Intelligent monitoring system for traditional Chinese medicine extraction process based on edge calculation

Abstract

The invention relates to the technical field of industrial control systems and Internet of Things operating systems, and particularly discloses an intelligent monitoring system for a traditional Chinese medicine extraction process based on edge computing. According to the system, a three-layer architecture composed of a field sensing layer, an edge intelligent layer and a cloud collaboration layer is constructed. The edge intelligent layer is deployed in a production field, performs real-time preprocessing, multi-scale feature extraction and lightweight intelligent diagnosis on original data of equipment, and realizes millisecond-level anomaly perception and local early warning; and the cloud collaboration layer is responsible for model optimization and global data analysis. According to the method, core calculation is sunk to the edge, the network load and delay are reduced, and the capturing capacity and the early warning real-time performance of early weak fault features of equipment are improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial control systems and Internet of Things operating system technology, specifically relating to an intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing. Background Technology

[0002] In the field of industrial automation and intelligent manufacturing, production process monitoring is a core element in achieving stable product quality, improving production efficiency, and ensuring production safety. Real-time sensing, analysis, and early warning of the operating status of key production equipment form the foundation of modern intelligent monitoring systems.

[0003] As a crucial step in the pharmaceutical industry, the extraction process of traditional Chinese medicine requires a monitoring system designed to achieve precise control and comprehensive understanding of the process status by collecting operating parameters from core equipment such as extraction tanks and pumps. The fundamental goal of this technology is to ensure the stable execution of the extraction process and prevent equipment failures from impacting production continuity.

[0004] Existing technologies typically rely on centralized data processing models deployed on central servers, collecting signals such as equipment vibration, temperature, and pressure through sensor networks and uploading them to the cloud for analysis. However, this architecture faces significant challenges in complex industrial environments such as traditional Chinese medicine extraction workshops: the real-time transmission of massive amounts of time-series data puts enormous pressure on network bandwidth, resulting in high latency from data acquisition to analysis and decision-making, making it difficult to meet the millisecond-level response requirements for sudden equipment failures.

[0005] Centralized processing struggles to effectively uncover early wear characteristics hidden within equipment operating data. These subtle features are easily drowned out or diluted by noise during long-distance transmission and massive data aggregation, rendering the system inadequate in predicting gradual equipment degradation. Furthermore, the complex production environment and unstable network conditions exacerbate the risks to data transmission integrity and real-time performance, resulting in blind spots and delayed warnings for cloud-based monitoring systems when facing early anomalies in critical equipment.

[0006] In scenarios like traditional Chinese medicine extraction, where production continuity and equipment reliability are extremely important, how to achieve real-time, accurate, and localized intelligent monitoring of early wear and potential failures of key equipment has become a pressing technical challenge. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing, so as to solve the problems of blind spots and delayed early warning of early equipment anomalies caused by data transmission delay, network bandwidth pressure and weak feature overload in the existing centralized monitoring architecture in the traditional Chinese medicine extraction workshop environment.

[0008] This invention provides an intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing, comprising: The field sensing layer is deployed on the main body of key production equipment such as Chinese medicine extraction tanks, transfer pumps, condensers and valve groups to collect high-fidelity raw operating status time-series data. The edge intelligence layer consists of edge computing nodes deployed in the control cabinets of each production unit. Each edge computing node is directly connected to the field perception layer sensors of its unit through an industrial bus, and is used to perform on-site real-time processing, feature extraction and basic intelligent diagnosis of raw data. The cloud collaboration layer, consisting of remote data center servers, communicates asynchronously and aggregatedly with all edge computing nodes through the factory backbone network, and is used for macro-situation analysis, model optimization, and long-term data management.

[0009] Preferably, each edge computing node in the edge intelligence layer includes a data preprocessing and caching module, a multi-scale feature adaptive extraction module, a local lightweight diagnostic model module, and a decision and communication agent module; The data preprocessing and caching module is used to perform sliding window segmentation on the received raw sensor signals. The window length is preset according to the sensor type. The data within the window is standardized and filtered for noise reduction. The processed data blocks are cached in the node's non-volatile memory. The multi-scale feature adaptive extraction module is used to run the time-domain statistical feature engine, the frequency domain transformation feature engine, and the time-frequency joint analysis engine in parallel to extract features from the preprocessed data window. The local lightweight diagnostic model module embeds a pruned and quantized deep neural network model. This deep neural network model takes a 17-dimensional feature vector as input, passes through two fully connected hidden layers for nonlinear transformation, and generates a 3-dimensional state probability vector at the output layer. The three dimensions of the state probability vector correspond to the predicted probabilities of three states: equipment health, early wear, and potential failure, respectively. The total number of parameters of the deep neural network model is compressed. The decision and communication agent module is used to receive the state probability vector output by the local lightweight diagnostic model module in real time and execute local decision logic; when the probability value of the healthy state is greater than 0.85, the device is determined to be operating normally, and the feature vector and diagnostic results are compressed and uploaded to the cloud collaboration layer at a frequency of once per hour. When the probability of early wear or potential failure exceeds the threshold, a local early warning is immediately triggered and an audible and visual alarm signal is sent to the on-site human-machine interface. At the same time, a high-priority data packet containing the original data window, complete feature vector, and high-probability anomaly diagnosis results is synchronized to the cloud collaboration layer in real time.

[0010] Preferably, the time-domain statistical feature engine is used to calculate the mean, variance, peak factor, and waveform index of each data window; The frequency domain transform feature engine is used to perform a fast Fourier transform on the data window to extract the amplitude and centroid frequency of the first five main frequency components. The time-frequency joint analysis engine is used to decompose the signal to the fourth level using wavelet packet transform and to calculate the energy proportion of the eight sub-bands on the fourth level. The total 17-dimensional feature vectors extracted by the time-domain statistical feature engine, the frequency-domain transform feature engine, and the time-frequency joint analysis engine constitute a feature set used to characterize the current microscopic operating state of the device.

[0011] Preferably, the cloud collaboration layer includes a model training and optimization center, a global knowledge base, and a macro-health management platform; The model training and optimization center is used to continuously receive high-priority data packets uploaded from each edge computing node and periodically uploaded normal status data, and to use this data to incrementally train and validate the baseline version of the deep neural network model. The global knowledge base is used to associate, store, and mine the device status sequences, feature evolution trends, and maintenance records reported by all edge nodes, and to construct a cross-device and cross-production line device health map. The macro-health management platform is used to provide production managers with an overview of the health status of all equipment in the plant, predictive maintenance scheduling suggestions, and process parameter optimization analysis reports based on data from a global knowledge base.

[0012] Preferably, the incremental training and verification process of the model training and optimization center is as follows: maintain an abnormal sample library containing all historical high-priority data packets, and a normal sample library randomly sampled from the normal data of each node; after accumulating 1000 new samples, start a model fine-tuning training once, use the backpropagation algorithm to update the model weights, and send the new model parameters with improved performance after training and verification to the corresponding edge computing nodes to complete the online update of the local model.

[0013] Preferably, the sensor configuration of the field sensing layer is designed for the core equipment for traditional Chinese medicine extraction; For the extraction tank, a triaxial vibration acceleration sensor is installed at the bearing housing of the driving end and non-driving end of the stirring motor, multiple temperature sensors are evenly arranged on the outer wall of the tank, and pressure sensors and flow meters are installed in the inlet and outlet pipelines. For the transfer pump, vibration sensors and temperature sensors are installed on the pump casing and motor casing; All sensors transmit raw analog or digital signals directly to the data preprocessing and caching module of their respective edge computing nodes via a 4mA to 20mA current loop or Modbus RTU protocol.

[0014] Preferably, the communication between the edge computing node and the cloud collaboration layer adopts an adaptive compression and differential synchronization protocol; For low-priority data that is uploaded periodically, the decision and communication agent module uses a dictionary-based lightweight compression algorithm to encode the feature vectors. For high-priority data packets that are synchronized in real time, the adaptive compression and differential synchronization protocol first checks whether the cloud collaboration layer already has the reference data packet sent in the previous cycle; if it does, it only transmits the differential data and difference index between the current data packet and the reference data packet. After successfully receiving the differential data, the cloud collaboration layer merges and reconstructs it with the locally stored baseline data packet to obtain the complete data.

[0015] Preferably, the system operates under a unified time synchronization framework; The cloud collaboration layer acts as a primary time server, broadcasting time synchronization signals to all edge computing nodes via the network time protocol. Each edge computing node integrates a high-precision clock chip, which performs local clock calibration after receiving a timing signal, and uses this as a reference to stamp all raw data collected from the field perception layer with a unified timestamp with microsecond-level precision.

[0016] Preferably, the local lightweight diagnostic model module adds a feature attention subnetwork before the original deep neural network model; The feature attention subnetwork is used to receive the 17-dimensional feature vector as input and output an attention weight vector of the same dimension. The 17-dimensional feature vector is multiplied element-wise with the attention weight vector to obtain a weighted feature vector, which is then fed into a subsequent deep neural network for state classification.

[0017] Preferably, the feature attention subnetwork and the deep neural network are jointly trained, and the objective function of the training is based on the cross-entropy loss, with the addition of a sparse regularization term for the attention weight vector to encourage the model to focus on a few key features.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a three-layer distributed architecture of on-site perception, edge intelligence, and cloud collaboration, decentralizing the core processing and analysis tasks of massive time-series data to edge computing nodes on the production site. This design fundamentally eliminates the network bandwidth pressure and transmission latency caused by uploading all raw data to the cloud. Edge nodes can perform data preprocessing, multi-dimensional feature extraction, and millisecond-level intelligent diagnosis locally and in real time, realizing the immediate capture and response to early weak abnormal characteristics of equipment, meeting the stringent requirements of millisecond-level monitoring of equipment status in the traditional Chinese medicine extraction process.

[0019] 2. The multi-scale adaptive feature extraction module proposed in this invention systematically extracts 17-dimensional feature vectors representing the micro-state of equipment from the original signal through a parallel runtime domain, frequency domain, and time-frequency joint analysis engine. This multi-scale, multi-dimensional feature extraction method can more comprehensively and precisely characterize the health and abnormal modes of equipment operation. In particular, it can effectively enhance the ability to capture weak and nonlinear features exhibited in the gradual degradation process such as early wear, thereby improving the system's predictive sensitivity and accuracy for potential faults.

[0020] 3. This invention deploys a deeply optimized lightweight diagnostic model at the edge intelligence layer. This model has a small parameter size and fast inference speed, perfectly adapting to the limited computing resources of edge nodes. Combined with local decision-making logic, the system can autonomously determine the device status and execute differentiated communication strategies, only synchronizing with the cloud with high priority when a potential anomaly is detected. This integrated "sensing-computing-decision" edge autonomy capability not only significantly reduces cloud load and communication overhead, but more importantly, it achieves a local closed loop for monitoring and decision-making, shortening the response link from anomaly perception to on-site early warning, and effectively avoiding early warning loss or delay due to network fluctuations.

[0021] 4. This invention establishes a cloud-based collaborative layer for continuous optimization of edge models and a global knowledge base. The cloud utilizes aggregating abnormal and normal samples from various edge nodes to periodically incrementally train and update the diagnostic model, and then distributes the optimized model parameters, enabling the edge intelligence deployed on-site to continuously evolve and adapt to state mode changes caused by equipment aging and process adjustments. The global knowledge base achieves cross-device and cross-time dimension health data fusion and mining, providing a solid data foundation and knowledge support for macro-level decision-making and predictive maintenance in production management, forming a complete technical closed loop from micro-level real-time monitoring to macro-level optimization management. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-scale feature adaptive extraction module in this invention; Figure 3 This is a logical flowchart of the integrated "sensing-computing-decision" processing of the edge intelligence layer in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between edge computing nodes and the cloud collaboration layer in this invention; Figure 5 This is a schematic diagram illustrating the principle framework of the cloud-based collaborative layer for continuous optimization and knowledge management of edge models in this invention. Detailed Implementation

[0023] Example 1: Reference Figures 1 to 5 The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing proposed in this invention has an overall technical architecture consisting of a three-layer distributed collaborative system, namely, a field perception layer, an edge intelligence layer, and a cloud collaboration layer.

[0024] This three-tiered distributed collaborative system decentralizes data collection, feature extraction, preliminary diagnosis, and decision execution to the edge of the production site, while retaining the cloud's capabilities for macro-level analysis of the overall situation and continuous model optimization. This enables millisecond-level real-time monitoring of the operating status of key equipment during the traditional Chinese medicine extraction process, accurate early anomaly identification, and predictive maintenance closed-loop management. Please refer to the appendix. Figure 1 , attached Figure 1 It fully demonstrates the logical connections and data flow between the three-tier architecture of this system.

[0025] The field sensing layer is deployed on the core production equipment within the traditional Chinese medicine extraction workshop, including but not limited to extraction tanks, transfer pumps, condensers, and valve assemblies. The core task of this field sensing layer is to acquire high-fidelity, high-frequency raw physical quantity signals from the equipment during operation and transmit the data to its corresponding edge computing node using standardized industrial communication protocols. The sensor configuration of the field sensing layer is highly specialized, taking into account the characteristics of the traditional Chinese medicine extraction process.

[0026] For the extraction tank, triaxial vibration acceleration sensors are installed at the bearing housings of the driving and non-driving ends of the stirring motor to synchronously monitor vibration acceleration signals in the X, Y, and Z directions. Four temperature sensors are evenly arranged circumferentially on the outer wall of the tank to capture the heat distribution in different areas of the tank. Pressure sensors and flow meters are installed on the inlet and outlet pipes to monitor fluid pressure fluctuations and medium flow rate changes, respectively.

[0027] For the transfer pump, vibration sensors and temperature sensors are installed on the pump casing and the matching motor housing, respectively, to comprehensively evaluate the operational stability of the pump's mechanical structure and drive system. All of these sensors transmit the raw analog or digital signals directly to the data preprocessing and caching module of their respective edge computing nodes via a 4mA to 20mA current loop or Modbus RTU protocol at a sampling frequency greater than 1kHz. This sampling frequency ensures effective detection of early abnormal characteristics such as high-frequency weak vibrations and transient pressure fluctuations in the equipment.

[0028] The edge intelligence layer consists of edge computing nodes deployed in the control cabinets of each production unit. Each edge computing node establishes a point-to-point or bus-based connection with all field sensing layer sensors in its unit through an industrial bus (such as CAN bus or RS485), forming an independent local monitoring subsystem.

[0029] Each edge computing node integrates a high-performance embedded processor, non-volatile memory, a high-precision real-time clock chip, and an industrial-grade communication interface. Its core functional modules include a data preprocessing and caching module, a multi-scale adaptive feature extraction module, a local lightweight diagnostic model module, and a decision and communication agent module.

[0030] These four modules work collaboratively in a pipeline manner within the edge node, realizing a unified local autonomous processing logic integrating "sensing-computing-decision". Please refer to the attached document. Figure 3 , attached Figure 3 It describes in detail the data flow and processing sequence between the various functional modules within the edge intelligence layer.

[0031] The data preprocessing and caching module first receives the raw sensor signal stream from the field sensing layer. Since the dynamic range and noise characteristics of signals corresponding to different sensor types differ, the data preprocessing and caching module first performs channel separation and normalization mapping on the input signal according to the preset sensor type identifier.

[0032] Subsequently, the module employs a sliding window mechanism to segment the continuous signal stream. The length of each window is dynamically set according to the sensor type: for vibration sensors, the window length is set to 2 seconds; for temperature, pressure, and flow sensors, the window length is set to 10 seconds.

[0033] The window sliding step size is uniformly set to 50% of the window length to ensure the temporal continuity of feature extraction. The raw data within each window then enters the standardization and filtering / denoising stage. Standardization adopts the Z-score method, which subtracts the mean from the data within the window and divides it by the standard deviation, so that the output data follows a distribution with zero mean and unit variance.

[0034] The filtering and noise reduction employs a fifth-order Butterworth low-pass filter, with a cutoff frequency set to 200 Hz based on the normal operating frequency spectrum characteristics of the equipment, effectively suppressing high-frequency electromagnetic interference and mechanical noise. The preprocessed data blocks are written to the non-volatile memory of the edge nodes, forming a timestamped local cache queue for subsequent use by the feature extraction module.

[0035] The multi-scale adaptive feature extraction module obtains the processed data window from the data preprocessing and caching module and launches three independent feature extraction engines in parallel: a time-domain statistical feature engine, a frequency-domain transform feature engine, and a time-frequency joint analysis engine. These three engines together constitute a multi-dimensional, multi-scale feature generation system, aiming to comprehensively characterize the current microscopic operating state of the equipment. Please refer to the appendix. Figure 2 , attached Figure 2 The parallel processing structure of the three engines and their output fusion mechanism are clearly demonstrated.

[0036] The time-domain statistical feature engine performs calculations on four basic statistical measures for each data window. These include: the arithmetic mean of the signal within the window, reflecting the DC offset level of the signal; variance, characterizing the dispersion of the signal around the mean; peak factor, defined as the ratio of the signal peak value to the root mean square value, used to measure the signal's impulsiveness; and waveform index, defined as the ratio of the root mean square value to the rectified average value, used to describe the sharpness of the signal waveform. The time-domain statistical feature engine outputs a total of four-dimensional features.

[0037] The frequency domain transformation feature engine performs a Fast Fourier Transform on the same data window, converting the time-domain signal to the frequency domain. After the transformation, the engine extracts the amplitudes of the top five frequency components with the highest energy proportions, serving as intensity indicators of the equipment's main vibration or wave modes. Simultaneously, it calculates the signal's centroid frequency, which is the average frequency of all frequency components weighted by their amplitudes, used to characterize the concentrated area of ​​signal energy. The frequency domain transformation feature engine outputs a total of 6 features (5 amplitude values ​​+ 1 centroid frequency).

[0038] The time-frequency joint analysis engine employs wavelet packet transform for more refined time-frequency decomposition of the signal. Using the db4 wavelet basis function, the engine decomposes the signal to level 4, resulting in 16 sub-bands. Since high-frequency sub-bands typically contain significant noise, the engine selects only the 8 sub-bands with the highest energy concentration in level 4 and calculates the percentage of energy in each sub-band relative to the total signal energy. The engine outputs a total of 8-dimensional features.

[0039] The outputs of the three engines are concatenated to form a 17-dimensional feature vector. This 17-dimensional feature vector fully covers the device's multi-scale operating characteristics in the time domain, frequency domain, and time-frequency domain, providing a high-information-density input representation for subsequent intelligent diagnostics.

[0040] The local lightweight diagnostic model module deploys a deeply compressed deep neural network model embedded within it. The local lightweight diagnostic model adopts a fully connected feedforward structure with an input layer dimension of 17, corresponding to the aforementioned 17-dimensional feature vector; two hidden layers with 32 and 16 nodes respectively; and an output layer dimension of 3, corresponding to the probability output of three device states: "healthy", "early wear", and "potential fault".

[0041] To adapt to the limited computing resources of edge computing nodes, the local lightweight diagnostic model underwent rigorous pruning and quantization after training. Pruning removed connections with weights whose absolute values ​​were less than 0.01, and quantization converted floating-point weights and activation values ​​into 8-bit integer representations. Ultimately, the total number of parameters in the local lightweight diagnostic model was compressed to 48,720, far less than the threshold of 50,000. On typical edge hardware platforms (such as NVIDIA Jetson Nano or Huawei Atlas 500), this model can complete a single forward inference within 50 milliseconds.

[0042] The reasoning process is as follows: The local lightweight diagnostic model receives a 17-dimensional feature vector as input, which is then passed through two hidden layers via linear transformation and the ReLU activation function. Finally, a 3-dimensional state probability vector is generated at the output layer using the Softmax function. The sum of the three elements of this state probability vector is 1, representing the confidence level of the device in one of the three states.

[0043] The decision and communication agent module monitors the output of the local lightweight diagnostic model module in real time. This module incorporates a deterministic local decision-making logic to trigger different response actions based on the state probability vector. When the probability value of the "healthy" state is greater than 0.85, the device is determined to be operating normally.

[0044] At this point, the decision-making and communication agent module performs lightweight compression (using a dictionary compression algorithm based on Huffman coding) on ​​the current 17-dimensional feature vector and 3-dimensional state probability vector, marks it as a "low-priority data packet," and uploads it to the cloud collaboration layer at a fixed interval of once per hour. When the probability value of the "early wear" or "potential failure" state exceeds the threshold, it is immediately determined to be an abnormal event.

[0045] The decision-making and communication agent module then performs three operations: First, send an audible and visual alarm signal to the human-machine interface of the production unit to alert the operator. Second, retrieve the original data window (including timestamp) that triggered this diagnosis from the local cache. Third, the original data window, the complete 17-dimensional feature vector, and the 3-dimensional state probability vector are packaged, marked as "high-priority data packets," and synchronized to the cloud collaboration layer in real time through the factory backbone network.

[0046] This differentiated communication strategy significantly reduces network load under normal operating conditions while ensuring zero-latency reporting of abnormal events.

[0047] Communication between edge computing nodes and the cloud collaboration layer employs an adaptive compression and differential synchronization protocol. For low-priority data packets, the decision and communication agent module uses a pre-shared static dictionary to encode feature vectors, achieving a compression rate of over 60%. For high-priority data packets, the protocol first queries the cloud collaboration layer to confirm whether the device successfully uploaded the baseline data packet in the previous cycle.

[0048] If present, only the differential data (i.e., the numerical difference) and the difference index (indicating which fields have changed) between the current data packet and the reference data packet are transmitted. After receiving the differential data, the cloud collaboration layer uses the locally stored reference data packet to merge and reconstruct it, restoring the complete high-priority data packet. This can reduce the transmission volume of the high-priority data packet by 40% to 70%, effectively alleviating instantaneous network congestion.

[0049] The entire system operates under a unified time synchronization framework. The cloud collaboration layer, acting as a primary time server, periodically broadcasts time signals to all edge computing nodes via a network time protocol. Each edge computing node integrates a high-precision temperature-compensated crystal clock chip. Upon receiving the time signal, it uses the least squares method to fit the local clock drift curve and performs dynamic calibration.

[0050] The calibrated local clock timestamps all raw data collected from the field sensing layer with microsecond-level precision, ensuring strict alignment of data across devices and nodes on the timeline and providing a reliable time reference for subsequent global event correlation analysis, multi-source signal fusion, and fault propagation path tracing.

[0051] The cloud-based collaboration layer consists of a remote data center server cluster, establishing asynchronous, aggregated communication connections with all edge computing nodes through the factory's backbone network. Its core functional components include a model training and optimization center, a global knowledge base, and a macro-level health management platform. Please refer to the appendix. Figure 4 With appendix Figure 5 These two figures respectively illustrate the interactive data flow between the edge and the cloud, and the knowledge management and model optimization mechanisms within the cloud.

[0052] The model training and optimization center continuously monitors data uploads from each edge node. Internally, the center maintains two core sample libraries: an anomaly sample library and a normal sample library. The anomaly sample library gathers feature vectors and label information from all "high-priority data packets" uploaded by all edge nodes; the normal sample library is constructed by uniformly sampling data over time from "low-priority data packets" uploaded periodically by each node.

[0053] Whenever 1000 valid samples are added to any sample database, the center automatically triggers a model fine-tuning training process. The training uses the backpropagation algorithm, with the cross-entropy loss function as the optimization objective, to update the parameters of the baseline version of the deep neural network model.

[0054] After training, the new model is evaluated on an independent validation set. If the accuracy improvement is greater than 1%, the new model parameters are packaged into an incremental update package and distributed to the corresponding edge computing nodes via a secure encrypted channel. Edge nodes automatically load the new model during idle periods and replace the old version, completing the online evolution of local diagnostic capabilities.

[0055] The global knowledge base provides structured storage and relational modeling for the equipment status sequences, feature evolution trends, early warning records, and manual maintenance logs reported by all edge nodes. Employing graph database technology, the knowledge base uses entities such as equipment, production lines, process batches, and maintenance events as nodes, and their spatiotemporal, causal, and hierarchical relationships as edges, constructing a cross-equipment and cross-production line equipment health graph. This equipment health graph supports complex path queries and pattern mining, such as "the correlation strength between the increase in vibration waveform indicators and the drift of the average temperature of a certain type of pump after 200 hours of continuous operation."

[0056] The macro-health management platform provides three core services to production managers based on data from a global knowledge base.

[0057] First, a comprehensive overview of the health status of all equipment in the plant, displaying the distribution of health scores for each production line in real time in the form of a heat map; Second, predictive maintenance scheduling is provided, which automatically generates a list of maintenance tasks and their priorities for the next 7 days based on the degradation trend prediction in the equipment health map. Third, the process parameter optimization analysis report identifies key process factors affecting equipment lifespan by linking equipment status with process parameters (such as extraction temperature, pressure, and time), and proposes parameter adjustment suggestions.

[0058] In summary, this embodiment constructs a highly real-time, highly sensitive, low-bandwidth-dependent, and self-evolving intelligent monitoring system for the extraction process of traditional Chinese medicine through the close collaboration of a three-layer architecture, system extraction of multi-scale features, edge deployment of a lightweight model, intelligent execution of differentiated communication strategies, and continuous evolution of the cloud model. This system effectively solves the latency, bandwidth, and weak feature overload problems faced by traditional centralized architectures in complex industrial environments, providing solid technical support for intelligent manufacturing of traditional Chinese medicine.

[0059] Example 2: Building upon Example 1, this example further optimizes the local lightweight diagnostic model module of the edge intelligence layer by introducing a dynamic feature weighting strategy based on an attention mechanism to improve the model's sensitivity to key features and diagnostic robustness. This optimized approach is particularly suitable for traditional Chinese medicine extraction scenarios where equipment operating conditions frequently switch or where there is strong environmental interference.

[0060] Specifically, the local lightweight diagnostic model module adds a feature attention sub-network before the original two fully connected layers. This feature attention sub-network receives a 17-dimensional original feature vector as input and outputs an attention weight vector of the same dimension. Each element of the attention weight vector ranges from 0 to 1, representing the importance of the corresponding feature under the current condition. The original feature vector and the attention weight vector are multiplied element-wise to obtain a weighted feature vector, which is then fed into the subsequent fully connected network for state classification.

[0061] The feature attention subnetwork employs a single-layer fully connected structure, with its output constrained to the 0-1 range by a Sigmoid activation function. This feature attention subnetwork is jointly trained with the main diagnostic network. The loss function, based on the original cross-entropy loss, adds a sparse regularization term to encourage the model to focus on a few key features and avoid attention dispersion. The objective function for joint training is... This can be expressed as: ; For cross-entropy loss, For attention weight vectors, is the sparse regularization coefficient, with a value set to 0.01. The formula shows that while optimizing classification accuracy, the model also tends to generate a sparse attention distribution, thereby automatically identifying and strengthening the feature dimensions that are most discriminative for diagnosis.

[0062] In actual operation, when the equipment is in a stable extraction phase, the attention subnetwork may assign higher weights to process-related features such as mean temperature and pressure variance; while when the equipment is starting up, stopping, or operating at variable frequency, the weights may be concentrated on mechanical state features such as vibration peak factor and waveform indicators. This dynamic weighting mechanism enables the diagnostic model to adapt to different operating conditions, significantly improving the anomaly detection rate under non-steady-state operating conditions.

[0063] To ensure that the locally lightweight diagnostic model still meets the resource constraints of edge deployment, the entire network (including the attention subnetwork) undergoes channel pruning and 8-bit integer quantization again after training. The final model has a total number of parameters controlled below 52,000, and the single inference time on typical edge hardware is still less than 60 milliseconds, meeting real-time requirements.

[0064] Furthermore, this embodiment refines the local decision-making logic of the decision and communication agent module. When determining the "early wear" state, the module not only checks the probability threshold but also incorporates a feature evolution trend judgment.

[0065] Specifically, the decision and communication agent module continuously caches the 17-dimensional feature vectors from the last five diagnostic cycles and calculates the slope of key features (such as vibration variance and mean temperature). If the absolute value of the slope is greater than a preset threshold (e.g., 0.05 / minute), even if the current probability of "early wear" is only 0.55, it is considered a valid warning and triggers a high-priority report. This prevents missed reports due to fluctuations in a single diagnosis and enhances the system's ability to track gradual degradation.

[0066] The model training and optimization center in the cloud-based collaborative layer also adjusted its training strategy accordingly. When constructing training samples, data from different operating conditions (such as extraction, cleaning, and standby) were augmented with labels to explicitly indicate the operating condition type. During training, a multi-task learning framework was adopted, with the primary task being equipment state classification and the auxiliary task being operating condition recognition. This allowed the model to learn state discrimination capabilities while also internalizing operating condition perception capabilities, further improving the accuracy of the attention mechanism.

[0067] Through the above enhancements, this embodiment significantly improves the environmental adaptability and diagnostic reliability of the edge diagnostic model while maintaining the overall system architecture, making it particularly suitable for modern Chinese medicine extraction workshops with complex and ever-changing operating conditions.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing, characterized in that, include: The field sensing layer is deployed on the main body of key production equipment such as Chinese medicine extraction tanks, transfer pumps, condensers and valve groups to collect high-fidelity raw operating status time-series data. The edge intelligence layer consists of edge computing nodes deployed in the control cabinets of each production unit. Each edge computing node is directly connected to the field perception layer sensors of its unit through an industrial bus, and is used to perform on-site real-time processing, feature extraction and basic intelligent diagnosis of raw data. The cloud collaboration layer, consisting of remote data center servers, communicates asynchronously and aggregatedly with all edge computing nodes through the factory backbone network, and is used for macro-situation analysis, model optimization, and long-term data management.

2. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 1, characterized in that, Each edge computing node in the edge intelligence layer includes a data preprocessing and caching module, a multi-scale feature adaptive extraction module, a local lightweight diagnostic model module, and a decision and communication agent module. The data preprocessing and caching module is used to perform sliding window segmentation on the received raw sensor signals. The window length is preset according to the sensor type. The data within the window is standardized and filtered for noise reduction. The processed data blocks are cached in the node's non-volatile memory. The multi-scale feature adaptive extraction module is used to run the time-domain statistical feature engine, the frequency domain transformation feature engine, and the time-frequency joint analysis engine in parallel to extract features from the preprocessed data window. The local lightweight diagnostic model module embeds a pruned and quantized deep neural network model. This deep neural network model takes a 17-dimensional feature vector as input, passes through two fully connected hidden layers for nonlinear transformation, and generates a 3-dimensional state probability vector at the output layer. The three dimensions of the state probability vector correspond to the predicted probabilities of three states: equipment health, early wear, and potential failure, respectively. The total number of parameters of the deep neural network model is compressed. The decision and communication agent module is used to receive the state probability vector output by the local lightweight diagnostic model module in real time and execute local decision logic; when the probability value of the healthy state is greater than 0.85, the device is determined to be operating normally, and the feature vector and diagnostic results are compressed and uploaded to the cloud collaboration layer at a frequency of once per hour. When the probability of early wear or potential failure exceeds the threshold, a local early warning is immediately triggered and an audible and visual alarm signal is sent to the on-site human-machine interface. At the same time, a high-priority data packet containing the original data window, complete feature vector, and high-probability anomaly diagnosis results is synchronized to the cloud collaboration layer in real time.

3. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 2, characterized in that, The time-domain statistical feature engine is used to calculate the mean, variance, peak factor, and waveform index of each data window. The frequency domain transform feature engine is used to perform a fast Fourier transform on the data window to extract the amplitude and centroid frequency of the first five main frequency components. The time-frequency joint analysis engine is used to decompose the signal to the fourth level using wavelet packet transform and to calculate the energy proportion of the eight sub-bands on the fourth level. The total 17-dimensional feature vectors extracted by the time-domain statistical feature engine, the frequency-domain transform feature engine, and the time-frequency joint analysis engine constitute a feature set used to characterize the current microscopic operating state of the device.

4. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 3, characterized in that, The cloud-based collaboration layer includes a model training and optimization center, a global knowledge base, and a macro-health management platform. The model training and optimization center is used to continuously receive high-priority data packets uploaded from each edge computing node and periodically uploaded normal status data, and to use this data to incrementally train and validate the baseline version of the deep neural network model. The global knowledge base is used to associate, store, and mine the device status sequences, feature evolution trends, and maintenance records reported by all edge nodes, and to construct a cross-device and cross-production line device health map. The macro-health management platform is used to provide production managers with an overview of the health status of all equipment in the plant, predictive maintenance scheduling suggestions, and process parameter optimization analysis reports based on data from a global knowledge base.

5. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 4, characterized in that, The incremental training and validation process of the model training and optimization center is as follows: maintain an abnormal sample library containing all historical high-priority data packets, and a normal sample library randomly sampled from the normal data of each node; after accumulating 1000 new samples, start a model fine-tuning training once, use the backpropagation algorithm to update the model weights, and send the new model parameters with improved performance after training and validation to the corresponding edge computing nodes to complete the online update of the local model.

6. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 5, characterized in that, The sensor configuration of the field perception layer is designed for core equipment for traditional Chinese medicine extraction; For the extraction tank, a triaxial vibration acceleration sensor is installed at the bearing housing of the driving end and non-driving end of the stirring motor, multiple temperature sensors are evenly arranged on the outer wall of the tank, and pressure sensors and flow meters are installed in the inlet and outlet pipelines. For the transfer pump, vibration sensors and temperature sensors are installed on the pump casing and motor casing; All sensors transmit raw analog or digital signals directly to the data preprocessing and caching module of their respective edge computing nodes via a 4mA to 20mA current loop or Modbus RTU protocol.

7. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 5, characterized in that, The communication between the edge computing nodes and the cloud collaboration layer adopts an adaptive compression and differential synchronization protocol; For low-priority data that is uploaded periodically, the decision and communication agent module uses a dictionary-based lightweight compression algorithm to encode the feature vectors. For high-priority data packets that are synchronized in real time, the adaptive compression and differential synchronization protocol first checks whether the cloud collaboration layer already has the reference data packet sent in the previous cycle; if it does, it only transmits the differential data and difference index between the current data packet and the reference data packet. After successfully receiving the differential data, the cloud collaboration layer merges and reconstructs it with the locally stored baseline data packet to obtain the complete data.

8. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 7, characterized in that, The system operates under a unified time synchronization framework; The cloud collaboration layer acts as a primary time server, broadcasting time synchronization signals to all edge computing nodes via the network time protocol. Each edge computing node integrates a high-precision clock chip, which performs local clock calibration after receiving a timing signal, and uses this as a reference to stamp all raw data collected from the field perception layer with a unified timestamp with microsecond-level precision.

9. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 8, characterized in that, The local lightweight diagnostic model module adds a feature attention subnetwork before the original deep neural network model; The feature attention subnetwork is used to receive the 17-dimensional feature vector as input and output an attention weight vector of the same dimension. The 17-dimensional feature vector is multiplied element-wise with the attention weight vector to obtain a weighted feature vector, which is then fed into a subsequent deep neural network for state classification.

10. The intelligent monitoring system for the extraction process of traditional Chinese medicine based on edge computing according to claim 9, characterized in that, The feature attention subnetwork is jointly trained with the deep neural network. The objective function of the training is based on the cross-entropy loss, with the addition of a sparse regularization term for the attention weight vector to encourage the model to focus on a few key features.

Images