlessly understand and utilize it. This refers to the abstract temporal characteristics extracted from the raw data. In many predictive maintenance scenarios, cloud modules may... Starting with zero input or given future hypothetical input, it autonomously performs multi-step iterative reasoning (e.g., calculating...). This is used to predict the evolution of the system's state over a future period. Ultimately, the inferred final hidden state (e.g., ...) is used to predict the system's state evolution over a future period. It is passed to the output layer.

[0061] The output layer is typically one or more fully connected (dense) layers, with a weight matrix of... , bias is For the fault classification task, the calculation is as follows: Output a probability distribution representing the probability of the system being in various fault modes (e.g., bearing normal, bearing slightly worn, bearing severely damaged). For regression tasks like pressure control, the calculation is as follows: It directly outputs a control quantity. In this embodiment, the conclusion is... It could be a specific control command, such as a structured JSON data: This means "increase the outlet valve opening by 3% within the next 5 seconds." Throughout the entire cloud computing process, the cloud server never accessed any raw, sensitive data such as pressure or temperature, yet it still arrived at a precise control conclusion.

[0062] Step S5: Send the output conclusion data to the execution equipment or monitoring terminal in the air compressor station to perform equipment control or operation early warning.

[0063] This translates intelligent decisions made in the cloud into actual control actions in the physical world or clear warnings to operators. The specific implementation process is as follows:

[0064] At the command issuance level, the cloud server generates the conclusion data. Then, it is encapsulated into a downlink command packet. The encapsulation format is usually a lightweight structured data format (such as Protocol Buffers or MessagePack), which has better serialization efficiency and lower network overhead compared to JSON. The command packet contains complete operation instructions, target device identifier, execution parameters, validity period, digital signature, and other information. Downlink communication adopts the MQTT protocol publish / subscribe model, where the cloud acts as the publisher to publish instructions to a specific topic ( / edge / device_id / command), and the edge gateway subscribes to this topic to receive instructions. This model supports one-to-many broadcasting and offline message caching, making it suitable for industrial IoT scenarios.

[0065] At the security verification level, the command receiving service within the edge gateway listens for MQTT topic messages. Upon receiving a command, it performs multi-layered security verification: first, it verifies the digital signature to ensure the command's source is trustworthy and has not been tampered with; then, it checks the command's validity period and rejects expired commands; next, it verifies the security of the execution parameters to ensure all operating parameters are within the device's allowed safe range (e.g., valve opening commands do not exceed 100%); finally, it performs operation conflict detection to prevent the simultaneous execution of contradictory operation commands (e.g., simultaneously opening and closing the valve).

[0066] At the control level, verified instructions are parsed into specific device control commands. The gateway sends control signals to the corresponding actuators via industrial control buses (such as Modbus TCP, PROFINET, EtherCAT, etc.). For analog devices, the gateway outputs 4-20mA control signals through analog output modules; for digital devices, it outputs switch signals through digital output modules; for intelligent devices, it writes signals to the control register via the corresponding industrial protocol (such as Modbus TCP). All operations are logged in detail, including operation time, operation content, operation result, operator (for system-automatic operations), and other information.

[0067] At the feedback confirmation level, the system implements a closed-loop control strategy: sensors on the executing device (such as valve position feedback sensors) monitor the command execution results in real time and return the actual status to the edge gateway; the gateway compares the command requirements with the actual status, and if the deviation exceeds the allowable range, it automatically performs compensation adjustments or retryes the operation; simultaneously, the execution results are fed back to the cloud through the uplink channel, forming a complete control closed loop. For important control operations, the system also supports a manual confirmation mechanism, requiring maintenance personnel to confirm the operation a second time before execution to prevent malfunctions of the automatic control system.

[0068] For early warning conclusions, the system implements a multi-layered alarm strategy: at the local level, the edge gateway triggers audible and visual alarms to provide on-site alerts; at the monitoring center level, the early warning information is pushed to the HMI monitoring system and displayed prominently on the monitoring screen; at the mobile terminal level, the early warning information is pushed to relevant maintenance personnel via SMS, WeChat, email, etc.; at the management system level, maintenance work orders are automatically created, and maintenance plans and resource allocation are arranged. All early warning information includes detailed contextual information, including the current status of the equipment, the prediction result, the confidence level, and suggested handling measures, helping maintenance personnel make quick and accurate decisions.

[0069] Through the complete technical implementation described above, this invention transforms cloud-based artificial intelligence decision-making into actual control over the physical world or effective operation and maintenance guidance, realizing a complete intelligent operation and maintenance closed loop from perception to decision-making to execution. The entire system not only ensures data security but also provides reliable execution guarantees, ensuring that intelligent operation and maintenance decisions can truly be implemented and generate value.

[0070] Example 1 details the complete application process of the method of the present invention in the pressure control scenario of an air compressor station, demonstrating a safe closed loop from data acquisition to control execution. The implementation of this process fundamentally relies on the ingenious segmentation of the complete GRU prediction model and its placement at both the edge and cloud ends. To enable those skilled in the art to fully understand and reproduce this core technology, this example will delve into its internal principles through Example 2, detailing the model segmentation implementation method based on the iterative characteristics of GRU neurons.

[0071] Example 2

[0072] Within the technical framework of this invention, the edge side only undertakes the preliminary computation tasks of the complete GRU model, that is, completing the transformation from the original multi-source operating condition data to the hidden state; while the cloud continues to perform inference on the hidden state to arrive at the final maintenance conclusion. The reason for adopting this "front-end and back-end decoupling" approach is not simply a difference in deployment strategy, but rather stems from the inherent limitations of the time-series computation mechanism of cyclic structure models.

[0073] It is important to emphasize that GRU is a typical recurrent neural network structure, and its time progression mechanism is drastically different from that of conventional feedforward networks. Compared to feedforward models like CNNs or MLPs, where all inputs must arrive simultaneously before a complete forward operation can be performed, GRU, by definition, allows for "asynchronous time steps." This means that as soon as the current input segment arrives, the update and reset gates will trigger the computation of the next hidden state, without waiting for updates from other modalities or delayed input sequences.

[0074] From an engineering perspective, the operation of a GRU can be viewed as a "local iteration sliding along the time axis." Assuming that different sensors, such as those for pressure, temperature, vibration, and current, have their own different refresh rates or data stabilization periods, the GRU will prioritize absorbing the earliest updated data features and accumulate them in the hidden state. As the time step continues, time steps of different modes may become misaligned: some modes, due to their higher update frequency, will leave deeper traces in the hidden state more quickly; while modes with longer update cycles may not have an effective impact on the hidden state for a period of time.

[0075] This characteristic is not a bug, but rather the natural operating logic of recurrent models: the hidden state may gradually shift towards the "faster-updating data source" during iteration. Therefore, when the model is split on an edge device, if the original sensor data is directly uploaded to the cloud and the cloud reconstructs the temporal relationships, the internal state evolution trajectory of the cloud GRU will be completely different from that on the edge side, ultimately leading to a mismatch in inference results. This is the fundamental reason why recurrent structure models are difficult to "reproduce without loss".

[0076] In some embodiments, each computational unit in a recurrent neural network can be considered as a node, and the logical directional relationships between all nodes can be determined through directed edges, thereby forming a directed graph. It should be noted that the term "edge" in this invention refers specifically to directed edges. However, due to the existence of iteration, the constructed directed graph cannot determine logical relationships other than directional relationships. More seriously, it cannot fully describe the directed graph because it is inherently dynamic. When external parameters (e.g., changing the iteration count of a module from 3 to 5) change, although the static directed graph appears unchanged on the surface, it has actually changed.

[0077] Suppose that in a connected iterative module, at least two nodes have inputs as stimuli, such as the first input and the second input. Then, the time-step misalignment mentioned above will occur. This is because if the first input's time step is faster, it becomes impossible to determine how many iterations the second input was added to the connected iterative module after the first input. Therefore, from a global perspective, the first and second algorithm modules at different ends effectively eliminate time steps, causing the iteration order in the second algorithm module to change, leading to different results. A connected iterative module refers to a set of algorithm modules where every node participates in the iteration and is bidirectionally connected. Understandably, nodes participating in the iteration must be bidirectional nodes, meaning they can iterate back to themselves along a path. Bidirectional connectivity means that any two nodes in the connected iterative module can reach each other along a directed edge path. To ensure consistency, the first and second algorithm modules can be split based on the concept of flow in graph algorithms. First, identify all connected iterative modules; second, expand the connected iterative modules until the time steps are eliminated by unidirectional nodes to form a stable graph; finally, identify cuts in the stable graph to distinguish between the first and second algorithm modules; it is understood that any cut in the stable graph meets the requirements.

[0078] exist Figure 3 In the connected iterative module u1, there are nodes a1, a2, a3, and a4. It is assumed that nodes a1 and a2 share two inputs as stimuli. Expanding the connected iterative module until a time step is eliminated by a unidirectional node means that these two inputs are used as stimuli by the same unidirectional node, thus canceling out time step errors. For convenience, the unidirectional node that cancels out time step errors is called a cutoff node. In this invention, it is broadly defined as a node with paths to all input stimuli of the connected iterative module. However, it should be understood that a cutoff node can actually have a more refined definition. Bidirectional nodes refer to nodes in the connected iterative module; while unidirectional nodes are non-bidirectional nodes. Figure 3As shown, the cutoff node corresponding to the connected iteration module can be b3. For each connected iteration module, the expanded nodes are merged until the cutoff node is reached. Therefore, the resulting stable graph is as indicated by the arrow. It can be understood that the expanded nodes are exactly the set of nodes on a certain path from the cutoff node to any bidirectional node of the connected iteration module, but do not include the cutoff node. The expanded nodes form new nodes and inherit the connection relationships to form a stable graph. Figure 3 If b3 is chosen as the termination node, then b4~b6, as well as the four nodes of the connected iterative module u1 itself, will all form a new node A1. Understandably, there can be multiple termination nodes. For example, if b2 is chosen as the termination node, then the expanded nodes include b3~b6 and the four nodes of the connected iterative module u1 itself, forming a new node.

[0079] In some embodiments, to minimize expansion and enhance robustness by avoiding an insufficient number of nodes in the stable graph, all cutoff nodes corresponding to the connected iteration module can be obtained. The expanded nodes are precisely the set of nodes in the paths from any cutoff node to any bidirectional node of the connected iteration module. That is, assuming a node 'a' to be expanded satisfies the condition that 'a' is included in the path from cutoff node 'b' to the connected iteration module, but 'a' is not included in any path from cutoff node 'c' to the connected iteration module, then 'a' is not a node to be expanded. In this way, the number of expanded nodes is greatly reduced.

[0080] It's important to note that in practical algorithm implementation, a theoretical extreme case needs to be emphasized: when the dependencies within the model exhibit a completely indivisible, strongly coupled structure, stable functional boundary partitioning (i.e., cuts in the stable graph: the set of directed edges between the first and second algorithm modules) may not exist. In this case, the parameter flow and state transfer between the first and second algorithm modules cannot form a separable topological partition, leading to unstable implementation of module decomposition and independent scheduling. However, from the perspective of computer architecture and deep learning model implementation, this situation is extremely rare. The reasons include: firstly, recurrent neural networks are typically implemented using Python frameworks (such as PyTorch and TensorFlow) in engineering practice, and their computational graph construction naturally follows a modular design. Forward propagation, backward gradient calculation, and state updates are all organized in the form of independent operator nodes and tensor flows, and the dependencies between operators are explicitly maintained by the computational graph, making it difficult to form an indivisible globally coupled structure. On the other hand, modern deep learning frameworks employ hierarchical module management mechanisms during the model definition phase, such as class inheritance structures, submodule registration systems, independent weight namespaces, and parameter mapping tables. These mechanisms ensure the separability of each functional component both logically and in memory. Therefore, even if the model develops strong feature interactions during learning, stable decomposition can still be achieved through graph truncation, node pruning, or state isolation. In summary, although theoretically there are highly coupled cases where cuts cannot form in a stable graph, in current mainstream Python modular recurrent neural network implementations, such globally indivisible structures are almost impossible to achieve. In engineering practice, module partitioning can be considered stable and feasible.

[0081] Example 3

[0082] This embodiment aims to elaborate on the core algorithm principle upon which the present invention relies, and to explain in detail how to utilize the inherent iterative computation characteristics of the gated recurrent unit (GRU) to securely divide a fully trained single neural network model into two physically separate but logically coordinated algorithm modules, and deploy them to edge devices and cloud servers respectively. However, it should be noted that this embodiment 3 is not the innovation of the present invention, because the GRU algorithm is not original to the present invention, and its description is only for the purpose of improving the content.

[0083] The GRU model of this invention adopts a single-layer GRU + fully connected output layer structure, with the hidden state dimension set to 128 and the input vector dimension set to 32. During training, the MSE loss function (for regression tasks) or the cross-entropy loss function (for classification tasks) is used, and the Adam optimizer is employed. The initial learning rate is 0.001, and the learning rate is decayed every 5 epochs by a decay factor of 0.9. The batch size is set to 64, the number of training epochs is set to 120, the activation function of GRU is tanh, the gating function is sigmoid, the weights are initialized using Xavier uniform initialization, and the biases are initialized to zero vectors.

[0084] The theoretical basis of this invention lies in the unique "memory" mechanism and state iteration characteristics of the GRU neural network model. GRU utilizes its internal hidden state vectors... It is used to remember and transmit historical information in sequence data. Its core computational logic is a recursive process, namely the hidden state at the current moment. The calculation depends on the current external input. and the hidden state of the previous moment Its functional relationship can be abstractly represented as: This mathematical property means that, It is a crucial variable that connects the past and the future: it is both a highly abstract and condensed representation of all past input sequence information, and the unique starting point for calculating all future outputs. It is precisely this characteristic that makes it possible to... This makes model segmentation possible. After segmenting the model at this point, the first algorithm module at the edge is responsible for receiving the raw data. And combined with its own maintenance Calculate The second algorithm module in the cloud only needs to receive... This can be used as the initial state for its own calculations, and all subsequent calculations can be completed. It is a high-dimensional floating-point vector and does not have any direct business semantics, thus naturally ensuring the security of data transmission while achieving functional continuity.

[0085] Model segmentation is an offline operation performed using a mature deep learning framework. First, during model development, frameworks such as TensorFlow and PyTorch are used to train a complete GRU time-series prediction model on a dataset containing a large amount of historical operational data from air compressor stations. This model typically consists of an input layer, one or more GRU units, and a final task-specific output layer (such as a fully connected layer). After the model training is complete and converges, logical segmentation can be performed either in the serialized model file or through the framework's model object interface. Specifically, the model is segmented at the output state of the last GRU layer. Taking TensorFlow as an example, when building the model, the `return_sequences=False` and `return_state=True` parameters of the GRU layer can be set to ensure that the layer returns its final hidden state along with the output sequence result. Then, the model is reconstructed using the framework's API (such as Keras' Functional API): the original model's state from the input layer to the output state of the last GRU layer is reconstructed. The portion is reconstructed into a new model, called the first GRU sub-model, whose output is... Next, with As an input layer, all layers following the GRU layer in the original model (such as fully connected layers) are used to construct another new model, called the second GRU sub-model. Finally, these two sub-models are exported as independent model files (such as SavedModel or ONNX format) for deployment.

[0086] The real-time computation of the first algorithm module in the edge computing device is essentially the specific execution of the GRU forward inference process. Its detailed mathematical derivation is as follows. For each input vector... The GRU unit calculates the new hidden state through a gating mechanism, which involves the following four core formulas:

[0087] ① Update Gate Calculation: ,in, This indicates that two vectors are concatenated. and It updates the weight matrix and bias vector corresponding to the gate. It's the sigmoid activation function. (Update gate) It determines how much past information is available. It needs to be preserved in the new state.

[0088] ② Reset Gate Calculation: Reset the door The calculation method is similar to that of the update gate, and it has its own parameters. and It controls how much past information needs to be "forgotten" in order to calculate candidate states.

[0089] ③ Calculation of Candidate Hidden State: ,in, This represents element-wise multiplication (Hadamard product). Reset gate. First, compare with the previous state Multiply the results to simulate selective "forgetting." Then multiply the result with the current input. splicing, and then going through a weighted process and bias The parameterized linear transformation is followed by the tanh activation function to obtain the candidate states. It represents the new state suggested based on the current input and some past states.

[0090] ④ Final hidden state calculation: This is the final update step. Update portal As a selector, in the previous state (Representing past memories) and candidate states Linear interpolation is performed between (representing the current proposal) to generate a new hidden state. At this point, a complete GRU computing unit is finished.

[0091] To ensure the continuity of the temporal context, the first algorithm module at the edge must persistently maintain the hidden state. During system initialization, It is initialized as a zero vector. After each calculation, the newly generated... It will be immediately written to a fixed storage area in the edge device's non-volatile memory and read back as a function during the next computation. Usage. This process needs to guarantee atomicity, typically achieved through mutexes or transactional memory operations, ensuring consistency and correctness of the state in a multi-threaded environment. The computation process in the cloud is relatively simple. The second algorithm module receives data sent from the edge. Then, it is used as the initial hidden state input for its own model. This module is usually composed of fully connected layers, and its computation process is represented as follows: (For classifying tasks) or (For regression tasks), where and This refers to the pre-trained weights and biases. Through this partitioning and collaboration mechanism, the cloud can achieve this without needing to know any original sensitive data, relying solely on the received abstract state. This allows for accurate prediction or classification, and the output of final control conclusions. This achieves a high degree of balance between safety and efficiency.

[0092] Example 4

[0093] Figure 2 This is a schematic diagram of a performance evaluation system for a current transformer provided according to an embodiment of this disclosure. (Refer to...) Figure 2 The robot includes:

[0094] The data acquisition module is configured at the air compressor station site and is used to collect multi-source operating data of the equipment in the air compressor station.

[0095] An edge computing module is integrated into the data acquisition device and has a built-in first algorithm module for processing the multi-source data to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector of the current time step;

[0096] A data communication module is used to transmit the intermediate data to an analysis device located in the cloud via a communication network;

[0097] The cloud-based analysis module, deployed on a cloud server, has a built-in second algorithm module for processing the received intermediate data to obtain output conclusion data; wherein, the second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as hidden state input and performs iterative calculations;

[0098] The instruction execution and feedback module is configured in the air compressor station and is used to receive the output conclusion data and send it to the execution equipment or monitoring terminal to perform equipment control or operation early warning.

[0099] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute the methods described above.

[0100] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided by the above methods.

[0102] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0103] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0104] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0105] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0106] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0107] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0108] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for the operation and maintenance of air compressor stations based on the GRU model, characterized in that, Includes the following steps: Step S1: Collect multi-source operational data of the equipment in the air compressor station using the data acquisition equipment located in the air compressor station; Step S2: The multi-source data is processed using the first algorithm module deployed on the acquisition device to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector of the current time step; Step S3: Transmit the intermediate data to the analysis device located in the cloud via a communication network; Step S4: The received intermediate data is processed using the second algorithm module deployed on the analysis device to obtain output conclusion data; wherein, the second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as the hidden state input for its calculation and performs forward calculation to obtain the output conclusion data; Step S5: Send the output conclusion data to the execution equipment or monitoring terminal in the air compressor station to perform equipment control or operation early warning; The first and second algorithm modules are divided based on the following method: all connected iterative modules are identified; the connected iterative modules are expanded until the time steps are eliminated by unidirectional nodes to form a stable graph; the cuts in the stable graph are identified to divide the first and second algorithm modules; wherein, a connected iterative module refers to a set in which every node in the algorithm module participates in the iteration and is bidirectionally connected, and each edge in the cut in the stable graph has at least one unidirectional node, and a unidirectional node refers to a node in a non-connected iterative module.

2. The air compressor station operation and maintenance method based on the GRU model according to claim 1, characterized in that, The complete recurrent neural network model is a GRU model; the first algorithm module consists of the part of the GRU model from the input layer to the last GRU layer; the second algorithm module consists of all the layers in the GRU model that follow the last GRU layer.

3. The air compressor station operation and maintenance method based on the GRU model according to claim 2, characterized in that, The second algorithm module uses the received intermediate data as the initial hidden state and autonomously performs multi-step iterative reasoning in the absence of subsequent original data input to predict the future operating status of the air compressor station equipment.

4. The air compressor station operation and maintenance method based on the GRU model according to claim 1, characterized in that, The connected iteration module is expanded until the time step is eliminated by the unidirectional node to form a stable graph. Specifically, the expanded nodes are the set of nodes on a certain path from the cutoff node to any bidirectional node of the connected iteration module, but do not include the cutoff node. The expanded nodes form new nodes and inherit the connection relationship to form a stable graph. The cutoff node is any unidirectional node that exists on all input stimuli of the connected iteration module, and the unidirectional node is a node in the non-connected iteration module.

5. The air compressor station operation and maintenance method based on the GRU model according to claim 4, characterized in that, The expanded nodes are exactly the set of nodes in the path from any cutoff node to any bidirectional node of the connectivity iteration module, wherein the cutoff node corresponds to the connectivity iteration module.

6. A GRU-based air compressor station operation and maintenance robot, used to perform the method described in any one of claims 1-5, characterized in that, The robot includes: The data acquisition module is configured at the air compressor station site and is used to collect multi-source operating data of the equipment in the air compressor station. An edge computing module is integrated into the data acquisition device and has a built-in first algorithm module for processing the multi-source data to obtain intermediate data; wherein, the first algorithm module is the first part of a pre-trained complete recurrent neural network model, and the intermediate data is the hidden state vector of the current time step; A data communication module is used to transmit the intermediate data to an analysis device located in the cloud via a communication network; The cloud-based analysis module, deployed on a cloud server, has a built-in second algorithm module for processing the received intermediate data to obtain output conclusion data. The second algorithm module is the latter part of the complete recurrent neural network model, which uses the received intermediate data as the hidden state input for its calculation and performs forward calculation to obtain the output conclusion data. The instruction execution and feedback module is configured in the air compressor station and is used to receive the output conclusion data and send it to the execution equipment or monitoring terminal to perform equipment control or operation early warning.

7. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method described in any one of claims 1-5.

Images