Search

CN-121998625-A - Equipment predictive maintenance system based on multi-mode data fusion and edge collaborative decision

CN121998625ACN 121998625 ACN121998625 ACN 121998625ACN-121998625-A

Abstract

A device predictive maintenance system based on multi-mode data fusion and edge collaborative decision-making comprises intelligent data acquisition cards and a unified management platform, wherein the intelligent data acquisition cards are connected with each other through an internal LoRa networking unit to form a dynamic Mesh network, a host intelligent data acquisition card is in communication connection with a cloud through a network unit, the unified management platform comprises a cloud server and a user terminal and provides an AI model management function for supporting remote issuing of a trained MSAD algorithm model file to a designated intelligent data acquisition card or device group in a differential upgrade package mode, the intelligent data acquisition card carries out real-time analysis on acquired multi-mode sensing data based on a built-in MSAD algorithm model to output a quantized abnormal confidence score, and a collaborative verification mechanism is triggered when the abnormal confidence score is in an uncertainty interval preset by the system. The invention obviously improves the accuracy and reliability of equipment fault early warning.

Inventors

  • WU XIAOMAO
  • NAI HUANHUAN

Assignees

  • 众华电子科技(太仓)有限公司

Dates

Publication Date
20260508
Application Date
20260331

Claims (10)

  1. 1. The equipment predictive maintenance system based on multi-mode data fusion and edge collaborative decision-making is characterized by comprising an edge side, wherein the edge side comprises at least two intelligent data acquisition cards, the intelligent data acquisition cards are deployed on equipment and used for data acquisition and control, and the intelligent data acquisition cards are mutually connected through an internal LoRa networking unit to form a dynamic Mesh network, one intelligent data acquisition card is selected as a host node, and the other intelligent data acquisition cards are collaborative nodes; the system comprises a unified management platform, an AI model management function and a data processing module, wherein the unified management platform comprises a cloud server and a user terminal, the cloud server performs data interaction with a host node, and the user terminal accesses the cloud server through the Internet to monitor and manage; The specific flow of differential upgrading is as follows: 1) The cloud server compares the new version MSAD algorithm model file with the old version model file byte by byte, extracts a difference part (Delta) to generate a differential upgrade packet, and calculates an MD5 value of the differential upgrade packet; 2) The cloud server splits the differential packet into data blocks and transmits the data blocks to a designated target intelligent data acquisition card one by one through a host node; in the scheme, a mesh network is adopted, each device in the networking has respective numbers (specifically operated by a wireless module), and a user can also correspond the numbers to the corresponding arrangement positions by himself/herself, and upgrade is performed according to the numbers when upgrade is required; 3) After the verification is passed, the acquisition card utilizes a built-in Bootloader program to combine the differential packet with the local old model to generate a new model and load and operate; 4) The safe rollback, namely if the new model fails to be loaded or the system is unstable, the acquisition card is automatically switched back to the old version, and the loss log is reported to the management platform; The intelligent data acquisition card carries out real-time analysis on locally acquired multi-mode sensing data based on a built-in MSAD algorithm model and outputs a quantized abnormal confidence score, when the abnormal confidence score is in an uncertainty interval between a first high threshold value and a second low threshold value preset by a system, a collaborative verification mechanism is triggered, the collaborative verification mechanism is completed by a node initiating a request, a dynamically elected host node and one or more collaborative nodes, data synchronous acquisition, distributed reasoning and result fusion are finally formed on the edge side.
  2. 2. The predictive maintenance system for equipment based on multi-mode data fusion and edge collaborative decision-making according to claim 1, wherein the intelligent data acquisition card comprises a data acquisition unit, a control unit, a core processing unit, a communication unit and an auxiliary unit, wherein the data acquisition unit, the control unit, the communication unit and the auxiliary unit are all in bidirectional connection with the core processing unit; The data acquisition unit comprises an analog quantity acquisition unit and a digital quantity acquisition unit, and the control unit comprises an analog quantity control unit and a digital quantity control unit; The core processing unit takes the analysis operation unit as a core, and the MSAD algorithm model is arranged in the analysis operation unit; The communication unit comprises a networking unit and a network unit, wherein the cooperation of the analysis operation unit and the networking unit jointly supports an MSAD algorithm model and a dynamic cooperation mechanism; the auxiliary unit includes a power supply unit and an indication unit.
  3. 3. The predictive maintenance system of equipment based on multi-mode data fusion and edge collaborative decision-making according to claim 2, wherein the analysis and operation unit adopts a heterogeneous computing architecture, which comprises a multi-core CPU and a special AI acceleration chip, and is used for running a multi-mode perception anomaly detection MSAD algorithm model.
  4. 4. The predictive maintenance system for devices based on multi-mode data fusion and edge collaborative decision-making according to claim 2, wherein the networking unit is configured to implement inter-node networking communication and data uplink transmission to a host node by adopting a communication mode including, but not limited to, a LoRa wireless communication module, or a Wi-Fi Mesh, zigbee, bluetooth Mesh or other communication technology supporting ad hoc network, and by adopting a data uplink transmission mode to a host node including, but not limited to, a wired ethernet, a wireless Wi-Fi, and a cellular network.
  5. 5. The equipment predictive maintenance system based on multi-mode data fusion and edge collaborative decision-making according to claim 1 is characterized in that an MSAD algorithm model adopts a double-branch deep neural network structure and comprises a numerical anomaly sensing branch for capturing local feature modes and a sequence anomaly sensing branch for capturing long-time sequence dependency relations, wherein the numerical anomaly sensing branch adopts a one-dimensional convolutional neural network, the sequence anomaly sensing branch adopts a gating circulation unit, the MSAD algorithm model carries out fusion analysis on acquired multi-mode time sequence data and outputs anomaly confidence scores and fault type labels, and different levels of responses are triggered according to preset multi-level threshold rules.
  6. 6. The predictive maintenance system for devices based on multi-modal data fusion and edge collaborative decisions of claim 5, wherein the multi-level threshold rules are specifically: If the abnormal confidence coefficient score is larger than a first high threshold value, triggering local control immediately and reporting high-risk alarms; if the abnormal confidence score is between the uncertainty interval between the first high threshold and the second low threshold, starting a collaborative verification mechanism to request verification of the adjacent nodes; If the abnormal confidence score is lower than a second low threshold, only recording data, and not triggering active response; The multi-level threshold calibration method is to calibrate by combining historical data, and is not fixed, and the system provides a threshold self-learning calibration function, which comprises the following steps: 1) Collecting historical data of the equipment in a normal state, inputting the historical data into an MSAD model to obtain confidence distribution, and setting 95% quantiles as a low threshold T_low; 2) Known fault injection or historical fault data is collected to obtain a confidence distribution, and the 5% quantile is set to be a high threshold T_high.
  7. 7. The predictive maintenance system for equipment based on multi-modal data fusion and edge collaborative decision-making according to claim 6, wherein the data processing flow of the MSAD algorithm model is as follows: 1) The method comprises the steps of preprocessing and aligning data, filtering high-noise industrial data by adopting a combined noise reduction algorithm based on wavelet transformation after the original data acquired by a data acquisition unit are received, carrying out 5-layer decomposition on the original signal by adopting a db4' wavelet basis function, carrying out noise reduction processing on each layer of high-frequency detail coefficient by adopting a soft threshold function, and carrying out time sequence asynchronous alignment processing on multi-mode sensor data by adopting a dynamic time alignment algorithm after noise reduction, wherein the size of a bending window is set to be 10 during alignment, and the distance measurement mode is Euclidean distance so as to ensure the synchronism of different mode data on a time axis; 2) Multi-scale feature extraction, namely, synchronously executing time domain feature, frequency domain feature and time-frequency domain feature multi-scale feature value extraction on the preprocessed data by utilizing the parallel computing capability of an AI acceleration chip in an analysis operation unit, and carrying out multi-vector fusion on the extracted feature values by a system to form a single feature vector; 3) The method comprises the steps of anomaly detection and reasoning, namely inputting a fused single feature vector into an MSAD algorithm model for reasoning, namely capturing a local feature mode by adopting a one-dimensional convolutional neural network of a numerical anomaly perception branch, capturing a long-time sequence dependency relationship by adopting a gating circulation unit of a sequence anomaly perception branch, summarizing outputs of the numerical anomaly perception branch and the two branches of the sequence anomaly perception branch in a decision fusion layer to form a fused comprehensive feature, and sending the comprehensive feature into a full-connection layer to finally generate a comprehensive anomaly confidence score 0-1 and a primary fault type label, wherein the anomaly confidence score is a direct basis for subsequently triggering different-level responses and collaborative verification; 4) The system executes different operations through the self-adaptive decision according to a preset multi-level threshold rule.
  8. 8. The predictive maintenance system of equipment based on multi-modal data fusion and edge collaborative decision-making according to claim 1, wherein the intelligent data acquisition card adopts a distributed host election algorithm based on resource perception to elect host nodes, and in the process of electing host nodes, each node resource state is interacted through heartbeat packets, and the host nodes are elected according to a dynamic scoring function, and the host nodes coordinate member nodes to perform data sharing and distributed collaborative calculation, specifically as follows: 1) The initial networking and state broadcasting, namely, entering a network maintenance state after each node is electrified, and broadcasting a heartbeat packet containing self equipment ID, computing capacity (CPU occupancy rate and memory residual), power supply state and signal strength (RSSI) outwards every interval T seconds (T is set to 30s by default and the period can be configured remotely through a management platform); 2) The dynamic host node election is that after the node receives heartbeat packets of other nodes, the node does not trigger the reselection immediately so as to avoid network jitter; Each node maintains an election timer (default duration is 3 times the heartbeat cycle, i.e., 90 seconds); When the timer times out, the node collects the latest heartbeat information of all neighbor nodes received in the window period, substitutes the latest heartbeat information into the following scoring function to calculate scores, and the highest scoring node is identified as the host node; If the host node is not specified, all nodes follow the rule, and the node with the highest score becomes the host node; because the dimensions of CPU_free (percent), memory_free (MB), battery_level (percent) and RSSI (dBm, usually negative value) are different, min-Max normalization processing is required before substitution into a scoring function, and the scoring function is uniformly mapped to a [0, 1] interval; The specific formula is as follows: Wherein, the Taking RSSI as an example, the set of all RSSI original values (negative values, such as-45, -62, -78.) stored in the current sliding window is set as the set of all RSSI original values (negative values, such as-45, -62, -78.) stored in the current sliding window Taking absolute values of all values in the set Find the maximum value in And minimum value Normalizing the RSSI absolute value of the current heartbeat of the current node by using a Min-Max normalization processing method, and completing normalization of other physical quantities based on respective sliding windows in the same way; the normalized parameters are taken into the following formula: The system comprises a CPU_free, a memory_free, a battery_level, an RSSI, an alpha, a beta, a gamma and a delta, wherein the CPU_free is a CPU idle rate, the memory_free is a Memory residual quantity, the battery_level is a Battery capacity percentage, the RSSI is a signal strength, the alpha, the beta, the gamma and the delta are weight coefficients for balancing the importance of each index, and a group of balance coefficients are provided by default, wherein the balance coefficients comprise alpha=0.3, beta=0.2, gamma=0.3 and delta=0.2; The node with the highest score becomes a host node, other nodes are used as members to automatically join the network, and the election mechanism ensures that a network coordinator is always the node with the optimal current resource; 3) When a member node falls into an uncertainty interval according to abnormal MSAD confidence scores output by an MSAD algorithm model, a cooperative verification request is initiated to a host node, after the host node receives the cooperative verification request, one or more other nodes with complementarity on physical positions or data types are instructed according to the request type and network resource state, the designated nodes become cooperative nodes at the moment, data are synchronously acquired and locally analyzed or distributed reasoning is carried out, each cooperative node feeds back an analysis result to the host node, and a final decision is made after information fusion is carried out by the host node.
  9. 9. The predictive maintenance system for a device based on multimodal data fusion and edge collaborative decisions according to claim 8, wherein the "complementarity" determining criteria includes: space complementation, namely, the physical linear distance between the cooperative node and the verified node is smaller than a preset threshold value; The mode complementation is that the sensor type of the cooperative node is different from that of the node to be verified, for example, the vibration node requests the temperature node and the current node to participate in verification, the host node selects the nodes according to a preconfigured cooperative strategy library, the cooperative verification mechanism defaults to select 2 cooperative nodes, the network overhead is balanced while the decision reliability is ensured, and a user can set the upper limit of the number of the cooperative nodes in a cooperative strategy configuration interface of the management platform, and the default is not more than 3.
  10. 10. The prediction method of the equipment predictive maintenance system based on the multi-mode data fusion and the edge collaborative decision is characterized in that the prediction method is based on the equipment predictive maintenance system based on the multi-mode data fusion and the edge collaborative decision according to any one of claims 1 to 9, and the specific prediction method is as follows: s1, an intelligent data acquisition card acquires multi-mode physical signals of industrial equipment; s2), an analysis operation unit in the core processing unit adopts an MSAD algorithm model to analyze the multi-mode physical signals acquired in the S1) in real time, and a decision fusion layer fuses the data to obtain quantized abnormal confidence scores and fault type labels; S3) judging according to a set multilevel threshold rule, and if the abnormal confidence coefficient is between an uncertainty interval between a high threshold value and a low threshold value, initiating a collaborative verification mechanism request to a host node through the LoRa networking unit; s4, the host node coordinates the relevant cooperative nodes to perform data synchronous acquisition and distributed reasoning; And S5, the host node gathers the local results of the cooperative nodes, performs data fusion and final decision, triggers corresponding control instructions and reports the decision abstract to the management platform.

Description

Equipment predictive maintenance system based on multi-mode data fusion and edge collaborative decision Technical Field The invention belongs to the technical field of data processing, and particularly relates to a device predictive maintenance system and a device predictive maintenance method based on multi-mode data fusion and edge collaborative decision. Background The traditional data acquisition scheme has the remarkable limitations that (1) the function is single, the traditional acquisition card only has the functions of analog-to-digital conversion and IO control, the data analysis completely depends on an upper computer, and the real-time intelligent analysis can not be performed on the edge side of the data generation. (2) The method is lack of synergy, the dimension of data acquired by single points is limited, multi-mode data (such as vibration, temperature and current) of cross-position and cross-type cannot be fused to carry out comprehensive decision, and all the acquisition points are isolated and cannot form a collaborative sensing and control network. (3) The flexibility is poor, and the acquisition logic, the control strategy and the AI model are all solidified by manufacturers or secondarily developed by depending on complex upper computer software, so that dynamic self-adaptive adjustment is difficult to carry out according to the change of the working condition of the site. The existing data acquisition and edge analysis scheme has inherent defects that (1) an edge AI model is isolated and single, namely the existing scheme for introducing edge calculation is mostly aimed at a single data type or simple rule, cross-modal complementary information cannot be effectively fused, the complex fault mode identification capability is limited, and the model output is lack of quantification (such as confidence) of uncertainty of self judgment, so that a decision mechanism is rough. (2) The networking mode is statically rigidified, that is, the existing distributed sensing network mostly adopts preset master-slave nodes or static topology, the network structure and task allocation cannot be dynamically optimized according to the on-site node resource states (calculation power, electric quantity and link quality), and the overall robustness and resource utilization efficiency of the system are low. (3) The decision mechanism lacks coordination, when the edge node encounters an ambiguous abnormal state, the existing scheme either directly reports to the cloud (introduces delay) or relies on the uncertain judgment of a single node to make a key decision, and the capability of utilizing adjacent nodes to perform fast and low-delay collaborative verification and joint decision on the edge side is lacking. Therefore, the prior art cannot realize intelligent operation and maintenance with high reliability, self-adaption and optimal resources on an industrial site. The invention provides an integrated solution capable of realizing edge side multi-mode data fusion, intelligent networking and collaborative decision-making aiming at the defects. Disclosure of Invention The invention aims to overcome the defects, and provides a device predictive maintenance system based on multi-mode data fusion and edge collaborative decision, which effectively solves the problems of insufficient single-point decision reliability, low utilization of static networking resources and singleness of decision mechanism in the traditional edge scheme, and remarkably improves the accuracy and the system reliability of fault early warning through intelligent collaboration among nodes while maintaining the advantage of low delay of the edge. The invention solves the problem that the current data acquisition system lacks intelligent coordination and has fuzzy decision mechanism, and remarkably improves the accuracy and reliability of the fault early warning of the industrial equipment. In order to achieve the above purpose, the invention provides a device predictive maintenance system based on multi-mode data fusion and edge collaborative decision-making, which comprises an edge side, wherein the edge side comprises at least two intelligent data acquisition cards, the intelligent data acquisition cards are deployed on the device for data acquisition and control, and are mutually connected through an internal LoRa networking unit to form a dynamic Mesh network, one intelligent data acquisition card is selected as a host node, and the other intelligent data acquisition cards corresponding to the host node are connected with a cloud through a network unit; The system comprises a unified management platform, an AI model management function and a data processing module, wherein the unified management platform comprises a cloud server and a user terminal, the cloud server performs data interaction with a host node, the user terminal accesses the cloud server through the Internet to monitor and manage, and the AI model management function is